ecocost 0.2.2__tar.gz → 0.2.4__tar.gz

{ecocost-0.2.2 → ecocost-0.2.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ecocost
-Version: 0.2.2
+Version: 0.2.4
 Summary: Per-run carbon, energy and water estimates for AI inference, with computed confidence and sourced inputs
 License: Apache-2.0
 License-File: LICENSE
@@ -66,8 +66,10 @@ r["confidence"]  # {"level": "low", "range_ratio": 16.6, "reasons": ["wue_assume
 Model ids are matched through aliases, so provider-specific ids such as
 `accounts/fireworks/models/gpt-oss-120b` resolve to the same record. A model
 that isn't in the knowledge base raises `UnknownModelError`, which suggests
-close matches and links to [adding it](CONTRIBUTING.md#adding-a-model). Ids
-are matched ignoring case.
+close matches and links to [adding it](CONTRIBUTING.md#adding-a-model). The
+exception is a model listed as a [proxy](API.md#unknown-models-and-providers)
+of a related record: it is estimated from that record and flagged, with a
+wider range and low confidence. Ids are matched ignoring case.
 
 ## Provider ids
 

{ecocost-0.2.2 → ecocost-0.2.4}/README.md

@@ -35,8 +35,10 @@ r["confidence"]  # {"level": "low", "range_ratio": 16.6, "reasons": ["wue_assume
 Model ids are matched through aliases, so provider-specific ids such as
 `accounts/fireworks/models/gpt-oss-120b` resolve to the same record. A model
 that isn't in the knowledge base raises `UnknownModelError`, which suggests
-close matches and links to [adding it](CONTRIBUTING.md#adding-a-model). Ids
-are matched ignoring case.
+close matches and links to [adding it](CONTRIBUTING.md#adding-a-model). The
+exception is a model listed as a [proxy](API.md#unknown-models-and-providers)
+of a related record: it is estimated from that record and flagged, with a
+wider range and low confidence. Ids are matched ignoring case.
 
 ## Provider ids
 

{ecocost-0.2.2 → ecocost-0.2.4}/ecocost/__init__.py

@@ -62,6 +62,10 @@ def estimate(
     region, including one implied by ``endpoint``; the provider still sets
     PUE, water use and hardware.
 
+    A model with no record that a reviewer listed as a proxy of a related one
+    is estimated from it: ``proxy`` in the result names both, the range is
+    wider and confidence is low.
+
     Ids are matched ignoring case. ``UnknownModelError``,
     ``UnknownProviderError`` and ``UnknownRegionError`` suggest close matches and where to add the
     record. Bad arguments raise ``TypeError`` or ``ValueError``.
@@ -76,7 +80,14 @@ def estimate(
         cached_input_tokens,
     )
     kb = get_kb()
-    m = kb.resolve_model(model)
+    proxy = None
+    try:
+        m = kb.resolve_model(model)
+    except UnknownModelError:
+        if not (hit := kb.resolve_proxy(model)):
+            raise
+        m, proxy_id, proxy = hit
+        m = calc.as_proxy(m)
     p, provider_reason = kb.resolve_provider(provider, endpoint=endpoint, model=m)
     if region is not None:
         # the caller knows the site: it is pinned, so no candidates or widening
@@ -100,7 +111,19 @@ def estimate(
     )
     if provider_reason:
         est.reasons.insert(0, provider_reason)
+    if proxy:
+        est.reasons.insert(0, "model_inferred_from_sibling")
     out = est.to_dict()
+    out["proxy"] = None
+    if proxy:
+        # a sibling's size is a guess however narrow the range comes out
+        out["confidence"]["level"] = "low"
+        out["proxy"] = {
+            "model": proxy_id,
+            "estimated_from": m.id,
+            "note": proxy.note,
+            "source": proxy.source,
+        }
     out["requested"] = {
         "model": model,
         "provider": provider,

{ecocost-0.2.2 → ecocost-0.2.4}/ecocost/calc.py

@@ -17,6 +17,7 @@ carbon range and explained by listing every tier-3 input that went into it.
 
 from __future__ import annotations
 
+import dataclasses
 from collections.abc import Sequence
 from dataclasses import dataclass
 
@@ -252,6 +253,25 @@ class Estimate:
         }
 
 
+def as_proxy(model: Model) -> Model:
+    """The record a proxy id is estimated from, made honest for other
+    weights: the size becomes an assumption, widened to at least the tier-3
+    half-width either way, and measured energy is dropped because it
+    measured this record's weights, not the proxy's."""
+    f = model.active_params_b
+    half_width = DEFAULT_HALF_WIDTH[Tier.assumed]
+    return dataclasses.replace(
+        model,
+        active_params_b=dataclasses.replace(
+            f,
+            tier=Tier.assumed,
+            min=min(f.min, f.value / half_width),
+            max=max(f.max, f.value * half_width),
+        ),
+        measured_wh_per_1k_output_tokens_h100=None,
+    )
+
+
 def _h100_seconds(
     model: Model, reference: Hardware, tokens: float, utilization: Range
 ) -> Range:

{ecocost-0.2.2 → ecocost-0.2.4}/ecocost/data/models.yaml

@@ -286,6 +286,30 @@ models:
       stale_after: '2027-03-31'
       sources:
       - resource: https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Thinking
+  qwen3-vl-235b-a22b-instruct:
+    label: Qwen3 VL 235B-A22B Instruct
+    creator: Qwen
+    open_weights: true
+    total_params_b: 235
+    aliases:
+    - accounts/fireworks/models/qwen3-vl-235b-a22b-instruct
+    - qwen/qwen3-vl-235b-a22b-instruct
+    active_params_b:
+      value: 22
+      unit: B
+      tier: 1
+      source: https://arxiv.org/html/2511.21631#S2
+      note: MoE, 235B total with 22B activated per token, per the Qwen3-VL Technical Report (Qwen Team, 2025), section 2; config.json has
+        128 experts, 8 per token.
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://arxiv.org/html/2511.21631#S2
+      - resource: https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct
+      - resource: https://fireworks.ai/models/fireworks/qwen3-vl-235b-a22b-instruct
   glm-5.3:
     label: GLM-5.3
     creator: Z.ai
@@ -331,6 +355,82 @@ models:
       stale_after: '2027-03-31'
       sources:
       - resource: https://huggingface.co/zai-org/GLM-5.3-Flash
+  glm-5.2:
+    label: GLM-5.2
+    creator: Z.ai
+    open_weights: true
+    total_params_b: 753
+    aliases:
+    - accounts/fireworks/models/glm-5p2
+    - z-ai/glm-5.2
+    - zai-org/glm-5.2
+    active_params_b:
+      value: 39.4
+      unit: B
+      tier: 2
+      source: https://huggingface.co/zai-org/GLM-5.2/blob/main/config.json
+      note: "Card states no size. Computed from config.json (glm_moe_dsa: 78 layers, first 3 dense; 256 routed + 1 shared expert, 8 routed per token; MLA; DSA indexer computed on the first 3 layers, then 1 of every 4): 39.4B active, excluding embeddings, LM head and the MTP layer. The same method gives 39.9B for GLM-5's config, which its card states as 40B active, and 32.1B for GLM-4.5's, stated as 32B. total_params_b is the HF safetensors count incl. the MTP layer (753.3B; 743B without it)."
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://huggingface.co/zai-org/GLM-5.2
+      - resource: https://huggingface.co/zai-org/GLM-5.2/blob/main/config.json
+      - resource: https://huggingface.co/zai-org/GLM-5
+      - resource: https://huggingface.co/zai-org/GLM-4.5
+      - resource: https://fireworks.ai/models/fireworks/glm-5p2
+  glm-5.1:
+    label: GLM-5.1
+    creator: Z.ai
+    open_weights: true
+    total_params_b: 754
+    aliases:
+    - accounts/fireworks/models/glm-5p1
+    - z-ai/glm-5.1
+    - zai-org/glm-5.1
+    active_params_b:
+      value: 39.9
+      unit: B
+      tier: 2
+      source: https://huggingface.co/zai-org/GLM-5.1/blob/main/config.json
+      note: "Card states no size. Computed from config.json, which matches GLM-5's except transformers_version (glm_moe_dsa: 78 layers, first 3 dense; 256 routed + 1 shared expert, 8 routed per token; MLA + DSA indexer): 39.9B active, excluding embeddings, LM head and the MTP layer; the GLM-5 card states 40B active for this architecture. total_params_b is the HF safetensors count incl. the MTP layer (753.9B; 744B without it). Fireworks serves Z.ai's FP8 checkpoint, zai-org/GLM-5.1-FP8."
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://huggingface.co/zai-org/GLM-5.1
+      - resource: https://huggingface.co/zai-org/GLM-5.1/blob/main/config.json
+      - resource: https://huggingface.co/zai-org/GLM-5
+      - resource: https://fireworks.ai/models/fireworks/glm-5p1
+  glm-4.7:
+    label: GLM-4.7
+    creator: Z.ai
+    open_weights: true
+    total_params_b: 358
+    aliases:
+    - accounts/fireworks/models/glm-4p7
+    - z-ai/glm-4.7
+    - zai-org/glm-4.7
+    active_params_b:
+      value: 32.1
+      unit: B
+      tier: 2
+      source: https://huggingface.co/zai-org/GLM-4.7/blob/main/config.json
+      note: "Card states no size. Computed from config.json, which matches GLM-4.5's except max_position_embeddings (glm4_moe: 92 layers, first 3 dense; 160 routed + 1 shared expert, 8 routed per token; GQA 96/8 heads): 32.1B active, excluding embeddings, LM head and the MTP layer; the GLM-4.5 card states 32B active for this architecture. total_params_b is the HF safetensors count incl. the MTP layer (358.3B; 355B per the GLM-4.5 card)."
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://huggingface.co/zai-org/GLM-4.7
+      - resource: https://huggingface.co/zai-org/GLM-4.7/blob/main/config.json
+      - resource: https://huggingface.co/zai-org/GLM-4.5
+      - resource: https://fireworks.ai/models/fireworks/glm-4p7
   mistral-medium-3-5:
     label: Mistral Medium 3.5
     creator: Mistral
@@ -438,6 +538,31 @@ models:
       stale_after: '2027-03-31'
       sources:
       - resource: https://docs.mistral.ai/models/mistral-small-4-0-26-03
+  mixtral-8x7b-instruct-v0.1:
+    label: Mixtral 8x7B Instruct
+    creator: Mistral
+    open_weights: true
+    total_params_b: 46.7
+    aliases:
+    - accounts/fireworks/models/mixtral-8x7b-instruct
+    - mistralai/mixtral-8x7b-instruct
+    - mistralai/mixtral-8x7b-instruct-v0.1
+    - mixtral-8x7b-32768
+    active_params_b:
+      value: 12.9
+      unit: B
+      tier: 1
+      source: https://mistral.ai/news/mixtral-of-experts
+      note: Sparse MoE, 2 of 8 experts per token; 46.7B total, 12.9B used per token per Mistral's release post (2023-12-11).
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://mistral.ai/news/mixtral-of-experts
+      - resource: https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1
+      - resource: https://fireworks.ai/models/fireworks/mixtral-8x7b-instruct
   sarvam-105b:
     label: Sarvam 105B
     creator: Sarvam AI
@@ -458,6 +583,28 @@ models:
       stale_after: '2027-03-31'
       sources:
       - resource: https://huggingface.co/sarvamai/sarvam-105b
+  sarvam-m:
+    label: Sarvam-M
+    creator: Sarvam AI
+    open_weights: true
+    total_params_b: 23.6
+    aliases:
+    - sarvamai/sarvam-m
+    active_params_b:
+      value: 23.6
+      unit: B
+      tier: 1
+      source: https://huggingface.co/sarvamai/sarvam-m
+      note: Dense (MistralForCausalLM), all parameters active; 23,572,403,200 BF16 parameters per the repo's safetensors metadata, a
+        text-only post-train of Mistral-Small-3.1-24B (no vision encoder).
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://huggingface.co/sarvamai/sarvam-m
+      - resource: https://huggingface.co/sarvamai/sarvam-m/blob/main/config.json
   sunflower-14b:
     label: Sunflower 14B
     creator: Sunbird AI
@@ -683,6 +830,81 @@ models:
       stale_after: '2027-03-31'
       sources:
       - resource: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro
+  deepseek-v3.2:
+    label: DeepSeek V3.2
+    creator: DeepSeek
+    open_weights: true
+    total_params_b: 671
+    aliases:
+    - accounts/fireworks/models/deepseek-v3p2
+    - deepseek-ai/DeepSeek-V3.2
+    - deepseek/deepseek-v3.2
+    active_params_b:
+      value: 37
+      unit: B
+      tier: 2
+      source: https://huggingface.co/deepseek-ai/DeepSeek-V3.1
+      note: MoE. The V3.2 card and report give no activated count; report sec. 2.1 (https://arxiv.org/html/2512.02556v1#S2.SS1)
+        says V3.2 keeps the V3.1-Terminus architecture except for DeepSeek Sparse Attention (a small lightning indexer).
+        37B is the V3.1 card's activated figure, so derived; the indexer adds roughly 1B active, so it slightly undercounts.
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://arxiv.org/html/2512.02556v1#S2.SS1
+      - resource: https://arxiv.org/abs/2512.02556
+      - resource: https://huggingface.co/deepseek-ai/DeepSeek-V3.2
+      - resource: https://huggingface.co/deepseek-ai/DeepSeek-V3.1
+      - resource: https://fireworks.ai/models/fireworks/deepseek-v3p2
+  deepseek-v3.1:
+    label: DeepSeek V3.1
+    creator: DeepSeek
+    open_weights: true
+    total_params_b: 671
+    aliases:
+    - accounts/fireworks/models/deepseek-v3p1
+    - deepseek-ai/DeepSeek-V3.1
+    - deepseek/deepseek-chat-v3.1
+    active_params_b:
+      value: 37
+      unit: B
+      tier: 1
+      source: https://huggingface.co/deepseek-ai/DeepSeek-V3.1
+      note: MoE, 671B total / 37B activated per the model card's download table.
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://huggingface.co/deepseek-ai/DeepSeek-V3.1
+      - resource: https://fireworks.ai/models/fireworks/deepseek-v3p1
+  deepseek-r1:
+    label: DeepSeek R1
+    creator: DeepSeek
+    open_weights: true
+    total_params_b: 671
+    aliases:
+    - accounts/fireworks/models/deepseek-r1
+    - deepseek-ai/DeepSeek-R1
+    - deepseek/deepseek-r1
+    active_params_b:
+      value: 37
+      unit: B
+      tier: 1
+      source: https://huggingface.co/deepseek-ai/DeepSeek-R1
+      note: MoE, 671B total / 37B activated per the model card. Original January 2025 R1 weights, not
+        R1-0528.
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://huggingface.co/deepseek-ai/DeepSeek-R1
+      - resource: https://fireworks.ai/models/fireworks/deepseek-r1
   nemotron-3.5-lightning-30b-a3b:
     label: Nemotron 3.5 Lightning 30B-A3B
     creator: Nvidia
@@ -751,6 +973,74 @@ models:
       stale_after: '2027-03-31'
       sources:
       - resource: https://huggingface.co/moonshotai/Kimi-K3
+  kimi-k2.6:
+    label: Kimi K2.6
+    creator: Moonshot AI
+    open_weights: true
+    total_params_b: 1000
+    aliases:
+    - accounts/fireworks/models/kimi-k2p6
+    - moonshotai/kimi-k2.6
+    active_params_b:
+      value: 32
+      unit: B
+      tier: 1
+      source: https://huggingface.co/moonshotai/Kimi-K2.6
+      note: MoE, 8 of 384 routed + 1 shared; 1T total, plus a 400M MoonViT vision encoder. Native INT4.
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://huggingface.co/moonshotai/Kimi-K2.6
+      - resource: https://fireworks.ai/models/fireworks/kimi-k2p6
+  kimi-k2-thinking:
+    label: Kimi K2 Thinking
+    creator: Moonshot AI
+    open_weights: true
+    total_params_b: 1000
+    aliases:
+    - accounts/fireworks/models/kimi-k2-thinking
+    - fireworks/kimi-k2-thinking
+    - moonshotai/kimi-k2-thinking
+    active_params_b:
+      value: 32
+      unit: B
+      tier: 1
+      source: https://huggingface.co/moonshotai/Kimi-K2-Thinking
+      note: MoE, 8 of 384 routed + 1 shared; 1T total. Native INT4 (QAT) weights for the MoE layers.
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://huggingface.co/moonshotai/Kimi-K2-Thinking
+      - resource: https://fireworks.ai/models/fireworks/kimi-k2-thinking
+  kimi-k2-instruct-0905:
+    label: Kimi K2 Instruct 0905
+    creator: Moonshot AI
+    open_weights: true
+    total_params_b: 1000
+    aliases:
+    - accounts/fireworks/models/kimi-k2-instruct-0905
+    - moonshotai/kimi-k2-0905
+    - moonshotai/kimi-k2-instruct-0905
+    active_params_b:
+      value: 32
+      unit: B
+      tier: 1
+      source: https://huggingface.co/moonshotai/Kimi-K2-Instruct-0905
+      note: MoE, 8 of 384 routed + 1 shared; 1T total. September 2025 update of Kimi K2 Instruct.
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://huggingface.co/moonshotai/Kimi-K2-Instruct-0905
+      - resource: https://fireworks.ai/models/fireworks/kimi-k2-instruct-0905
   inkling:
     label: Inkling
     creator: Thinking Machines Lab
@@ -795,12 +1085,97 @@ models:
       stale_after: '2027-03-31'
       sources:
       - resource: https://huggingface.co/MiniMaxAI/MiniMax-M3
+  minimax-m3.1-flash:
+    label: MiniMax M3.1 Flash
+    creator: MiniMax
+    open_weights: true
+    total_params_b: 428
+    aliases:
+    - accounts/fireworks/models/minimax-m3p1-flash
+    - minimax/minimax-m3.1-flash
+    active_params_b:
+      value: 23
+      unit: B
+      tier: 1
+      source: https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash#model-at-a-glance
+      note: MoE, ~428B total, 23B activated per token per model card (Model at a Glance). Sparse attention throughout, quantization-aware training.
+    provenance:
+      generated: {by: 'agent:claude-coverage-routine', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash
+  minimax-m2.7:
+    label: MiniMax M2.7
+    creator: MiniMax
+    open_weights: true
+    total_params_b: 229.9
+    aliases:
+    - accounts/fireworks/models/minimax-m2p7
+    - fireworks/minimax-m2p7
+    - MiniMaxAI/MiniMax-M2.7
+    - minimax/minimax-m2.7
+    active_params_b:
+      value: 9.8
+      unit: B
+      tier: 2
+      source: https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/main/config.json
+      note: "MoE, 256 experts, 8 per token. The card states no size. Computed from config.json, architecturally identical to MiniMax-M2 (same layers, dims, experts): 9.80B non-embedding activated. The MiniMax-M2 series tech report (https://arxiv.org/html/2605.26494v2#S2.SS1) gives 229.9B total / 9.8B activated for M2 and says M2.7 runs on ~10B activated parameters."
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://arxiv.org/html/2605.26494v2#S2.SS1
+      - resource: https://arxiv.org/abs/2605.26494
+      - resource: https://huggingface.co/MiniMaxAI/MiniMax-M2.7
+      - resource: https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/main/config.json
+      - resource: https://fireworks.ai/models/fireworks/minimax-m2p7
+  minimax-m2.1:
+    label: MiniMax M2.1
+    creator: MiniMax
+    open_weights: true
+    total_params_b: 229.9
+    aliases:
+    - accounts/fireworks/models/minimax-m2p1
+    - MiniMaxAI/MiniMax-M2.1
+    - minimax/minimax-m2.1
+    active_params_b:
+      value: 9.8
+      unit: B
+      tier: 2
+      source: https://huggingface.co/MiniMaxAI/MiniMax-M2.1/blob/main/config.json
+      note: MoE. The M2.1 card states no size. Computed from config.json (62 layers, hidden 3072, 256
+        experts, 8 per token, expert FFN 1536, GQA 48/8 heads) gives 9.80B non-embedding activated. Config is
+        architecturally identical to MiniMax-M2 (same layers, dims, experts), whose tech report gives 9.8B activated / 229.9B total
+        (https://arxiv.org/html/2605.26494v2#S2.SS1).
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://huggingface.co/MiniMaxAI/MiniMax-M2.1/blob/main/config.json
+      - resource: https://huggingface.co/MiniMaxAI/MiniMax-M2.1
+      - resource: https://huggingface.co/MiniMaxAI/MiniMax-M2/blob/main/config.json
+      - resource: https://arxiv.org/html/2605.26494v2#S2.SS1
+      - resource: https://fireworks.ai/models/fireworks/minimax-m2p1
   gpt-4o:
     label: GPT-4o
     creator: OpenAI
     open_weights: false
     aliases:
     - gpt-4o-2024-08-06
+    - openai-gpt-4o-prod-eastus2-1  # an Azure OpenAI deployment name
+    proxies:
+    - ids: [chatgpt-4o-latest]
+      note: "OpenAI: a model alias for the GPT-4o snapshot used in ChatGPT; deprecated, no size of its own."
+      source: https://developers.openai.com/api/docs/models/chatgpt-4o-latest
+    - ids: [gpt-4o-realtime-preview, gpt-4o-realtime-preview-2025-06-03, gpt-4o-realtime-preview-2024-12-17, gpt-4o-realtime-preview-2024-10-01]
+      note: "OpenAI: preview release of the GPT-4o Realtime model (audio and text); deprecated, no size of its own."
+      source: https://developers.openai.com/api/docs/models/gpt-4o-realtime-preview
     active_params_b:
       value: 100
       unit: B
@@ -822,6 +1197,7 @@ models:
     open_weights: false
     aliases:
     - gpt-4o-mini-2024-07-18
+    - openai-gpt-4o-mini-prod-eastus2-1  # an Azure OpenAI deployment name
     active_params_b:
       value: 12
       unit: B
@@ -837,6 +1213,84 @@ models:
       stale_after: '2027-03-31'
       sources:
       - resource: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+  gpt-4-turbo:
+    label: GPT-4 Turbo
+    creator: OpenAI
+    total_params_b: 880
+    open_weights: false
+    aliases:
+    - gpt-4-turbo-2024-04-09
+    - gpt-4-turbo-preview
+    - gpt-4-0125-preview
+    - gpt-4-1106-vision-preview
+    - gpt-4-vision-preview
+    - openai-gpt-4-turbo-prod-ca-1  # an Azure OpenAI deployment name
+    active_params_b:
+      value: 152
+      unit: B
+      tier: 2
+      min: 88
+      max: 264
+      source: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+      note: "Undisclosed. EcoLogits (gpt-4-turbo and gpt-4-turbo-2024-04-09): MoE 880B total, 88-264B active, flagged model-arch-not-released and model-arch-multimodal. Point = geometric mean. Aliases are the snapshots OpenAI lists under GPT-4 Turbo, plus gpt-4-vision-preview, inferred to be the alias of gpt-4-1106-vision-preview: OpenAI's deprecations page lists the two together at the same price and shutdown date. Shut down 2026-10-23."
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+      - resource: https://developers.openai.com/api/docs/models/gpt-4-turbo
+      - resource: https://developers.openai.com/api/docs/deprecations
+  gpt-4:
+    label: GPT-4
+    creator: OpenAI
+    total_params_b: 1760
+    open_weights: false
+    aliases:
+    - gpt-4-0613
+    - gpt-4-0314
+    - openai-gpt-4-prod-ca-1  # an Azure OpenAI deployment name
+    active_params_b:
+      value: 305
+      unit: B
+      tier: 2
+      min: 176
+      max: 528
+      source: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+      note: "Undisclosed. EcoLogits (gpt-4 and gpt-4-0613): MoE 1760B total, 176-528B active, flagged model-arch-not-released and model-arch-multimodal. Point = geometric mean. Aliases are the snapshots OpenAI lists under GPT-4. Shut down 2026-10-23."
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+      - resource: https://developers.openai.com/api/docs/models/gpt-4
+  gpt-3.5-turbo:
+    label: GPT-3.5 Turbo
+    creator: OpenAI
+    open_weights: false
+    aliases:
+    - gpt-3.5-turbo-0125
+    - gpt-3.5-turbo-1106
+    - openai-gpt-35-turbo-prod-ca-1  # an Azure OpenAI deployment name
+    active_params_b:
+      value: 37
+      unit: B
+      tier: 2
+      min: 20
+      max: 70
+      source: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+      note: "Undisclosed. EcoLogits (gpt-3.5-turbo, and gpt-3.5-turbo-0125 with the same figures): dense, 20-70B, flagged model-arch-not-released. Point = geometric mean. Aliases are the snapshots OpenAI lists under GPT-3.5 Turbo."
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+      - resource: https://developers.openai.com/api/docs/models/gpt-3.5-turbo
   o4-mini:
     label: o4-mini
     creator: OpenAI
@@ -858,6 +1312,30 @@ models:
       stale_after: '2027-03-31'
       sources:
       - resource: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+  gpt-5.4:
+    label: GPT-5.4
+    creator: OpenAI
+    total_params_b: 450
+    open_weights: false
+    aliases:
+    - gpt-5.4-2026-03-05
+    - openai/gpt-5.4
+    active_params_b:
+      value: 82
+      unit: B
+      tier: 2
+      min: 45
+      max: 150
+      source: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+      note: "Undisclosed. EcoLogits (gpt-5.4 and gpt-5.4-2026-03-05): MoE 450B total, 45-150B active, flagged model-arch-not-released and model-arch-multimodal. gpt-5.4-2026-03-05 is the snapshot OpenAI lists for GPT-5.4. Point = geometric mean of that range. Tier 2: a published estimate, not a disclosure."
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+      - resource: https://developers.openai.com/api/docs/models/gpt-5.4
   gpt-5.4-mini:
     label: GPT-5.4 mini
     creator: OpenAI
@@ -1120,6 +1598,10 @@ models:
     creator: OpenAI
     total_params_b: 440
     open_weights: false
+    proxies:
+    - ids: [gpt-realtime, gpt-realtime-2025-08-28]
+      note: "OpenAI's first GA realtime model, which gpt-realtime-1.5 replaced; same prices ($4/$16 text, $32/$64 audio per 1M tokens), no size of its own."
+      source: https://developers.openai.com/api/docs/models/gpt-realtime
     active_params_b:
       value: 76
       unit: B
@@ -1375,6 +1857,33 @@ models:
       stale_after: '2027-03-31'
       sources:
       - resource: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+  gemini-4-argon:
+    label: Gemini 4 Argon
+    creator: Google
+    open_weights: false
+    aliases:
+    - google/gemini-4-argon
+    - gemini/gemini-4-argon
+    active_params_b:
+      value: 268
+      unit: B
+      tier: 3
+      min: 120
+      max: 600
+      source: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+      note: "Undisclosed; model card: 'a large scale sparse mixture-of-experts (MoE)', 'not a modification or a fine-tune of a prior model', no size or serving hardware. No published size estimate as of 2026-10-18: EcoLogits (latest models.json commit 2026-10-14) has no Gemini 4 entry, and Epoch AI lists parameters as Unknown. Proxy from Google's previous flagship Pro-tier models in EcoLogits: gemini-3.1-pro-preview (MoE 1200B total, 120-360B active; Google's migration guide moves 3.1 Pro users to 4 Argon) and gemini-2.5-pro (MoE 2000B total, 200-600B active). Range spans both (120-600B); point = geometric mean (268). Tier 3: a proxy from sibling models, not an estimate of these weights."
+    provenance:
+      generated: {by: 'agent:claude-coverage-routine', at: '2026-10-18'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-18'
+      sources:
+      - resource: https://deepmind.google/models/model-cards/gemini-4-argon/
+      - resource: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+      - resource: https://ai.google.dev/gemini-api/docs/models/gemini-4-argon
+      - resource: https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/4-argon
+      - resource: https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/guides/gemini-4-argon#migration-guide
+      - resource: https://epoch.ai/models/gemini-4-argon
   gemini-3.8-flash:
     label: Gemini 3.8 Flash
     creator: Google
@@ -1383,8 +1892,6 @@ models:
     aliases:
     - google/gemini-3.8-flash
     - gemini/gemini-3.8-flash
-    - google/gemini-3.7-flash
-    - gemini/gemini-3.7-flash
     active_params_b:
       value: 122
       unit: B
@@ -1400,6 +1907,31 @@ models:
       stale_after: '2027-03-31'
       sources:
       - resource: https://deepmind.google/models/model-cards/gemini-3-8-flash/
+  gemini-3.7-flash:
+    label: Gemini 3.7 Flash
+    creator: Google
+    total_params_b: 750
+    open_weights: false
+    aliases:
+    - google/gemini-3.7-flash
+    - gemini/gemini-3.7-flash
+    active_params_b:
+      value: 122
+      unit: B
+      tier: 2
+      min: 75
+      max: 200
+      source: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+      note: "Undisclosed; model card: 'based on Gemini 3.6 Flash', no size or serving hardware. EcoLogits (gemini-3.7-flash): MoE 750B total, 75-200B active, flagged model-arch-not-released, the same estimate as gemini-3.6-flash and gemini-3.8-flash. Point = geometric mean of that range. Tier 2: a published estimate, not a disclosure. Deprecated in the Gemini API; requests to gemini-3.7-flash are now routed to gemini-3.8-flash, which has the same estimate."
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://deepmind.google/models/model-cards/gemini-3-7-flash/
+      - resource: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+      - resource: https://ai.google.dev/gemini-api/docs/models/gemini-3.7-flash
   gemini-3.6-flash:
     label: Gemini 3.6 Flash
     creator: Google
@@ -1500,6 +2032,101 @@ models:
       stale_after: '2027-03-31'
       sources:
       - resource: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+  gemini-3-flash-preview:
+    label: Gemini 3 Flash preview
+    creator: Google
+    total_params_b: 300
+    open_weights: false
+    aliases:
+    - google/gemini-3-flash-preview
+    - gemini/gemini-3-flash-preview
+    active_params_b:
+      value: 173
+      unit: B
+      tier: 2
+      min: 100
+      max: 300
+      source: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+      note: "Undisclosed. EcoLogits (gemini-3-flash-preview): MoE 300B total, 100-300B active, flagged model-arch-not-released and model-arch-multimodal. Point = geometric mean of that range. Tier 2: a published estimate, not a disclosure."
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+      - resource: https://ai.google.dev/gemini-api/docs/models/gemini-3-flash-preview
+  gemini-3.1-flash-lite-preview:
+    label: Gemini 3.1 Flash-Lite preview
+    creator: Google
+    open_weights: false
+    aliases:
+    - google/gemini-3.1-flash-lite-preview
+    - gemini/gemini-3.1-flash-lite-preview
+    active_params_b:
+      value: 56
+      unit: B
+      tier: 2
+      min: 30
+      max: 105
+      source: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+      note: "Undisclosed. EcoLogits (gemini-3.1-flash-lite-preview): dense, 30-105B, flagged model-arch-not-released and model-arch-multimodal; the same estimate as gemini-3.1-flash-lite. A separate record because the preview and GA weights are not documented as identical. Shut down in the Gemini API. Point = geometric mean of that range. Tier 2: a published estimate, not a disclosure."
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+      - resource: https://ai.google.dev/gemini-api/docs/models/gemini-3.1-flash-lite-preview
+  gemini-2.5-pro:
+    label: Gemini 2.5 Pro
+    creator: Google
+    total_params_b: 2000
+    open_weights: false
+    aliases:
+    - google/gemini-2.5-pro
+    - gemini/gemini-2.5-pro
+    active_params_b:
+      value: 346
+      unit: B
+      tier: 2
+      min: 200
+      max: 600
+      source: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+      note: "Undisclosed. EcoLogits (gemini-2.5-pro): MoE 2000B total, 200-600B active, flagged model-arch-not-released and model-arch-multimodal. Point = geometric mean of that range. Tier 2: a published estimate, not a disclosure."
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+      - resource: https://ai.google.dev/gemini-api/docs/models/gemini-2.5-pro
+  gemini-2.5-flash:
+    label: Gemini 2.5 Flash
+    creator: Google
+    total_params_b: 440
+    open_weights: false
+    aliases:
+    - google/gemini-2.5-flash
+    - gemini/gemini-2.5-flash
+    active_params_b:
+      value: 76
+      unit: B
+      tier: 2
+      min: 44
+      max: 132
+      source: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+      note: "Undisclosed. EcoLogits (gemini-2.5-flash): MoE 440B total, 44-132B active, flagged model-arch-not-released and model-arch-multimodal. Point = geometric mean of that range. Tier 2: a published estimate, not a disclosure."
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+      - resource: https://ai.google.dev/gemini-api/docs/models/gemini-2.5-flash
   gemini-3.1-pro-preview:
     label: Gemini 3.1 Pro
     creator: Google
@@ -1507,6 +2134,10 @@ models:
     total_params_b: 1200
     aliases:
     - google/gemini-3.1-pro-preview
+    proxies:
+    - ids: [gemini-3-pro-preview]
+      note: "Google's model card: 'Gemini 3.1 Pro is based on Gemini 3 Pro' and defers to it for architecture; no EcoLogits estimate of 3 Pro itself."
+      source: https://deepmind.google/models/model-cards/gemini-3-1-pro/
     active_params_b:
       value: 200
       unit: B
@@ -1541,6 +2172,30 @@ models:
       stale_after: '2027-03-31'
       sources:
       - resource: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
+  muse-wonder-1.0:
+    label: Muse Wonder
+    creator: Meta AI
+    open_weights: false
+    aliases:
+    - meta/muse-wonder
+    active_params_b:
+      value: 164
+      unit: B
+      tier: 3
+      source: https://artificialanalysis.ai/models/muse-wonder
+      note: "Undisclosed ('Meta has not disclosed the model size or parameter count'); released 2026-10-12; no EcoLogits entry as of 2026-10-18. Proxy gpt-5.6-sol: same list price, $4/$20 per 1M tokens (promotional through at least 2026-11-21); EcoLogits MoE 900B total, 90-300B active. Point = rounded geometric mean of that range."
+      min: 90
+      max: 300
+    provenance:
+      generated: {by: 'agent:claude-coverage-routine', at: '2026-10-18'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-18'
+      sources:
+      - resource: https://artificialanalysis.ai/models/muse-wonder
+      - resource: https://dev.meta.ai/models/muse-wonder
+      - resource: https://developers.openai.com/api/docs/models/gpt-5.6-sol
+      - resource: https://raw.githubusercontent.com/genai-impact/ecologits/main/ecologits/data/models.json
   muse-spark-1.3:
     label: Muse Spark 1.3
     creator: Meta AI
@@ -1802,3 +2457,152 @@ models:
       stale_after: '2027-03-31'
       sources:
       - resource: https://huggingface.co/meta-llama/Llama-3.1-405B-Instruct
+
+  llama-4-maverick:
+    label: Llama 4 Maverick 17B-128E Instruct
+    creator: Meta
+    open_weights: true
+    total_params_b: 400
+    aliases:
+    - accounts/fireworks/models/llama4-maverick-instruct-basic
+    - meta-llama/Llama-4-Maverick-17B-128E-Instruct
+    - meta-llama/llama-4-maverick
+    active_params_b:
+      value: 17
+      unit: B
+      tier: 1
+      source: https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct
+      note: MoE, 128 experts; 17B activated, 400B total per the model card's model-information table. The safetensors
+        metadata counts 401,583,781,376 BF16 parameters (includes the vision encoder).
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct
+  llama-4-scout:
+    label: Llama 4 Scout 17B-16E Instruct
+    creator: Meta
+    open_weights: true
+    total_params_b: 109
+    aliases:
+    - accounts/fireworks/models/llama4-scout-instruct-basic
+    - meta-llama/Llama-4-Scout-17B-16E-Instruct
+    - meta-llama/llama-4-scout
+    active_params_b:
+      value: 17
+      unit: B
+      tier: 1
+      source: https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct
+      note: MoE, 16 experts; 17B activated, 109B total per the model card's model-information table. The safetensors
+        metadata counts 108,641,793,536 BF16 parameters (includes the vision encoder).
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct
+      - resource: https://raw.githubusercontent.com/ml-energy/leaderboard/master/public/data/tasks/lm-arena-chat.json
+    measured_wh_per_1k_output_tokens_h100:
+      value: 0.3212
+      min: 0.0934
+      max: 1.105
+      unit: Wh per 1k output tokens, H100, GPU-only, PUE 1.0
+      tier: 1
+      source: https://raw.githubusercontent.com/ml-energy/leaderboard/master/public/data/tasks/lm-arena-chat.json
+      note: ML.ENERGY v3.0, lm-arena-chat, meta-llama/Llama-4-Scout-17B-16E-Instruct, bfloat16, 4xH100 EP4 (an 8xH100 EP8 sweep, 0.126-1.465, is not used).
+        min = max-batch steady state (max_num_seqs 512), max = batch-8; value = geometric mean; production
+        sits between.
+  llama-3-70b:
+    label: Llama 3 70B Instruct
+    creator: Meta
+    open_weights: true
+    total_params_b: 70.55
+    aliases:
+    - accounts/fireworks/models/llama-v3-70b-instruct
+    - llama3-70b-8192
+    - meta-llama/llama-3-70b-instruct
+    - meta-llama/Meta-Llama-3-70B-Instruct
+    active_params_b:
+      value: 70.55
+      unit: B
+      tier: 1
+      source: https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct
+      note: Dense, 8k context. 70,553,706,496 BF16 parameters per the model card's safetensors metadata. The original
+        Llama 3 (April 2024), not Llama 3.1 or 3.3.
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct
+  llama-3-8b:
+    label: Llama 3 8B Instruct
+    creator: Meta
+    open_weights: true
+    total_params_b: 8.03
+    aliases:
+    - llama3-8b-8192
+    - meta-llama/llama-3-8b-instruct
+    - meta-llama/Meta-Llama-3-8B-Instruct
+    active_params_b:
+      value: 8.03
+      unit: B
+      tier: 1
+      source: https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct
+      note: Dense, 8k context. 8,030,261,248 BF16 parameters per the model card's safetensors metadata. The original
+        Llama 3 (April 2024), not Llama 3.1.
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct
+  llama-3-groq-70b-tool-use:
+    label: Llama 3 Groq 70B Tool Use
+    creator: Groq
+    open_weights: true
+    total_params_b: 70.55
+    aliases:
+    - Groq/Llama-3-Groq-70B-Tool-Use
+    - llama3-groq-70b-8192-tool-use-preview
+    active_params_b:
+      value: 70.55
+      unit: B
+      tier: 1
+      source: https://huggingface.co/Groq/Llama-3-Groq-70B-Tool-Use
+      note: Dense. The card calls it the 70B version, fully fine-tuned (plus DPO) from Llama 3 70B; 70,553,804,800
+        BF16 parameters per the safetensors metadata (a few added tokens over the base). Fine-tune, not the
+        base weights.
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://huggingface.co/Groq/Llama-3-Groq-70B-Tool-Use
+  afrollama_v1:
+    label: AfroLlama V1
+    creator: Jacaranda Health
+    open_weights: true
+    total_params_b: 8.03
+    aliases:
+    - Jacaranda/AfroLlama_V1
+    active_params_b:
+      value: 8.03
+      unit: B
+      tier: 1
+      source: https://huggingface.co/Jacaranda/AfroLlama_V1
+      note: Dense. The card states an 8B-parameter model built on Meta-Llama-3-8B (LoRA continual pre-training per
+        language, TIES merge, instruction tuning); 8,030,261,248 BF16 parameters per the safetensors metadata.
+    provenance:
+      generated: {by: 'agent:claude-model-research', at: '2026-10-27'}
+      verified: []
+      status: draft
+      stale_after: '2027-04-27'
+      sources:
+      - resource: https://huggingface.co/Jacaranda/AfroLlama_V1

{ecocost-0.2.2 → ecocost-0.2.4}/ecocost/data/sources.yaml

@@ -159,6 +159,7 @@ sources:
       - novita.ai
       - fal.ai
       - developer.meta.com
+      - dev.meta.ai
       - nscale.com
       - thinkingmachines.ai
       - sec.gov

{ecocost-0.2.2 → ecocost-0.2.4}/ecocost/loader.py

@@ -28,6 +28,7 @@ from .schema import (
     Model,
     Provenance,
     Provider,
+    Proxy,
     Region,
     Tier,
     Verification,
@@ -102,6 +103,11 @@ class KnowledgeBase:
         self.models: dict[str, Model] = {}
         # lower-cased id or alias -> canonical id, for case-insensitive lookup
         self._model_index: dict[str, str] = {}
+        # last path segment of a prefixed alias -> canonical id, or None when
+        # aliases of different models share it
+        self._model_tail_index: dict[str, str | None] = {}
+        # lower-cased proxy id -> (canonical id it is estimated from, proxy)
+        self._proxy_index: dict[str, tuple[str, Proxy]] = {}
         self._provider_index: dict[str, str] = {}
         self._region_index: dict[str, str] = {}
         self.fallback_provider_id: str | None = None
@@ -172,6 +178,7 @@ class KnowledgeBase:
                 kb._index_host(host, pid)
         mdoc = _read("models.yaml")
         kb.first_party_provider = mdoc.get("first_party_provider") or {}
+        proxies: list[tuple[str, Proxy]] = []
         for mid, raw in mdoc.get("models", {}).items():
             measured = raw.get("measured_wh_per_1k_output_tokens_h100")
             kb.models[mid] = Model(
@@ -188,6 +195,12 @@ class KnowledgeBase:
             )
             for key in [mid, *(raw.get("aliases") or [])]:
                 kb._index_model(key, mid)
+            for proxy in raw.get("proxies") or []:
+                proxies.append((mid, _proxy(mid, proxy)))
+        # after every model, so a proxy can't shadow an id or alias
+        for mid, proxy in proxies:
+            for key in proxy.ids:
+                kb._index_proxy(key, mid, proxy)
         kb._validate()
         return kb
 
@@ -220,12 +233,25 @@ class KnowledgeBase:
         """The model record for an id or alias, ignoring case; raises
         UnknownModelError."""
         # the id as given, then without a prefix like accounts/fireworks/models/
-        for key in (model_id, model_id.rsplit("/", 1)[-1]):
+        tail = model_id.rsplit("/", 1)[-1]
+        for key in (model_id, tail):
             if mid := self._model_index.get(key.lower()):
                 return self.models[mid]
-        tail = model_id.rsplit("/", 1)[-1]
+        # the same id behind another prefix: routers/x vs models/x
+        if mid := self._model_tail_index.get(tail.lower()):
+            return self.models[mid]
         raise UnknownModelError(model_id, _similar(tail, self._model_index))
 
+    def resolve_proxy(self, model_id: str) -> tuple[Model, str, Proxy] | None:
+        """For an id listed under a record's ``proxies``: that record, the
+        proxy id as listed, and the proxy. None for any other id."""
+        for key in (model_id, model_id.rsplit("/", 1)[-1]):
+            if hit := self._proxy_index.get(key.lower()):
+                mid, proxy = hit
+                listed = next(i for i in proxy.ids if i.lower() == key.lower())
+                return self.models[mid], listed, proxy
+        return None
+
     def resolve_region(self, region_id: str) -> Region:
         """The grid region for an id from regions.yaml, ignoring case; raises
         UnknownRegionError."""
@@ -288,6 +314,40 @@ class KnowledgeBase:
         other = self._model_index.setdefault(key.lower(), mid)
         if other != mid:
             raise ValueError(f"model id or alias {key!r} is used by {other} and {mid}")
+        if "/" in key:
+            tail = key.rsplit("/", 1)[-1].lower()
+            if self._model_tail_index.setdefault(tail, mid) != mid:
+                self._model_tail_index[tail] = None
+
+    def _index_proxy(self, key: str, mid: str, proxy: Proxy):
+        # resolve_model tries the id without its prefix, then that segment of
+        # any prefixed alias, before proxies; either would win
+        tail = key.rsplit("/", 1)[-1].lower()
+        if (
+            key.lower() in self._model_index
+            or tail in self._model_index
+            or self._model_tail_index.get(tail)
+        ):
+            raise ValueError(f"proxy {key!r} of {mid} is already a model id or alias")
+        if hit := self._proxy_index.get(key.lower()):
+            raise ValueError(f"proxy {key!r} of {mid} is already listed by {hit[0]}")
+        self._proxy_index[key.lower()] = (mid, proxy)
+
+
+def _proxy(mid: str, raw: dict) -> Proxy:
+    ids = raw.get("ids") if isinstance(raw, dict) else None
+    if not (
+        isinstance(ids, list) and ids and all(isinstance(i, str) and i for i in ids)
+    ):
+        raise ValueError(f"model {mid}: a proxy needs `ids`, a list of model ids")
+    proxy = Proxy(
+        ids=tuple(ids),
+        note=raw.get("note", "") or "",
+        source=raw.get("source", "") or "",
+    )
+    if not (proxy.note and proxy.source.startswith(("https://", "http://"))):
+        raise ValueError(f"model {mid}: a proxy needs a note and a source URL")
+    return proxy
 
 
 def _similar(key: str, index: dict[str, str]) -> list[str]:

{ecocost-0.2.2 → ecocost-0.2.4}/ecocost/result.py

@@ -83,6 +83,13 @@ class Provenance(TypedDict):
     hardware: RecordProvenance
 
 
+class ProxyInfo(TypedDict):
+    model: str
+    estimated_from: str
+    note: str
+    source: str
+
+
 class Requested(TypedDict):
     model: str
     provider: str | None
@@ -103,4 +110,5 @@ class EstimateResult(TypedDict):
     confidence: Confidence
     grid: Grid
     provenance: Provenance
+    proxy: ProxyInfo | None
     requested: Requested

{ecocost-0.2.2 → ecocost-0.2.4}/ecocost/schema.py

@@ -318,6 +318,17 @@ class Provider:
     provenance: Provenance = field(default_factory=Provenance)
 
 
+@dataclass(frozen=True)
+class Proxy:
+    """Ids with no record of their own, estimated from a related model's
+    record: an earlier version or sibling the vendor ties to it. Unlike an
+    alias these are different weights, so the estimate is flagged."""
+
+    ids: tuple[str, ...]
+    note: str  # why this model stands in for them
+    source: str  # where the vendor ties them together
+
+
 @dataclass(frozen=True)
 class Model:
     id: str

{ecocost-0.2.2 → ecocost-0.2.4}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "ecocost"
-version = "0.2.2"
+version = "0.2.4"
 description = "Per-run carbon, energy and water estimates for AI inference, with computed confidence and sourced inputs"
 authors = ["Gooey.AI <devs@dara.network>"]
 license = "Apache-2.0"