abstractcore 2.34.0__tar.gz → 2.35.0__tar.gz

{abstractcore-2.34.0 → abstractcore-2.35.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: abstractcore
-Version: 2.34.0
+Version: 2.35.0
 Summary: Unified interface to all LLM providers with essential infrastructure for tool calling, streaming, and model management
 Author-email: Laurent-Philippe Albou <contact@abstractcore.ai>
 Maintainer-email: Laurent-Philippe Albou <contact@abstractcore.ai>
@@ -196,20 +196,14 @@ Dynamic: license-file
 Unified LLM Interface
 > Write once, run everywhere
 
+What's new in 2.35.0: the model capability registry covers the current vendor lineups (GPT-6 and GPT-6.1, Claude 5.5 and Fable 5.1, Grok 4.3 to 4.7, Qwen 3.8, DeepSeek V4.1 Flash, GLM 5.3, MiniMax M3.1 Flash, Kimi K3 and K3.1 Pro, LFM3, Spark X2.5/X2.6, Nemotron 3.5, Bonsai 2 and more), so an image, video or voice route can name them; and a GGUF or MLX re-packaging of a registered model (`unsloth/...-GGUF:Q4_K_M`, `mlx-community/...-4bit`, bartowski's `Org_Model-GGUF`) resolves to that model's entry instead of the text-only default. Transcriptions report `language` and `detected_language`. See the [changelog](CHANGELOG.md#2350---2026-11-07).
+
 What's new in 2.34.0: a local model that no call has used for 60 minutes is unloaded from memory and loads
 again on its next call; a model kept loaded (Keep loaded) never is, and no running call is interrupted
 (`process_residency.set_idle_unload(seconds)`, `0` turns it off). A provider key you send to the server's discovery
 routes now reaches only the provider you name, never every provider those routes check. The web console's themes
 match ui-kit 0.9.14. See the [changelog](CHANGELOG.md#2340---2026-11-03).
 
-What's new in 2.33.0: GGUF prompt-cache snapshots drop an unused 0.5 GB copy each, stay under a byte budget
-(`gguf_prompt_cache_max_gb`), and are never evicted under a prefill that is reading them; a model that is not
-downloaded raises `ModelNotDownloadedError` naming it; an OpenAI-compatible key check refuses a wrong key even where
-the model list is public (OVH); hosted decision models (`typesafe`, `systemone`) are connection families, listed and
-routed as decision models, and a local decision model answers its own decisions; a provider profile this version
-cannot read is kept and reported instead of resetting the whole config. See the
-[changelog](CHANGELOG.md#2330---2026-11-03).
-
 AbstractCore is an offline-capable, open-source-first LLM infrastructure layer
 for Python applications. It gives you one `create_llm(...)` API across local
 runtimes, self-hosted servers, cloud APIs, and OpenAI-compatible gateways.
@@ -329,7 +323,29 @@ graph LR
 
 ## Install
 
-AbstractCore has three install settings:
+**Install in one line** (no admin rights, no Python needed):
+
+- **macOS / Linux:**
+
+  ```bash
+  curl -LsSf https://abstractframework.ai/core/install.sh | sh
+  ```
+
+- **Windows 10 / 11** (PowerShell):
+
+  ```powershell
+  powershell -ExecutionPolicy ByPass -c "irm https://abstractframework.ai/core/install.ps1 | iex"
+  ```
+
+The script detects the machine (Apple Silicon with macOS 14+: `abstractcore[apple]`; an NVIDIA or
+ROCm card: `abstractcore[gpu]`; anything else: the light install) and installs AbstractCore alone
+with uv and Python 3.12, giving you the `abstractcore`, `abstractcore-chat` and
+`abstractcore-endpoint` commands and the apps; run it again to upgrade. It never starts a compiler,
+so echo cancellation and stable-diffusion.cpp are left out (Diffusers / MLX-Gen cover image
+generation; with a C compiler you can add them yourself). Options and details:
+[One-line install](docs/installation.md#one-line-install).
+
+**With pip**, in your own Python environment, AbstractCore has three install settings:
 
 ```bash
 # Light: every remote provider (OpenAI, Anthropic, OpenRouter, Portkey, any OpenAI-compatible

{abstractcore-2.34.0 → abstractcore-2.35.0}/README.md

@@ -9,20 +9,14 @@
 Unified LLM Interface
 > Write once, run everywhere
 
+What's new in 2.35.0: the model capability registry covers the current vendor lineups (GPT-6 and GPT-6.1, Claude 5.5 and Fable 5.1, Grok 4.3 to 4.7, Qwen 3.8, DeepSeek V4.1 Flash, GLM 5.3, MiniMax M3.1 Flash, Kimi K3 and K3.1 Pro, LFM3, Spark X2.5/X2.6, Nemotron 3.5, Bonsai 2 and more), so an image, video or voice route can name them; and a GGUF or MLX re-packaging of a registered model (`unsloth/...-GGUF:Q4_K_M`, `mlx-community/...-4bit`, bartowski's `Org_Model-GGUF`) resolves to that model's entry instead of the text-only default. Transcriptions report `language` and `detected_language`. See the [changelog](CHANGELOG.md#2350---2026-11-07).
+
 What's new in 2.34.0: a local model that no call has used for 60 minutes is unloaded from memory and loads
 again on its next call; a model kept loaded (Keep loaded) never is, and no running call is interrupted
 (`process_residency.set_idle_unload(seconds)`, `0` turns it off). A provider key you send to the server's discovery
 routes now reaches only the provider you name, never every provider those routes check. The web console's themes
 match ui-kit 0.9.14. See the [changelog](CHANGELOG.md#2340---2026-11-03).
 
-What's new in 2.33.0: GGUF prompt-cache snapshots drop an unused 0.5 GB copy each, stay under a byte budget
-(`gguf_prompt_cache_max_gb`), and are never evicted under a prefill that is reading them; a model that is not
-downloaded raises `ModelNotDownloadedError` naming it; an OpenAI-compatible key check refuses a wrong key even where
-the model list is public (OVH); hosted decision models (`typesafe`, `systemone`) are connection families, listed and
-routed as decision models, and a local decision model answers its own decisions; a provider profile this version
-cannot read is kept and reported instead of resetting the whole config. See the
-[changelog](CHANGELOG.md#2330---2026-11-03).
-
 AbstractCore is an offline-capable, open-source-first LLM infrastructure layer
 for Python applications. It gives you one `create_llm(...)` API across local
 runtimes, self-hosted servers, cloud APIs, and OpenAI-compatible gateways.
@@ -142,7 +136,29 @@ graph LR
 
 ## Install
 
-AbstractCore has three install settings:
+**Install in one line** (no admin rights, no Python needed):
+
+- **macOS / Linux:**
+
+  ```bash
+  curl -LsSf https://abstractframework.ai/core/install.sh | sh
+  ```
+
+- **Windows 10 / 11** (PowerShell):
+
+  ```powershell
+  powershell -ExecutionPolicy ByPass -c "irm https://abstractframework.ai/core/install.ps1 | iex"
+  ```
+
+The script detects the machine (Apple Silicon with macOS 14+: `abstractcore[apple]`; an NVIDIA or
+ROCm card: `abstractcore[gpu]`; anything else: the light install) and installs AbstractCore alone
+with uv and Python 3.12, giving you the `abstractcore`, `abstractcore-chat` and
+`abstractcore-endpoint` commands and the apps; run it again to upgrade. It never starts a compiler,
+so echo cancellation and stable-diffusion.cpp are left out (Diffusers / MLX-Gen cover image
+generation; with a C compiler you can add them yourself). Options and details:
+[One-line install](docs/installation.md#one-line-install).
+
+**With pip**, in your own Python environment, AbstractCore has three install settings:
 
 ```bash
 # Light: every remote provider (OpenAI, Anthropic, OpenRouter, Portkey, any OpenAI-compatible

{abstractcore-2.34.0 → abstractcore-2.35.0}/abstractcore/architectures/detection.py

@@ -54,10 +54,26 @@ _VARIANT_TOKEN_RE = re.compile(
     r"(?i)^(?:"
     r"gguf|mlx|awq|gptq|bnb|quant|quantized|"
     r"(?:nv|mx)?fp\d+|bf\d+|f(?:16|32|64)|int\d+|uint\d+|nf4|"
-    r"[1-9]bit|(?:i|t)?q[1-8](?:_[a-z0-9]+)*"
+    r"[1-9]bit|(?:i|t)?q[1-8](?:_[a-z0-9]+)*|"
+    # MLX quantizer markers: mlx-community "OptiQ" (always followed by "-Nbit") and the
+    # "oQ4e"/"oQ2e" mixed-affine tags used by Jundot and mlx-community conversions.
+    r"optiq|oq[1-8]e?|"
+    # mradermacher imatrix markers ("-i1-GGUF", "-imatrix").
+    r"i1|imatrix|"
+    # Google QAT / compressed-tensors packagings ("-it-qat-q4_0-gguf", "-it-qat-w4a16-ct",
+    # "-it-qat-q4_0-unquantized", "-it-qat-mobile-transformers").
+    r"qat|unquantized|ct|mobile|transformers|w\d+a\d+(?:c\d+)?"
     r")$"
 )
+# A quant tag after a colon: Ollama / LM Studio / `hf.co/org/repo:TAG` ids such as
+# "unsloth/Qwen3.8-27B-GGUF:Q4_K_M", "...-GGUF:UD-Q4_K_XL" or "...:latest". Stripped before
+# the trailing-token walk so the repo name is normalized like any other variant.
+_COLON_QUANT_TAG_RE = re.compile(r"(?i):(?:ud-)?(?:(?:i|t)?q[1-8](?:_[a-z0-9]+)*|(?:nv|mx)?fp\d+|bf\d+|f(?:16|32)|[1-9]bit|latest)$")
+# bartowski / mradermacher style "Org_Model-GGUF": the upstream org folded into the repo
+# name with an underscore. Only a plain org-looking head is split off.
+_UNDERSCORED_ORG_PREFIX_RE = re.compile(r"^([A-Za-z][A-Za-z0-9.-]{1,40})_(?=[A-Za-z0-9])")
 _TRAILING_VARIANT_SEGMENT_RE = re.compile(r"(?i)([-_.@])([a-z0-9_]+)$")
+_MAX_TRAILING_VARIANT_TOKENS = 6
 _FORMAT_VARIANT_TOKENS = {
     "gguf",
     "safetensors",
@@ -102,6 +118,10 @@ def _variant_normalizations(name: str) -> List[str]:
     _add(s)
 
     working = s
+    tag_match = _COLON_QUANT_TAG_RE.search(working)
+    if tag_match:
+        working = working[:tag_match.start()]
+        _add(working)
     trailing_tokens_reversed: List[tuple[str, str]] = []
 
     extension_match = _VARIANT_EXTENSION_RE.search(working)
@@ -119,6 +139,10 @@ def _variant_normalizations(name: str) -> List[str]:
         ud_quant = token.lower() == "ud" and any(re.match(r"(?i)^(?:i|t)?q[1-8]", t) for _, t in trailing_tokens_reversed)
         if not _is_variant_token(token) and not ud_quant:
             break
+        if len(trailing_tokens_reversed) >= _MAX_TRAILING_VARIANT_TOKENS:
+            # The subset walk below is 2^n; real repo names carry at most four or
+            # five tags, and an adversarial name must not stall capability lookup.
+            break
         trailing_tokens_reversed.append((match.group(1), token))
         working = working[:match.start()].rstrip(".-_@")
 
@@ -396,12 +420,28 @@ def resolve_model_alias(model_name: str, models: Dict[str, Any]) -> str:
                 continue
             for normalized in _variant_normalizations(s):
                 out.append(normalized)
-                out.extend(_colon_variants(normalized))
+                for colon_variant in _colon_variants(normalized):
+                    out.extend(_variant_normalizations(colon_variant))
                 t = _tail(normalized)
                 if t:
                     for tail_normalized in _variant_normalizations(t):
                         out.append(tail_normalized)
-                        out.extend(_colon_variants(tail_normalized))
+                        for colon_variant in _colon_variants(tail_normalized):
+                            out.extend(_variant_normalizations(colon_variant))
+        # "Org_Model" repo names (bartowski, mradermacher) also resolve as "Org/Model" and
+        # as the bare model name. Alias matching is a set intersection, so these compete
+        # with the other candidates on equal terms; only the canonical-key fallback walks
+        # the list in order.
+        for s in list(out):
+            if "/" in s or "_" not in s:
+                continue
+            m = _UNDERSCORED_ORG_PREFIX_RE.match(s)
+            if not m:
+                continue
+            rest = s[m.end():]
+            if rest:
+                out.append(f"{m.group(1)}/{rest}")
+                out.append(rest)
         # Deduplicate while preserving order
         uniq: List[str] = []
         seen: set[str] = set()
@@ -451,6 +491,10 @@ def resolve_model_alias(model_name: str, models: Dict[str, Any]) -> str:
         if not isinstance(aliases, list) or not aliases:
             continue
         alias_set = {str(a).strip().lower() for a in aliases if isinstance(a, str) and str(a).strip()}
+        # A declared "org/Model" alias also answers to the bare "Model" tail, so a
+        # re-packaged repo ("unsloth/Model-GGUF", "mlx-community/Model-4bit") that
+        # strips to the tail still finds it. Only aliases with exactly one "/" qualify.
+        alias_set |= {a.split("/", 1)[1] for a in alias_set if a.count("/") == 1 and a.split("/", 1)[1]}
         if alias_set and cand_set and alias_set.intersection(cand_set):
             alias_matches.append(canonical_name)
 

{abstractcore-2.34.0 → abstractcore-2.35.0}/abstractcore/assets/architecture_formats.json

@@ -285,7 +285,13 @@
         "qwen3-32b",
         "qwen3:",
         "qwen3-4b-thinking",
-        "qwen3-4b-instruct"
+        "qwen3-4b-instruct",
+        "bonsai-8b",
+        "bonsai-4b",
+        "bonsai-1.7b",
+        "ternary-bonsai-8b",
+        "ternary-bonsai-4b",
+        "ternary-bonsai-1.7b"
       ]
     },
     "qwen3_5": {
@@ -449,7 +455,6 @@
       },
       "patterns": [
         "qwen3.8-flash-next",
-        "qwen3.8-flash",
         "flash-next",
         "qwen4exp",
         "qwen4-exp",
@@ -750,7 +755,8 @@
         "gemma4",
         "gemma4_unified",
         "gemma4-unified",
-        "google/gemma-4"
+        "google/gemma-4",
+        "nemotron-3.5-bolt-e4b"
       ]
     },
     "gemma2": {
@@ -951,7 +957,13 @@
         "kimi-k2.6",
         "moonshotai/kimi-k2.6",
         "kimi-k2",
-        "kimi_k25"
+        "kimi_k25",
+        "kimi-k3",
+        "moonshotai/kimi-k3",
+        "kimi_k3",
+        "kimi_k31",
+        "kimi-k2.7",
+        "kimi-k2.5"
       ]
     },
     "deepseek": {
@@ -1011,7 +1023,10 @@
         "minimax/minimax-m2.5",
         "minimax-m2",
         "minimaxai/minimax-m2",
-        "minimax/minimax-m2"
+        "minimax/minimax-m2",
+        "minimax-m2.7",
+        "minimaxai/minimax-m2.7",
+        "minimax/minimax-m2.7"
       ]
     },
     "lfm2": {
@@ -1085,7 +1100,12 @@
         "claude-opus-4-8",
         "claude-sonnet-5",
         "claude-fable-5",
-        "claude-mythos-5"
+        "claude-mythos-5",
+        "claude-haiku-5-5",
+        "claude-fable-5-1",
+        "claude-mythos-5-1",
+        "claude-opus-5-5",
+        "claude-sonnet-5-5"
       ]
     },
     "gpt": {
@@ -1099,7 +1119,14 @@
         "gpt-5.6",
         "gpt-5.6-sol",
         "gpt-5.6-terra",
-        "gpt-5.6-luna"
+        "gpt-5.6-luna",
+        "gpt-6",
+        "gpt-6-astra",
+        "gpt-6-sol",
+        "gpt-6-luna",
+        "gpt-6.1",
+        "gpt-6.1-sol",
+        "gpt-6.1-terra"
       ]
     },
     "gpt_oss": {
@@ -1151,7 +1178,12 @@
         "nemotron-cascade-2",
         "nemotron-labs",
         "nvidia/nemotron",
-        "nvidia-nemotron"
+        "nvidia-nemotron",
+        "nemotron-3-ultra",
+        "nemotron-3.5-lightning",
+        "nemotron-3.5-bolt-74b",
+        "nemotron-3.5-super-vl",
+        "nemotron-3.5-transcribe"
       ]
     },
     "qwen3_5_agentic": {
@@ -1207,7 +1239,10 @@
         "zai-org/glm-5.1",
         "zai-org/glm-5.2",
         "z-ai/glm-5.1",
-        "z-ai/glm-5.2"
+        "z-ai/glm-5.2",
+        "glm-5.3",
+        "zai-org/glm-5.3",
+        "z-ai/glm-5.3"
       ]
     },
     "minimax_m3": {
@@ -1227,7 +1262,10 @@
       "patterns": [
         "minimax-m3",
         "minimaxai/minimax-m3",
-        "minimax/minimax-m3"
+        "minimax/minimax-m3",
+        "minimax-m3.1",
+        "minimaxai/minimax-m3.1-flash",
+        "minimax/minimax-m3.1-flash"
       ]
     },
     "hunyuan3": {
@@ -1418,7 +1456,7 @@
       ]
     },
     "bonsai": {
-      "description": "Prism Bonsai-27B low-bit Qwen3.6 derivative; parent chat template, separate model identity.",
+      "description": "Prism Bonsai low-bit Qwen3.5/3.6/3.8 derivatives (Bonsai-27B, Ternary-Bonsai-27B, Bonsai 2); parent chat template, separate model identities. The Qwen3-derived Bonsai-8B/4B/1.7B and Ternary-Bonsai-8B/4B/1.7B use the qwen3 block.",
       "message_format": "im_start_end",
       "system_prefix": "<|im_start|>system\n",
       "system_suffix": "<|im_end|>\n",
@@ -1447,7 +1485,13 @@
         "top_k": 20
       },
       "patterns": [
-        "bonsai-27b"
+        "bonsai-27b",
+        "bonsai-2-",
+        "ternary-bonsai-2-",
+        "ternary-bonsai-27b",
+        "prism-ml/bonsai-27b",
+        "prism-ml/ternary-bonsai-27b",
+        "prism-ml/ternary-bonsai-2-"
       ]
     },
     "spark2_5": {