@iowarp/clio-coder 0.3.0

package/src/domains/providers/models/local-models/clio-local-coding-targets.yaml

@@ -0,0 +1,993 @@
+# Sources:
+# https://huggingface.co/Qwen/Qwen3.6-35B-A3B
+# https://huggingface.co/Qwen/Qwen3.6-27B
+# https://huggingface.co/google/gemma-4-26B-A4B
+# https://huggingface.co/LilaRest/gemma-4-31B-it-NVFP4-turbo
+# https://huggingface.co/Jackrong/Gemopus-4-31B-it-GGUF
+# https://huggingface.co/Jackrong/Qwopus3.6-27B-v1-preview-GGUF
+# https://huggingface.co/Jackrong/Qwopus3.6-35B-A3B-Coder-MTP-GGUF
+# https://huggingface.co/nvidia/Nemotron-Cascade-2-30B-A3B
+# https://huggingface.co/unsloth/NVIDIA-Nemotron-3-Nano-Omni-30B-A3B-Reasoning-GGUF
+# https://huggingface.co/mradermacher/Qwen3.5-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-i1-GGUF
+# https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B-GGUF
+#
+# This knowledge base is intentionally narrow. It describes only the local
+# models curated for Clio's local coding workflows and leaves cloud GPT models
+# to pi-ai's native OpenAI and openai-codex catalogs.
+#
+# Thinking semantics for these local families: the chain-of-thought is emitted
+# by the model's chat template (Qwen-style <think> blocks, Gemma 4 thinking
+# template, Nemotron reasoning template). LM Studio's SDK and llama.cpp's
+# OpenAI-compatible surface do not expose a reliable "no thinking" flag, so
+# Clio cannot disable template-driven reasoning at the API layer. With
+# `--thinking off` Clio simply does not request reasoning; if the model emits
+# it anyway, the chain-of-thought is captured into a ThinkingContent block,
+# counted as reasoning tokens (separate from output tokens) on the receipt
+# and TUI footer, and surfaced to the user truthfully rather than hidden.
+- family: openai-gpt-oss
+  matchPatterns:
+    - openai/gpt-oss
+    - gpt-oss
+    - gpt-oss-20b
+    - gpt-oss-120b
+  capabilities:
+    chat: true
+    tools: true
+    toolCallFormat: openai
+    reasoning: true
+    thinkingFormat: harmony
+    structuredOutputs: json-schema
+    vision: false
+    audio: false
+    embeddings: false
+    rerank: false
+    fim: false
+    contextWindow: 131072
+    maxTokens: 32768
+  quirks:
+    sampling:
+      thinking:
+        temperature: 1
+        topP: 1
+        maxTokens: 32768
+      instruct:
+        temperature: 1
+        topP: 1
+        maxTokens: 8192
+    runtimePreference:
+      llamaCpp: "Use the OpenAI-compatible chat-completions surface with chat_template_kwargs.reasoning_effort set to low, medium, or high."
+      openaiCompat: "Preferred for Harmony reasoning-effort control when the local gateway exposes chat_template_kwargs."
+    thinking:
+      mechanism: effort-levels
+      effortByLevel:
+        low: low
+        medium: medium
+        high: high
+      guidance: |
+        GPT-OSS uses OpenAI Harmony formatting and supports low, medium, and
+        high reasoning effort. Clio coerces off/minimal to low, captures
+        non-final Harmony channels as ThinkingContent, and strips Harmony
+        control tokens from visible output.
+
+- family: agenticqwen-30b-a3b-i1
+  matchPatterns:
+    - agenticqwen-30b-a3b-i1
+    - agenticqwen-30b-a3b-i1-q4_k_m
+    - agenticqwen
+  capabilities:
+    chat: true
+    tools: true
+    toolCallFormat: qwen
+    reasoning: true
+    thinkingFormat: qwen-chat-template
+    structuredOutputs: json-schema
+    vision: false
+    audio: false
+    embeddings: false
+    rerank: false
+    fim: false
+    contextWindow: 262144
+    maxTokens: 65536
+  quirks:
+    sampling:
+      thinking:
+        temperature: 0.5
+        topP: 0.9
+        topK: 20
+        maxTokens: 16384
+        reasoningBudget: 4096
+      instruct:
+        temperature: 0.5
+        topP: 0.9
+        topK: 20
+        maxTokens: 4096
+    gpuTiers:
+      "32gb": "Default 32 GB local llama.cpp orchestrator profile; benchmarked with roughly 1 GiB VRAM headroom on a 32 GiB class card."
+    runtimePreference:
+      llamaCpp: "OpenAI-compatible chat completions with --jinja. Recommended local profile uses ctx=262144, parallel=4, batch=2048, ubatch=512, and q8 KV cache."
+      openaiCompat: "Use only when a gateway fronts the same llama.cpp chat-completions surface."
+    llamaCpp:
+      ctxSize: 262144
+      cacheTypeK: q8_0
+      cacheTypeV: q8_0
+      parallel: 4
+      batchSize: 2048
+      ubatchSize: 512
+    thinking:
+      mechanism: budget-tokens
+      budgetByLevel:
+        low: 1024
+        medium: 4096
+        high: 16384
+      guidance: |
+        AgenticQwen is the tuned local default for Clio source work on
+        a local llama.cpp server. Thinking emits through the qwen chat template;
+        keep high reasoning for main-agent work and preserve visible thinking
+        blocks in receipts.
+    serving: "Use the llama.cpp settings listed above for a 262144-token local profile."
+
+- family: nemotron-3-nano-omni-30b-a3b-reasoning
+  matchPatterns:
+    - nvidia-nemotron-3-nano-omni-30b-a3b-reasoning
+    - nemotron-3-nano-omni-30b-a3b-reasoning-ud-q4_k_m
+    - nemotron-3-nano-omni-30b-a3b-reasoning
+    - nemotron-3-nano-omni
+    - nemotron-nano-omni
+  capabilities:
+    chat: true
+    tools: true
+    toolCallFormat: qwen
+    reasoning: true
+    thinkingFormat: qwen-chat-template
+    structuredOutputs: json-schema
+    vision: true
+    audio: true
+    embeddings: false
+    rerank: false
+    fim: false
+    contextWindow: 1048576
+    maxTokens: 131072
+  quirks:
+    sampling:
+      thinking:
+        temperature: 0.6
+        topP: 0.95
+        topK: 20
+        maxTokens: 20480
+        reasoningBudget: 16384
+        gracePeriod: 1024
+      instruct:
+        temperature: 0.2
+        topK: 1
+        maxTokens: 1024
+    gpuTiers:
+      "24gb": "Use a 4-bit quant with reduced load context for evaluation. Prefer the 32GB tier for sustained Clio writes."
+      "32gb": "Primary strong local model for multimodal main-agent work. Keep openai-compat available for tool-call fallback."
+    runtimePreference:
+      lmstudioNative: "Use only when the target's server lifecycle is Clio-managed or the model is already loaded at the requested context."
+      openaiCompat: "Preferred fallback for LM Studio gateway tool-call extraction."
+      llamaCpp: "OpenAI-compatible chat completions with --jinja."
+    llamaCpp:
+      ctxSize: 819200
+      cacheTypeK: q8_0
+      cacheTypeV: q8_0
+      flashAttn: true
+      nGpuLayers: 99
+      parallel: 4
+    thinking:
+      mechanism: budget-tokens
+      budgetByLevel:
+        low: 1024
+        medium: 4096
+        high: 16384
+      guidance: |
+        Reasoning is template-driven; thinking blocks emit through the qwen
+        chat template. Preserve thinking across coding-task turns when the
+        chain matters; otherwise let the budget cap output growth.
+    serving: "Use qwen3_coder tool parsing when served through OpenAI-compatible HTTP. LM Studio native uses SDK rawTools, but openai-compat remains the fallback for template-specific tool extraction issues."
+
+- family: qwen3.6-35b-a3b
+  matchPatterns:
+    - qwen3.6-35b-a3b
+    - qwen3_6-35b-a3b
+    - qwen3-6-35b-a3b
+    - qwen3.6-35b
+    - qwen3.6-35b-a3b-ud
+    - qwen3.6-35b-a3b-ud-q4_k_xl
+  capabilities:
+    chat: true
+    tools: true
+    toolCallFormat: qwen
+    reasoning: true
+    thinkingFormat: qwen-chat-template
+    structuredOutputs: json-schema
+    vision: true
+    audio: false
+    embeddings: false
+    rerank: false
+    fim: false
+    contextWindow: 262144
+    maxTokens: 65536
+  quirks:
+    sampling:
+      thinking:
+        temperature: 0.6
+        topP: 0.95
+        topK: 20
+        maxTokens: 16384
+        reasoningBudget: 4096
+      instruct:
+        temperature: 0.7
+        topP: 0.80
+        topK: 20
+        minP: 0.0
+    gpuTiers:
+      "32gb": "Main-agent candidate on 32GB targets when Nemotron Omni is unavailable or tool behavior is better for text-only work."
+    runtimePreference:
+      lmstudioNative: "Usable for text and vision when loaded in LM Studio. Keep openai-compat available for tool fallback."
+      openaiCompat: "Preferred local gateway path for tool-call stress tests."
+      llamaCpp: "OpenAI-compatible chat completions with --jinja. Local llama.cpp profile emits reasoning_content without a separate --reasoning flag."
+    llamaCpp:
+      ctxSize: 262144
+      cacheTypeK: q8_0
+      cacheTypeV: q8_0
+      flashAttn: true
+      nGpuLayers: 99
+    thinking:
+      mechanism: budget-tokens
+      budgetByLevel:
+        low: 1024
+        medium: 4096
+        high: 16384
+      guidance: |
+        Qwen chat template emits thinking blocks. The card recommends at least
+        128K context for thinking quality. Preserve thinking blocks across
+        coding-task turns so iterative reasoning chains do not restart cold.
+    serving: "Official card recommends at least 128K context for thinking quality. Use 262K when memory allows."
+
+- family: ornith-1.0-35b
+  matchPatterns:
+    - ornith-1.0-35b
+    - ornith-1.0-35b-q4_k_m
+    - ornith-1.0-35b-gguf
+    - deepreinforce-ai/ornith-1.0-35b-gguf
+    - ornith
+  capabilities:
+    chat: true
+    tools: true
+    toolCallFormat: qwen
+    reasoning: true
+    thinkingFormat: qwen-chat-template
+    structuredOutputs: json-schema
+    vision: false
+    audio: false
+    embeddings: false
+    rerank: false
+    fim: false
+    contextWindow: 262144
+    maxTokens: 65536
+  quirks:
+    kvCache:
+      kQuant: q8_0
+      vQuant: q8_0
+    sampling:
+      thinking:
+        temperature: 0.6
+        topP: 0.95
+        topK: 20
+        maxTokens: 16384
+        reasoningBudget: 4096
+      instruct:
+        temperature: 0.6
+        topP: 0.95
+        topK: 20
+        repeatPenalty: 1.05
+        maxTokens: 4096
+    gpuTiers:
+      "32gb": "32 GB local profile with Q4_K_M, ctx=262144, q8 KV, flash attention, and parallel=1."
+    runtimePreference:
+      llamaCpp: "OpenAI-compatible chat completions with --jinja and --reasoning on. Recommended local profile uses ctx=262144, parallel=1, batch=2048, ubatch=512, flash attention, and q8 KV cache."
+      openaiCompat: "Preferred when a gateway fronts the same llama.cpp chat-completions surface with qwen reasoning/tool parsing."
+    llamaCpp:
+      ctxSize: 262144
+      cacheTypeK: q8_0
+      cacheTypeV: q8_0
+      flashAttn: true
+      nGpuLayers: 99
+      parallel: 1
+      batchSize: 2048
+      ubatchSize: 512
+    thinking:
+      mechanism: always-on
+      guidance: |
+        Ornith emits a Qwen-style <think> block by default. The local llama.cpp
+        router is configured with --reasoning on, so the OpenAI-compatible
+        response separates this trace into reasoning_content while content
+        contains the final answer.
+    serving: "Ornith-1.0-35B is an agentic coding model; the card recommends 262K context, qwen3 reasoning parsing, qwen3_coder/qwen3_xml tool parsing, and sampler temperature=0.6, top_p=0.95, top_k=20."
+
+- family: qwen3.6-27b
+  matchPatterns:
+    - qwen3.6-27b
+    - qwen3_6-27b
+    - qwen3-6-27b
+  capabilities:
+    chat: true
+    tools: true
+    toolCallFormat: qwen
+    reasoning: true
+    thinkingFormat: qwen-chat-template
+    structuredOutputs: json-schema
+    vision: true
+    audio: false
+    embeddings: false
+    rerank: false
+    fim: false
+    contextWindow: 262144
+    maxTokens: 32768
+  quirks:
+    sampling:
+      thinking:
+        temperature: 0.6
+        topP: 0.95
+        topK: 20
+        minP: 0.0
+        presencePenalty: 0.0
+        repetitionPenalty: 1.0
+        maxTokens: 32768
+      instruct:
+        temperature: 0.7
+        topP: 0.80
+        topK: 20
+        minP: 0.0
+        presencePenalty: 1.5
+        repetitionPenalty: 1.0
+        maxTokens: 32768
+    gpuTiers:
+      "32gb": "Fits at f16 KV with parallel=1 and ctx=262144 (Q4_K_XL weights ~19.5 GB plus ~32 GB KV)."
+    runtimePreference:
+      lmstudioNative: "Primary path; standard <think> tagging works correctly with the SDK classifier."
+      openaiCompat: "Verified parity for tool calls; reasoning_content surfaces and completion_tokens_details.reasoning_tokens is reported."
+    thinking:
+      mechanism: budget-tokens
+      budgetByLevel:
+        low: 1024
+        medium: 4096
+        high: 16384
+      guidance: |
+        Qwen chat template drives thinking via standard <think> blocks.
+        Coding sampler is the precise variant. Preserve thinking blocks across
+        turns when the workflow is iterative; the budget caps runaway chains.
+    serving: "Official Qwen card recommends max output 32768 tokens for standard queries (81920 only for math/code competitions). Coding sampler is the precise variant; preserve thinking blocks across turns for iterative workflows."
+
+- family: gemma-4-31b-it-nvfp4-turbo
+  matchPatterns:
+    - gemma-4-31b-it-nvfp4-turbo
+    - gemma-4-31b-it-nvfp4
+    - gemma-4-31b-it-turbo
+    - gemma-4-31b-it
+    - gemma-4-31b
+  capabilities:
+    chat: true
+    tools: true
+    toolCallFormat: openai
+    reasoning: true
+    thinkingFormat: qwen-chat-template
+    structuredOutputs: json-schema
+    vision: true
+    audio: false
+    embeddings: false
+    rerank: false
+    fim: false
+    contextWindow: 122880
+    maxTokens: 32768
+  quirks:
+    kvCache:
+      kQuant: q8_0
+      vQuant: q8_0
+    sampling:
+      thinking:
+        temperature: 1.0
+        topP: 0.95
+        topK: 64
+        minP: 0.0
+        presencePenalty: 0.0
+        repetitionPenalty: 1.0
+        maxTokens: 32768
+      instruct:
+        temperature: 0.7  # 0.7 keeps tool-call JSON parseable; gemma card's 1.0 is for chat
+        topP: 0.95
+        topK: 64
+        minP: 0.0
+        maxTokens: 32768
+    chatTemplate: "gemma-channel"
+    leakageNote: |
+      LM Studio's gemma4 chat template emits chain-of-thought as plain text
+      bracketed by literal '<|channel>thought\n' and '<channel|>' markers
+      rather than reasoningType=reasoning fragments. Clio's lmstudio-native
+      adapter buffers and re-classifies these regions into ThinkingContent
+      so the markers do not leak into the visible TUI text and reasoning
+      tokens are accounted via the chars/4 estimator.
+    gpuTiers:
+      "48gb": "Fits at f16 KV when parallel=1 and ctx=122880; drop to q8_0 KV at parallel=4."
+      "80gb": "Comfortable at f16 KV with parallel=1 and ctx=122880."
+    runtimePreference:
+      lmstudioNative: "Verified path. Pin parallel=1 in the LM Studio per-model preset to give a single agent the full 122880 context."
+      openaiCompat: "Fallback path; the openai-compat surface does not expose the channel-marker classifier so visible text shows the raw markers."
+    thinking:
+      mechanism: on-off
+      guidance: |
+        Gemma 4 chat template exposes thinking only as on or off; intermediate
+        levels coerce to on. NVFP4 quantization can corrupt JSON tool-call
+        output under aggressive sampling, so the instruct profile pins
+        temperature at 0.7. The lmstudio-native adapter re-classifies the
+        channel-marker thought region into ThinkingContent so markers do not
+        leak into visible text.
+    serving: "LilaRest publishes no per-model sampler; values inherit the gemma family default per the official Google Gemma 4 31B card. KV cache budget at f16, ctx=122880 is roughly 57 GB on top of ~21 GB of NVFP4 weights."
+
+# The qat/UD-Q4_K_XL/MTP build is the long-context llama.cpp gemma profile and is a
+# different serving profile from the NVFP4 turbo family above: 262K context
+# (not 122880), q8 KV, F16 gemma4v mmproj vision, and MTP speculative decoding.
+# Its longest matchPattern (gemma-4-31b-it-qat-...) is longer than the NVFP4
+# family's `gemma-4-31b-it` (14 chars), so the longest-substring matcher in
+# knowledge-base.ts resolves this id here rather than to the NVFP4 profile.
+- family: gemma-4-31b-it-qat-mtp
+  matchPatterns:
+    - gemma-4-31b-it-qat-ud-q4_k_xl-mtp-262k
+    - gemma-4-31b-it-qat-ud-q4_k_xl-mtp
+    - gemma-4-31b-it-qat-ud
+    - gemma-4-31b-it-qat
+  capabilities:
+    chat: true
+    tools: true
+    toolCallFormat: openai
+    reasoning: true
+    thinkingFormat: qwen-chat-template
+    structuredOutputs: json-schema
+    vision: true
+    audio: false
+    embeddings: false
+    rerank: false
+    fim: false
+    contextWindow: 262144
+    maxTokens: 32768
+  quirks:
+    kvCache:
+      kQuant: q8_0
+      vQuant: q8_0
+    sampling:
+      thinking:
+        temperature: 1.0
+        topP: 0.95
+        topK: 64
+        minP: 0.0
+        presencePenalty: 0.0
+        repetitionPenalty: 1.0
+        maxTokens: 32768
+      instruct:
+        temperature: 0.7  # 0.7 keeps tool-call JSON parseable; router profiles may serve this qat build at 0.8
+        topP: 0.95
+        topK: 64
+        minP: 0.0
+        repeatPenalty: 1.05
+        maxTokens: 32768
+    chatTemplate: "gemma-channel"
+    gpuTiers:
+      "32gb": "32 GB Vulkan llama.cpp profile with UD-Q4_K_XL weights, ctx=262144, q8 KV, flash attention, parallel=1, F16 gemma4v mmproj (vision), and MTP draft (spec-type draft-mtp, n_max 4) both active."
+    runtimePreference:
+      llamaCpp: "OpenAI-compatible chat completions with --jinja and --reasoning off. Recommended local profile serves at ctx=262144, q8 KV, flash attention, parallel=1, batch=1024, ubatch=256, with F16 mmproj vision and draft-mtp speculative decoding; sampler temperature 0.8, top_p 0.95, top_k 64, repeat_penalty 1.05."
+      openaiCompat: "Fallback when a gateway fronts the same llama.cpp chat-completions surface; the openai-compat surface does not expose the channel-marker classifier, so visible text may show raw thought markers."
+    llamaCpp:
+      ctxSize: 262144
+      cacheTypeK: q8_0
+      cacheTypeV: q8_0
+      flashAttn: true
+      nGpuLayers: 99
+      parallel: 1
+      batchSize: 1024
+      ubatchSize: 256
+      mmproj: true
+    thinking:
+      mechanism: on-off
+      guidance: |
+        Gemma 4 31B IT QAT exposes thinking only as on or off via the gemma-4
+        enable_thinking template; intermediate levels coerce to on. A local
+        llama.cpp profile serves it with --reasoning off (default_reasoning off),
+        so the chain-of-thought is not requested; if the template emits one it
+        surfaces via reasoning_content and is accounted as reasoning tokens. The
+        QAT weights hold tool-call JSON together better than the NVFP4 turbo
+        build, but the instruct profile still pins a conservative temperature.
+    serving: "Gemma 4 31B IT QAT, UD-Q4_K_XL, served as a local llama.cpp router model (vulkan) with MTP speculative decoding (spec-type draft-mtp, n_max 4, draft mtp-gemma-4-31B-it) and F16 gemma4v mmproj vision both active at 262k. Reasoning off by default; gemma-4 thinking is on/off. Sampler temperature 0.8, top_p 0.95, top_k 64, repeat_penalty 1.05."
+
+- family: gemopus-4-31b-it
+  matchPatterns:
+    - gemopus-4-31b-it
+    - gemopus-4-31b
+    - gemopus-4
+  capabilities:
+    chat: true
+    tools: true
+    toolCallFormat: openai
+    reasoning: true
+    thinkingFormat: qwen-chat-template
+    structuredOutputs: json-schema
+    vision: false
+    audio: false
+    embeddings: false
+    rerank: false
+    fim: false
+    contextWindow: 122880
+    maxTokens: 32768
+  quirks:
+    kvCache:
+      kQuant: q8_0
+      vQuant: q8_0
+    sampling:
+      thinking:
+        temperature: 1.0
+        topP: 0.95
+        topK: 64
+        minP: 0.0
+        presencePenalty: 0.0
+        repetitionPenalty: 1.0
+        maxTokens: 32768
+      instruct:
+        temperature: 0.7  # 0.7 keeps tool-call JSON parseable; gemma card's 1.0 is for chat
+        topP: 0.95
+        topK: 64
+        minP: 0.0
+        maxTokens: 32768
+    chatTemplate: "gemma-channel"
+    thinkingControl: |
+      Jackrong's distilled gemma-4 31B uses the same channel-marker template.
+      Thinking mode is gated by the literal '<|think|>' token in the system
+      prompt; Clio does not currently inject it, so the model emits an empty
+      thought region followed by the answer.
+    gpuTiers:
+      "48gb": "Fits at f16 KV with parallel=1; q8_0 KV recommended at parallel=4."
+      "80gb": "Comfortable at f16 KV with parallel=1 and ctx=122880."
+    runtimePreference:
+      lmstudioNative: "Verified path. Channel-marker handler in clio's adapter recovers reasoning tokens that the SDK classifier drops."
+      openaiCompat: "Fallback; reasoning_content is absent from the openai-compat reasoning probe for this family."
+    thinking:
+      mechanism: always-on
+      guidance: |
+        Distilled traces emit unconditionally. The thinking level setting is
+        ignored at the API surface; chain-of-thought arrives whether or not
+        Clio asks for it. Reasoning tokens are accounted via the channel-marker
+        re-classifier in lmstudio-native.
+    serving: "Distilled from Claude 4.6 Opus reasoning traces onto Google Gemma 4 31B; sampler matches the gemopus card values."
+
+- family: qwopus3.6-27b-v1-preview
+  matchPatterns:
+    - qwopus3.6-27b-v1-preview
+    - qwopus3.6-27b
+    - qwopus3.6
+  capabilities:
+    chat: true
+    tools: true
+    toolCallFormat: qwen
+    reasoning: true
+    thinkingFormat: qwen-chat-template
+    structuredOutputs: json-schema
+    vision: false
+    audio: false
+    embeddings: false
+    rerank: false
+    fim: false
+    contextWindow: 262144
+    maxTokens: 32768
+  quirks:
+    sampling:
+      thinking:
+        temperature: 0.6
+        topP: 0.95
+        topK: 20
+        minP: 0.0
+        presencePenalty: 0.0
+        repetitionPenalty: 1.0
+        maxTokens: 32768
+      instruct:
+        temperature: 0.7
+        topP: 0.80
+        topK: 20
+        minP: 0.0
+        presencePenalty: 1.5
+        repetitionPenalty: 1.0
+        maxTokens: 32768
+    gpuTiers:
+      "32gb": "Fits at f16 KV with parallel=1 and ctx=262144 thanks to the Q4_K_M weights (~16.5 GB)."
+    runtimePreference:
+      lmstudioNative: "Primary path; thinking is emitted via standard <think> tags so the SDK classifier accounts reasoning correctly."
+      openaiCompat: "Verified for tool-call extraction parity; reasoning_content surfaces."
+    thinking:
+      mechanism: budget-tokens
+      budgetByLevel:
+        low: 1024
+        medium: 4096
+        high: 16384
+      guidance: |
+        Distilled Opus reasoning traces over Qwen3.6 27B; thinking emits via
+        standard <think> tags. Coding-task sampler is the precise variant.
+        Preserve thinking blocks across turns for iterative workflows.
+    serving: "Distilled from Claude 4.6 Opus reasoning traces onto Qwen3.6 27B. Coding-task sampler is the precise variant: temperature 0.6, top_p 0.95, top_k 20."
+
+- family: qwopus3.6-35b-a3b-coder
+  matchPatterns:
+    - qwopus3.6-35b-a3b-coder-mtp-q4_k_m-262k
+    - qwopus3.6-35b-a3b-coder-mtp
+    - qwopus3.6-35b-a3b-coder
+    - qwopus3.6-35b-a3b
+  capabilities:
+    chat: true
+    tools: true
+    toolCallFormat: qwen
+    # Reasoning class: effort-levels, measured against an LM Studio host on
+    # 2026-08-08 over its OpenAI-compatible surface. The card
+    # frames the Coder-MTP line as thinking-off execution, but the wire
+    # disagrees: on "what is 17+25" this model spends 98 of 103 completion
+    # tokens on reasoning, and a wiki planning dispatch spent 89,501. It is an
+    # always-reasoning model unless told otherwise.
+    #
+    # chat_template_kwargs enable_thinking is inert here (verified with unique
+    # prompts to rule out cache hits). reasoning_effort "none" is the control
+    # that works, taking the same prompt from 98 reasoning tokens to 0, which
+    # is why `off` is mapped explicitly below rather than left unsent.
+    #
+    # reasoning=false previously resolved to mechanism "none", and mechanism
+    # "none" makes openai-completions strip reasoning_effort from the payload.
+    # The flag set to suppress thinking was the reason it could not be
+    # suppressed.
+    reasoning: true
+    thinkingFormat: qwen-chat-template
+    structuredOutputs: json-schema
+    vision: true
+    audio: false
+    embeddings: false
+    rerank: false
+    fim: false
+    contextWindow: 262144
+    maxTokens: 32768
+  quirks:
+    sampling:
+      instruct:
+        temperature: 0.2
+        topP: 0.9
+        topK: 20
+        repeatPenalty: 1.05
+        # Upstream Qwen3.6 card recommends presence_penalty 1.5 for
+        # non-thinking modes to suppress repetition loops. Measured on live
+        # dispatch (2026-07-06): without it a coder worker repeated one
+        # code_nav call into the loop-guard abort on 3 of 3 runs; with it the
+        # same task passed 3 of 3 with clean edit-then-validate trajectories.
+        presencePenalty: 1.5
+        maxTokens: 32768
+    gpuTiers:
+      "32gb": "Default 32 GB local profile: Q4_K_M, ctx=262144, q8 KV, flash attention, parallel=1, with F32 mmproj (vision) and draft-mtp speculative decoding both active. Leaves several GiB free at 262k on a 32 GiB class card."
+    runtimePreference:
+      llamaCpp: "OpenAI-compatible chat completions with --jinja. Recommended local profile serves text-only with reasoning off and enable_thinking false; coder sampler temperature 0.2, top_p 0.9, top_k 20."
+      openaiCompat: "Use when a gateway fronts the same llama.cpp chat-completions surface with qwen tool parsing."
+    llamaCpp:
+      ctxSize: 262144
+      cacheTypeK: q8_0
+      cacheTypeV: q8_0
+      flashAttn: true
+      nGpuLayers: 99
+      parallel: 1
+      batchSize: 2048
+      ubatchSize: 512
+      mmproj: true
+      chatTemplateKwargs:
+        enable_thinking: false
+    thinking:
+      mechanism: effort-levels
+      effortByLevel:
+        # "none" is the only value that silences this family. It is carried on
+        # the wire because the model reasons by default, so sending nothing is
+        # in effect a request to keep reasoning.
+        off: none
+        low: low
+        medium: medium
+        high: high
+      guidance: |
+        Qwopus3.6 35B-A3B Coder reasons unless reasoning_effort is set to
+        "none", despite the creator's card describing the Coder-MTP variants as
+        thinking-off. That dial only reaches the wire on openai-compat; the LM
+        Studio native transport has no field for it, so a native-routed run
+        reasons at full rate whatever the configured thinking level says.
+    serving: "Jackrong Qwopus3.6 35B-A3B Coder MTP, Q4_K_M. Thinking-off coder/agent tuned for tool-use loops. Served as a local router default with MTP speculative decoding (spec-type draft-mtp, n_max 2) and F32 mmproj vision both active at 262k."
+
+- family: qwopus3.6-27b-coder
+  matchPatterns:
+    - qwopus3.6-27b-coder-mtp-q5_k_m-262k
+    - qwopus3.6-27b-coder-mtp
+    - qwopus3.6-27b-coder
+  capabilities:
+    chat: true
+    tools: true
+    toolCallFormat: qwen
+    # Reasoning class: never, same Coder-MTP thinking-off design as the 35B.
+    reasoning: false
+    thinkingFormat: qwen-chat-template
+    structuredOutputs: json-schema
+    vision: false
+    audio: false
+    embeddings: false
+    rerank: false
+    fim: false
+    contextWindow: 262144
+    maxTokens: 32768
+  quirks:
+    sampling:
+      instruct:
+        temperature: 0.2
+        topP: 0.9
+        topK: 20
+        repeatPenalty: 1.05
+        # Same Coder-MTP non-thinking design as the 35B; see the measured
+        # presence-penalty note there.
+        presencePenalty: 1.5
+        maxTokens: 32768
+    runtimePreference:
+      llamaCpp: "Mini router serves it with --jinja and enable_thinking false; coder sampler temperature 0.2, top_p 0.9, top_k 20."
+      lmstudioNative: "Dense 27B Coder-MTP; the creator's template ships thinking-off, and no per-request field re-enables it."
+    thinking:
+      mechanism: none
+      guidance: |
+        Qwopus3.6 27B Coder-MTP is the dense thinking-off coder variant; the
+        dial clamps to off at every level.
+    serving: "Jackrong Qwopus3.6 27B Coder MTP, Q5_K_M. Thinking-off dense coder for tool-use loops."
+
+- family: qwopus3.5-9b-coder
+  matchPatterns:
+    - qwopus3.5-9b-coder-q8_0-262k
+    - qwopus3.5-9b-coder
+  capabilities:
+    chat: true
+    tools: true
+    toolCallFormat: qwen
+    # Reasoning class: never (Coder thinking-off design, 3.5 generation).
+    reasoning: false
+    thinkingFormat: qwen-chat-template
+    structuredOutputs: json-schema
+    vision: false
+    audio: false
+    embeddings: false
+    rerank: false
+    fim: false
+    contextWindow: 262144
+    maxTokens: 32768
+  quirks:
+    sampling:
+      instruct:
+        temperature: 0.2
+        topP: 0.9
+        topK: 20
+        repeatPenalty: 1.05
+        maxTokens: 32768
+    runtimePreference:
+      llamaCpp: "Mini router serves it with --jinja and enable_thinking false; coder sampler temperature 0.2, top_p 0.9, top_k 20."
+    thinking:
+      mechanism: none
+      guidance: |
+        Qwopus3.5 9B Coder is the small thinking-off coder; the dial clamps
+        to off at every level.
+    serving: "Jackrong Qwopus3.5 9B Coder, Q8_0. Thinking-off small coder for fast tool loops."
+
+- family: gemma4-26b-a4b
+  matchPatterns:
+    - gemma-4-26b-a4b
+    - gemma4-26b-a4b
+    - gemma-4-26b-a4b-it
+    - gemma4-26b-a4b-it
+    - gemma-4-26b-a4b-it-q4
+    - gemma-4-26b-a4b-it-q4_k_m
+  capabilities:
+    chat: true
+    tools: true
+    toolCallFormat: openai
+    reasoning: true
+    thinkingFormat: qwen-chat-template
+    structuredOutputs: json-schema
+    vision: true
+    audio: false
+    embeddings: false
+    rerank: false
+    fim: false
+    contextWindow: 262144
+    maxTokens: 65536
+  quirks:
+    sampling:
+      thinking:
+        temperature: 0.3
+        topP: 0.9
+        topK: 20
+        maxTokens: 8192
+        reasoningBudget: 4096
+      instruct:
+        temperature: 0.7  # 0.7 keeps tool-call JSON parseable; gemma card's 1.0 is for chat
+        topP: 0.95
+        topK: 64
+        minP: 0.0
+    gpuTiers:
+      "32gb": "Main-agent candidate for 32 GB and larger local systems when Gemma behavior is preferred."
+    runtimePreference:
+      lmstudioNative: "Use for already-loaded LM Studio sessions and multimodal checks."
+      openaiCompat: "Preferred fallback when SDK rawTools do not match the template."
+      llamaCpp: "OpenAI-compatible chat completions with --jinja, --reasoning on, and the matching mmproj."
+    llamaCpp:
+      ctxSize: 262144
+      cacheTypeK: q8_0
+      cacheTypeV: q8_0
+      flashAttn: true
+      nGpuLayers: 99
+      mmproj: true
+    thinkingControl: "Gemma 4 exposes thinking through chat template enable_thinking support."
+    thinking:
+      mechanism: on-off
+      guidance: |
+        Gemma 4 chat template exposes thinking only as on or off via
+        enable_thinking. Intermediate levels coerce to on. The MoE A4B variant
+        keeps the same template behavior as the dense gemma-4 family.
+
+- family: nemotron-cascade-2-30b-a3b
+  matchPatterns:
+    - nemotron-cascade-2
+    - nemotron-cascade-2-30b-a3b
+    - nemotron-cascade-2-30b-a3b-i1
+    - nemotron-cascade-2-30b-a3b-i1-q4_k_m
+    - nemotron-cascade2
+    - nemotron-cascade-2-30b
+  capabilities:
+    chat: true
+    tools: true
+    toolCallFormat: qwen
+    reasoning: true
+    thinkingFormat: qwen-chat-template
+    structuredOutputs: json-schema
+    vision: false
+    audio: false
+    embeddings: false
+    rerank: false
+    fim: false
+    contextWindow: 1048576
+    maxTokens: 65536
+  quirks:
+    sampling:
+      instruct:
+        temperature: 0.3
+        topP: 0.9
+        topK: 20
+        maxTokens: 8192
+    gpuTiers:
+      "32gb": "Preferred local worker model when a text-only worker is enough."
+    runtimePreference:
+      lmstudioNative: "Use when LM Studio SDK tool extraction is verified for the loaded template."
+      openaiCompat: "Preferred fallback for LM Studio gateway tool calls."
+      anthropicCompat: "Do not select for the current llama.cpp target profile; /v1/messages returned 404 while /v1/models and OpenAI-compatible chat worked."
+      llamaCpp: "OpenAI-compatible chat completions with --jinja. Recommended local profile disables thinking in chat_template_kwargs."
+    llamaCpp:
+      ctxSize: 1048576
+      cacheTypeK: q8_0
+      cacheTypeV: q8_0
+      flashAttn: true
+      nGpuLayers: 99
+      parallel: 4
+      chatTemplateKwargs:
+        enable_thinking: false
+    thinking:
+      mechanism: on-off
+      guidance: |
+        Nemotron Cascade 2 chat template exposes thinking through
+        chat_template_kwargs.enable_thinking. Intermediate levels coerce to on.
+        The recommended llama.cpp profile disables thinking in the template kwargs by
+        default; flip via the runtime override when reasoning is needed.
+    serving: "NVIDIA card recommends vLLM with nemotron_v3 reasoning and qwen3_coder tool parsing."
+
+- family: qwen3.5-35b-a3b-claude-4.6-opus-reasoning-distilled
+  matchPatterns:
+    - qwen3.5-35b-a3b-claude-4.6-opus-reasoning-distilled
+    - qwen3.5-35b-a3b-claude-4.6-opus-reasoning-distilled-i1
+    - qwen3.5-35b-a3b-claude-4.6-opus-reasoning-distilled-i1-q4_k_m
+    - qwen35-distilled
+    - qwen35-distilled-i1
+    - qwen35-distilled-i1-q4_k_m
+  capabilities:
+    chat: true
+    tools: true
+    toolCallFormat: qwen
+    reasoning: true
+    thinkingFormat: qwen-chat-template
+    structuredOutputs: json-schema
+    vision: false
+    audio: false
+    embeddings: false
+    rerank: false
+    fim: false
+    contextWindow: 262144
+    maxTokens: 32768
+  quirks:
+    sampling:
+      thinking:
+        temperature: 0.6
+        topP: 0.95
+        topK: 20
+        maxTokens: 16384
+        reasoningBudget: 4096
+      instruct:
+        temperature: 0.2
+        topP: 0.9
+        topK: 20
+        maxTokens: 4096
+    gpuTiers:
+      "32gb": "Distilled Opus-style reasoning at Qwen3.5 35B A3B scale. Use as a strong local main agent when GPU room is available."
+    runtimePreference:
+      lmstudioNative: "Verified usable on LM Studio when loaded with the Qwen-style preset."
+      openaiCompat: "Preferred fallback when SDK rawTools yield empty arguments for the distilled template."
+      llamaCpp: "OpenAI-compatible chat completions with --jinja. Match qwen3 reasoning parser flags upstream."
+    thinking:
+      mechanism: budget-tokens
+      budgetByLevel:
+        low: 1024
+        medium: 4096
+        high: 16384
+      guidance: |
+        Distilled Opus reasoning over Qwen3.5 35B A3B; thinking emits via the
+        qwen chat template. Preserve thinking blocks across coding-task turns;
+        the budget caps runaway chains while keeping iterative reasoning
+        coherent.
+    serving: "Distilled from Claude 4.6 Opus reasoning traces; expect stronger thinking-then-answer behavior than vanilla Qwen3.5 35B A3B."
+
+- family: qwopus3.5-9b-v3
+  matchPatterns:
+    - qwopus3.5-9b-v3
+    - qwopus-9b
+    - qwopus 9b
+  capabilities:
+    chat: true
+    tools: true
+    toolCallFormat: qwen
+    reasoning: true
+    thinkingFormat: qwen-chat-template
+    structuredOutputs: json-schema
+    vision: false
+    audio: false
+    embeddings: false
+    rerank: false
+    fim: false
+    contextWindow: 262144
+    maxTokens: 32768
+  quirks:
+    sampling:
+      instruct:
+        temperature: 0.2
+        topP: 0.9
+        topK: 20
+        maxTokens: 4096
+      thinking:
+        temperature: 0.6
+        topP: 0.95
+        topK: 20
+        maxTokens: 8192
+        reasoningBudget: 2048
+    gpuTiers:
+      "16gb": "Use this as the local 16GB emulation target. It keeps the output cap below the 32GB-class MoE defaults."
+    runtimePreference:
+      lmstudioNative: "Primary local laptop path. Verified through LM Studio /api/v1/models with 262144 loaded context, flash attention, GPU KV offload, and trained_for_tool_use."
+      openaiCompat: "Verified through the LM Studio gateway for read and artifact tool calls; use when the native SDK produces thinking-only length stops or malformed tools."
+      anthropicCompat: "LM Studio's Anthropic Messages surface streams simple text with the server root URL as baseUrl; keep as a protocol fallback until tool and reasoning behavior are proven."
+    thinking:
+      # Measured 2026-08-11 against LM Studio on this exact model: baseline 56
+      # reasoning tokens, reasoning_effort none 120, reasoning_effort minimal
+      # 243, chat_template_kwargs.enable_thinking false 45, both together 44.
+      # No spelling silences it, and the budget-tokens mechanism it used to
+      # claim is informational on both LM Studio and llama.cpp, so nothing ever
+      # reached the wire. Always-on is what the model actually does.
+      mechanism: always-on
+      guidance: |
+        Qwopus 9B distilled from Opus reasoning; the chain-of-thought comes out
+        of the chat template unconditionally and the dial cannot stop it, so
+        Clio shows the level as forced. Output cap is tighter than the
+        32GB-class MoE families because this is the 16GB emulation target;
+        budget for a reasoning preamble ahead of every answer.
+    serving: "LM Studio reports trained_for_tool_use for the local qwopus3.5-9b-v3 target. Native model override successfully switched residency to granite-4.0-h-350m through the SDK."