abstractcore 2.26.3__tar.gz → 2.31.0__tar.gz

{abstractcore-2.26.3 → abstractcore-2.31.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: abstractcore
-Version: 2.26.3
+Version: 2.31.0
 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>
@@ -49,10 +49,10 @@ Requires-Dist: uvicorn[standard]<1.0.0,>=0.23.0
 Requires-Dist: python-multipart<1.0.0,>=0.0.6
 Requires-Dist: sse-starlette<2.0.0,>=1.6.0
 Requires-Dist: cryptography>=41.0.0
-Requires-Dist: abstractvoice>=0.15.0
-Requires-Dist: abstractvision>=0.3.33
-Requires-Dist: abstractmusic>=0.1.16; python_version >= "3.10"
-Requires-Dist: abstract3d>=0.1.0; python_version >= "3.10"
+Requires-Dist: abstractvoice>=0.18.2
+Requires-Dist: abstractvision>=0.3.34
+Requires-Dist: abstractmusic>=0.1.17; python_version >= "3.10"
+Requires-Dist: abstract3d>=0.3.2; python_version >= "3.10"
 Provides-Extra: apple
 Requires-Dist: mlx<1.0.0,>=0.32.2; extra == "apple"
 Requires-Dist: mlx-lm<1.0.0,>=0.31.3; extra == "apple"
@@ -68,10 +68,10 @@ Requires-Dist: accelerate>=1.0.0; extra == "apple"
 Requires-Dist: sentence-transformers<6.0.0,>=5.1.0; extra == "apple"
 Requires-Dist: numpy<3.0.0,>=2.1.0; extra == "apple"
 Requires-Dist: Pillow<13.0.0,>=12.1.1; extra == "apple"
-Requires-Dist: abstractvoice[all-apple]>=0.15.0; extra == "apple"
+Requires-Dist: abstractvoice[all-apple]>=0.18.2; extra == "apple"
 Requires-Dist: omnivoice>=0.1.5; extra == "apple"
-Requires-Dist: abstractvision[all-apple]>=0.3.33; extra == "apple"
-Requires-Dist: abstractmusic[all-apple]>=0.1.16; extra == "apple"
+Requires-Dist: abstractvision[all-apple]>=0.3.34; extra == "apple"
+Requires-Dist: abstractmusic[all-apple]>=0.1.17; extra == "apple"
 Requires-Dist: playwright<2.0.0,>=1.49.0; extra == "apple"
 Provides-Extra: gpu
 Requires-Dist: vllm<1.0.0,>=0.6.0; sys_platform == "linux" and extra == "gpu"
@@ -85,10 +85,10 @@ Requires-Dist: llama-cpp-python<1.0.0,>=0.3.23; sys_platform != "win32" and extr
 Requires-Dist: sentence-transformers<6.0.0,>=5.1.0; extra == "gpu"
 Requires-Dist: numpy<3.0.0,>=1.20.0; python_version < "3.13" and extra == "gpu"
 Requires-Dist: numpy<3.0.0,>=2.1.0; python_version >= "3.13" and extra == "gpu"
-Requires-Dist: abstractvoice[all-gpu]>=0.15.0; extra == "gpu"
+Requires-Dist: abstractvoice[all-gpu]>=0.18.2; extra == "gpu"
 Requires-Dist: omnivoice>=0.1.5; extra == "gpu"
-Requires-Dist: abstractvision[all-gpu]>=0.3.33; extra == "gpu"
-Requires-Dist: abstractmusic[all-gpu]>=0.1.16; extra == "gpu"
+Requires-Dist: abstractvision[all-gpu]>=0.3.34; extra == "gpu"
+Requires-Dist: abstractmusic[all-gpu]>=0.1.17; extra == "gpu"
 Requires-Dist: playwright<2.0.0,>=1.49.0; extra == "gpu"
 Provides-Extra: remote
 Provides-Extra: openai
@@ -133,11 +133,11 @@ Requires-Dist: sentence-transformers<6.0.0,>=5.1.0; extra == "embeddings"
 Requires-Dist: numpy<3.0.0,>=1.20.0; python_version < "3.13" and extra == "embeddings"
 Requires-Dist: numpy<3.0.0,>=2.1.0; python_version >= "3.13" and extra == "embeddings"
 Provides-Extra: vision-diffusers
-Requires-Dist: abstractvision[huggingface]>=0.3.31; extra == "vision-diffusers"
+Requires-Dist: abstractvision[huggingface]>=0.3.34; extra == "vision-diffusers"
 Provides-Extra: vision-sdcpp
-Requires-Dist: abstractvision[sdcpp]>=0.3.31; extra == "vision-sdcpp"
+Requires-Dist: abstractvision[sdcpp]>=0.3.34; extra == "vision-sdcpp"
 Provides-Extra: vision-local
-Requires-Dist: abstractvision[local]>=0.3.31; extra == "vision-local"
+Requires-Dist: abstractvision[local]>=0.3.34; extra == "vision-local"
 Provides-Extra: browser
 Requires-Dist: playwright<2.0.0,>=1.49.0; extra == "browser"
 Provides-Extra: pdf-pymupdf-commercial
@@ -193,6 +193,10 @@ Dynamic: license-file
 Unified LLM Interface
 > Write once, run everywhere
 
+What's new in 2.31.0: streamed tool calls carry the declared MCP tool name, the MLX prefix cache
+survives per-call system riders in tool loops, and `POST /v1/voice/clone` takes `language`. See the
+[changelog](CHANGELOG.md#2310---2026-10-27).
+
 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.
@@ -248,6 +252,20 @@ For AI assistants and doc-indexing tools, this repository also publishes
 [`llms.txt`](llms.txt) and [`llms-full.txt`](llms-full.txt). MCP clients such as
 Context7 can query the public documentation directly as well.
 
+## For AI agents
+
+Register and use the [`abstractcore` agent skill](https://github.com/lpalbou/AbstractCore/tree/main/skills/abstractcore)
+to write integrations and operate the Python API, CLI and HTTP server. Start by reading
+[`skills/abstractcore/SKILL.md`](https://raw.githubusercontent.com/lpalbou/AbstractCore/main/skills/abstractcore/SKILL.md);
+keep its `references/` folder alongside it when copying the skill into your agent's skill directory.
+In a source checkout, tell another agent: **Read `skills/abstractcore/SKILL.md` and use its
+references for this task.**
+
+The skill covers provider and artifact selection, token/time/cost limits, host-controlled
+tools, sessions, typed output, media and diagnostics. The canonical
+[integration and operations guide](docs/agent-usage.md) contains the same supported workflows;
+source contributors should also read [Contributing](CONTRIBUTING.md).
+
 ## Why AbstractCore
 
 Many libraries can call an LLM. AbstractCore is for the messy middle of real
@@ -455,7 +473,7 @@ You can also persist settings (including API keys) via the config CLI:
 - Embeddings and semantic search → [Embeddings](docs/embeddings.md)
 - Observability: global event bus + interaction traces → [Architecture](docs/architecture.md), [API Reference (Events)](docs/api-reference.md#eventtype), [Interaction Tracing](docs/interaction-tracing.md)
 - MCP (Model Context Protocol): discover tools from MCP servers (HTTP/stdio) → [MCP](docs/mcp.md)
-- Email: one IMAP/SMTP account (password or OAuth2; servers found from the address), encrypted credentials, recipient policy, send limits, read-only mailbox and email tools → [Email](docs/email.md)
+- Email: one IMAP/SMTP account (password or OAuth2; servers found from the address), encrypted credentials, recipient policy, send limits, Read only or Read and write access, and email tools → [Email](docs/email.md)
 - OpenAI-compatible server: one `/v1` gateway for chat + optional `/v1/images/*` and `/v1/audio/*` endpoints → [Server](docs/server.md)
 
 ## Tool calling (passthrough by default)
@@ -512,7 +530,7 @@ Notes:
 - **Images**: use a vision-capable model, or configure **vision fallback** for text-only models (`abstractcore --config`; `abstractcore --set-vision-provider PROVIDER MODEL`).
 - **Video**: `video_policy="auto"` (default) uses native video when supported, otherwise samples frames (requires `ffmpeg`/`ffprobe`) and routes them through image/vision handling (so you still need a vision-capable model or vision fallback configured).
 - **Audio**: use an audio-capable model, or set `audio_policy="auto"`/`"speech_to_text"` for speech-to-text through AbstractVoice (part of the light install).
-  `abstractvoice` 0.11.0+ can install its base plugin path on Python 3.9 without OmniVoice, torch, or torchaudio. Local voice engines and clone backends are part of explicit local aggregate profiles.
+  `abstractvoice` installs its base plugin path on Python 3.9 without OmniVoice, torch, or torchaudio. Local voice engines and clone backends are part of explicit local aggregate profiles.
 
 Configure defaults (optional):
 
@@ -645,11 +663,30 @@ stt_models = llm.voice.list_stt_models()
 music_models = llm.capabilities.list_models("music", task="text_to_music")
 ```
 
+Voices are navigable by engine and by language, and you can create your own:
+
+```python
+french = llm.voice.list_voices(language="fr")      # native French voices first
+piper = llm.voice.list_voices(provider="piper")    # one engine's voices
+languages = llm.voice.list_languages(kind="tts")   # languages per engine, no model loaded
+# Each record: voice_id (pass as voice=), name, kind (profile|clone), provider, model, languages, language
+
+mira = llm.voice.design_voice("A calm, warm female voice in her thirties.", name="Mira")
+wav = llm.voice.tts("Good morning.", voice=mira["voice_id"], one_shot=True)
+llm.voice.rename_cloned_voice(mira["voice_id"], "Mira (narrator)")
+llm.voice.delete_cloned_voice(mira["voice_id"])
+```
+
+`one_shot=True` synthesizes the whole text as one generation, so the take keeps one
+voice from start to end (Qwen3-TTS voices, up to 1000 characters).
+
 The HTTP server exposes equivalent discovery at
-`/v1/vision/providers/`, `/v1/vision/models`, `/v1/audio/voices`,
-`/v1/audio/speech/models`, `/v1/audio/transcriptions/models`, and
-`/v1/voice/clone/providers`, plus `/v1/audio/music/providers` and
-`/v1/audio/music/models`.
+`/v1/vision/providers/`, `/v1/vision/models`, `/v1/audio/voices` (with `language`),
+`/v1/audio/languages`, `/v1/audio/speech/models`, `/v1/audio/transcriptions/models`,
+`/v1/voice/clone/providers` and `/v1/voice/clones`, plus `/v1/audio/music/providers`
+and `/v1/audio/music/models`. Voices are created and managed with
+`POST /v1/voice/clone`, `POST /v1/voice/design` and
+`GET|PATCH|DELETE /v1/voice/clones/{voice_id}`; see [Server](docs/server.md).
 Generated media HTTP routes include `/v1/images/generations`,
 `/v1/images/edits`, `/v1/videos/generations`, `/v1/videos/edits`, and
 async polling routes under `/v1/vision/jobs/images/*` and

{abstractcore-2.26.3 → abstractcore-2.31.0}/README.md

@@ -9,6 +9,10 @@
 Unified LLM Interface
 > Write once, run everywhere
 
+What's new in 2.31.0: streamed tool calls carry the declared MCP tool name, the MLX prefix cache
+survives per-call system riders in tool loops, and `POST /v1/voice/clone` takes `language`. See the
+[changelog](CHANGELOG.md#2310---2026-10-27).
+
 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.
@@ -64,6 +68,20 @@ For AI assistants and doc-indexing tools, this repository also publishes
 [`llms.txt`](llms.txt) and [`llms-full.txt`](llms-full.txt). MCP clients such as
 Context7 can query the public documentation directly as well.
 
+## For AI agents
+
+Register and use the [`abstractcore` agent skill](https://github.com/lpalbou/AbstractCore/tree/main/skills/abstractcore)
+to write integrations and operate the Python API, CLI and HTTP server. Start by reading
+[`skills/abstractcore/SKILL.md`](https://raw.githubusercontent.com/lpalbou/AbstractCore/main/skills/abstractcore/SKILL.md);
+keep its `references/` folder alongside it when copying the skill into your agent's skill directory.
+In a source checkout, tell another agent: **Read `skills/abstractcore/SKILL.md` and use its
+references for this task.**
+
+The skill covers provider and artifact selection, token/time/cost limits, host-controlled
+tools, sessions, typed output, media and diagnostics. The canonical
+[integration and operations guide](docs/agent-usage.md) contains the same supported workflows;
+source contributors should also read [Contributing](CONTRIBUTING.md).
+
 ## Why AbstractCore
 
 Many libraries can call an LLM. AbstractCore is for the messy middle of real
@@ -271,7 +289,7 @@ You can also persist settings (including API keys) via the config CLI:
 - Embeddings and semantic search → [Embeddings](docs/embeddings.md)
 - Observability: global event bus + interaction traces → [Architecture](docs/architecture.md), [API Reference (Events)](docs/api-reference.md#eventtype), [Interaction Tracing](docs/interaction-tracing.md)
 - MCP (Model Context Protocol): discover tools from MCP servers (HTTP/stdio) → [MCP](docs/mcp.md)
-- Email: one IMAP/SMTP account (password or OAuth2; servers found from the address), encrypted credentials, recipient policy, send limits, read-only mailbox and email tools → [Email](docs/email.md)
+- Email: one IMAP/SMTP account (password or OAuth2; servers found from the address), encrypted credentials, recipient policy, send limits, Read only or Read and write access, and email tools → [Email](docs/email.md)
 - OpenAI-compatible server: one `/v1` gateway for chat + optional `/v1/images/*` and `/v1/audio/*` endpoints → [Server](docs/server.md)
 
 ## Tool calling (passthrough by default)
@@ -328,7 +346,7 @@ Notes:
 - **Images**: use a vision-capable model, or configure **vision fallback** for text-only models (`abstractcore --config`; `abstractcore --set-vision-provider PROVIDER MODEL`).
 - **Video**: `video_policy="auto"` (default) uses native video when supported, otherwise samples frames (requires `ffmpeg`/`ffprobe`) and routes them through image/vision handling (so you still need a vision-capable model or vision fallback configured).
 - **Audio**: use an audio-capable model, or set `audio_policy="auto"`/`"speech_to_text"` for speech-to-text through AbstractVoice (part of the light install).
-  `abstractvoice` 0.11.0+ can install its base plugin path on Python 3.9 without OmniVoice, torch, or torchaudio. Local voice engines and clone backends are part of explicit local aggregate profiles.
+  `abstractvoice` installs its base plugin path on Python 3.9 without OmniVoice, torch, or torchaudio. Local voice engines and clone backends are part of explicit local aggregate profiles.
 
 Configure defaults (optional):
 
@@ -461,11 +479,30 @@ stt_models = llm.voice.list_stt_models()
 music_models = llm.capabilities.list_models("music", task="text_to_music")
 ```
 
+Voices are navigable by engine and by language, and you can create your own:
+
+```python
+french = llm.voice.list_voices(language="fr")      # native French voices first
+piper = llm.voice.list_voices(provider="piper")    # one engine's voices
+languages = llm.voice.list_languages(kind="tts")   # languages per engine, no model loaded
+# Each record: voice_id (pass as voice=), name, kind (profile|clone), provider, model, languages, language
+
+mira = llm.voice.design_voice("A calm, warm female voice in her thirties.", name="Mira")
+wav = llm.voice.tts("Good morning.", voice=mira["voice_id"], one_shot=True)
+llm.voice.rename_cloned_voice(mira["voice_id"], "Mira (narrator)")
+llm.voice.delete_cloned_voice(mira["voice_id"])
+```
+
+`one_shot=True` synthesizes the whole text as one generation, so the take keeps one
+voice from start to end (Qwen3-TTS voices, up to 1000 characters).
+
 The HTTP server exposes equivalent discovery at
-`/v1/vision/providers/`, `/v1/vision/models`, `/v1/audio/voices`,
-`/v1/audio/speech/models`, `/v1/audio/transcriptions/models`, and
-`/v1/voice/clone/providers`, plus `/v1/audio/music/providers` and
-`/v1/audio/music/models`.
+`/v1/vision/providers/`, `/v1/vision/models`, `/v1/audio/voices` (with `language`),
+`/v1/audio/languages`, `/v1/audio/speech/models`, `/v1/audio/transcriptions/models`,
+`/v1/voice/clone/providers` and `/v1/voice/clones`, plus `/v1/audio/music/providers`
+and `/v1/audio/music/models`. Voices are created and managed with
+`POST /v1/voice/clone`, `POST /v1/voice/design` and
+`GET|PATCH|DELETE /v1/voice/clones/{voice_id}`; see [Server](docs/server.md).
 Generated media HTTP routes include `/v1/images/generations`,
 `/v1/images/edits`, `/v1/videos/generations`, `/v1/videos/edits`, and
 async polling routes under `/v1/vision/jobs/images/*` and

{abstractcore-2.26.3 → abstractcore-2.31.0}/abstractcore/architectures/detection.py

@@ -280,6 +280,9 @@ def _partial_match_is_prefix_aligned(name_lower: str, key_lower: str) -> bool:
     return longer[len(shorter)] in _MODEL_NAME_BOUNDARY_CHARS
 
 
+_HF_CACHE_SNAPSHOT_RE = re.compile(r"models--([^/]+?)--([^/]+)/snapshots/[^/]+(?:/(?:.*/)?([^/]+))?/?$")
+
+
 def resolve_model_alias(model_name: str, models: Dict[str, Any]) -> str:
     """
     Resolve a model name to its canonical name by checking aliases.
@@ -310,6 +313,21 @@ def resolve_model_alias(model_name: str, models: Dict[str, Any]) -> str:
     if model_name in models:
         _resolved_aliases_cache[model_name] = model_name
         return model_name
+
+    # A Hugging Face cache snapshot PATH (`.../models--org--name/snapshots/<rev>[/file]`)
+    # names its repo only in the `models--` segment; the tail is a revision hash (or a
+    # file). Resolve it as the repo id (then the file name), so a request made through
+    # the local path gets the same registry entry as one made through the repo id.
+    snapshot = _HF_CACHE_SNAPSHOT_RE.search(str(model_name or "").replace("\\", "/"))
+    if snapshot:
+        repo_id = f"{snapshot.group(1)}/{snapshot.group(2)}"
+        for candidate in (repo_id, snapshot.group(3)):
+            if not candidate:
+                continue
+            resolved = resolve_model_alias(candidate, models)
+            if resolved in models:
+                _resolved_aliases_cache[model_name] = resolved
+                return resolved
     direct_lower_match = canonical_lower_map.get(str(model_name or "").strip().lower())
     if direct_lower_match is not None:
         _resolved_aliases_cache[model_name] = direct_lower_match

{abstractcore-2.26.3 → abstractcore-2.31.0}/abstractcore/architectures/thinking_controls.py

@@ -17,6 +17,7 @@ text. The typed object form makes the surface kind explicit:
   "low_effort_template_kwarg": "low_effort",
   "request_param": "reasoning_effort",
   "effort_template_kwarg": "reasoning_effort",
+  "effort_disable_value": "no_think",
   "effort_system_lines": {"xhigh": "…", "medium": "", "low": "…"}
 }
 ```
@@ -41,6 +42,13 @@ Semantics:
 - ``effort_template_kwarg``: string chat-template variable selecting a reasoning-effort
   level (Qwen3.8 ``reasoning_effort``), sent via ``chat_template_kwargs`` by providers whose
   backend renders the template.
+- ``effort_disable_value``: the value of ``effort_template_kwarg`` that turns reasoning OFF,
+  for templates whose ONE effort variable is also the on/off switch and that have no
+  boolean ``template_kwarg`` (HY V3 / Hy-Memory: ``reasoning_effort`` in
+  ``high | low | no_think``). Providers that render the template send this value for
+  ``thinking="off"`` and, for a bare ``thinking="on"``, the model's preferred declared
+  ``reasoning_levels`` entry (``medium``, then ``high``, then ``low``) — never a boolean
+  ``enable_thinking`` the template would silently ignore. Not a level name.
 - ``effort_system_lines``: level -> exact instruction sentence the model's own chat template
   prepends to the system block for that effort level. Lets locally-serializing providers
   (MLX) reproduce the template's effort control byte-for-byte. An EMPTY string value means
@@ -62,6 +70,7 @@ _SURFACE_KEYS = (
     "low_effort_template_kwarg",
     "request_param",
     "effort_template_kwarg",
+    "effort_disable_value",
 )
 
 _MAPPING_SURFACE_KEYS = ("effort_system_lines",)
@@ -78,11 +87,21 @@ class ThinkingControlSurfaces:
     low_effort_template_kwarg: Optional[str] = None
     request_param: Optional[str] = None
     effort_template_kwarg: Optional[str] = None
+    effort_disable_value: Optional[str] = None
     effort_system_lines: Optional[Mapping[str, str]] = None
 
     def any_declared(self) -> bool:
         return any(getattr(self, f.name) is not None for f in fields(self))
 
+    def effort_is_switch(self) -> bool:
+        """True when the effort template variable is ALSO the on/off switch.
+
+        Declared by ``effort_disable_value`` together with ``effort_template_kwarg`` and
+        no boolean ``template_kwarg``: thinking off/on are both expressed as values of
+        the one effort variable (HY V3 ``reasoning_effort=no_think`` / ``low`` / ``high``).
+        """
+        return bool(self.effort_template_kwarg and self.effort_disable_value and not self.template_kwarg)
+
 
 def _coerce_surface_value(value: Any) -> Optional[str]:
     if not isinstance(value, str):
@@ -168,7 +187,38 @@ def resolve_thinking_control_surfaces(
     return ThinkingControlSurfaces(**merged)
 
 
+_EFFORT_ON_PREFERENCE = ("medium", "high", "low")
+
+
+def effort_switch_value(
+    surfaces: ThinkingControlSurfaces,
+    *,
+    enabled: Optional[bool],
+    level: Optional[str],
+    declared_levels: Optional[list] = None,
+) -> Optional[str]:
+    """Value for an effort-switch template variable (see ``effort_is_switch``).
+
+    - ``enabled is False`` -> ``effort_disable_value``;
+    - an explicit ``level`` -> that level when declared (or when no levels are declared);
+    - a bare on -> the first of medium/high/low present in ``declared_levels``.
+    Returns None when the surface is not an effort switch or no honest value exists
+    (callers must then decline the claim, never guess).
+    """
+    if not surfaces.effort_is_switch():
+        return None
+    declared = [str(x).strip().lower() for x in (declared_levels or []) if isinstance(x, str) and x.strip()]
+    if enabled is False:
+        return surfaces.effort_disable_value
+    if isinstance(level, str) and level:
+        return level if (not declared or level in declared) else None
+    if enabled is True:
+        return next((x for x in _EFFORT_ON_PREFERENCE if x in declared), None)
+    return None
+
+
 __all__ = [
     "ThinkingControlSurfaces",
+    "effort_switch_value",
     "resolve_thinking_control_surfaces",
 ]

{abstractcore-2.26.3 → abstractcore-2.31.0}/abstractcore/assets/architecture_formats.json

@@ -1248,6 +1248,51 @@
         "tencent/hy3"
       ]
     },
+    "hy_v3": {
+      "description": "Tencent HY V3 compact MoE (HYV3ForCausalLM; model_type 'hy_v3', GGUF general.architecture 'hy_v3'): Tencent Hy-Memory-30B-A3B long-term-memory writer (~30B total / ~3B active, 128 routed + 1 shared experts; October 2026). BOS then bare system text, <｜hy_User｜>/<｜hy_Assistant｜> role tokens, plain <think> tags, ONE reasoning_effort template variable (high|low|no_think; unset or unknown = no_think, so thinking is OFF by default) and <tool_calls><tool_call>name<tool_sep><arg_key>/<arg_value> tool blocks. Distinct from the 295B Hy3 (hunyuan3), which uses ':opensource'-suffixed tokens",
+      "message_format": "special_tokens",
+      "system_prefix": "<｜hy_begin▁of▁sentence｜>",
+      "system_suffix": "",
+      "user_prefix": "<｜hy_User｜>",
+      "user_suffix": "",
+      "assistant_prefix": "<｜hy_Assistant｜>",
+      "assistant_suffix": "<eos:6124c78e>",
+      "output_wrappers": {
+        "end": "<eos:6124c78e>"
+      },
+      "tool_format": "glm_xml",
+      "tool_prefix": "<tool_calls>",
+      "tool_calling_format": "<tool_calls>\n<tool_call>function_name<tool_sep>\n<arg_key>key</arg_key>\n<arg_value>value</arg_value>\n</tool_call>\n</tool_calls>",
+      "default_tool_support": "native",
+      "thinking_tags": [
+        "<think>",
+        "</think>"
+      ],
+      "thinking_output_field": "reasoning_content",
+      "thinking_control": {
+        "effort_template_kwarg": "reasoning_effort",
+        "effort_disable_value": "no_think",
+        "request_param": "reasoning_effort"
+      },
+      "reasoning_support": true,
+      "reasoning_levels": [
+        "low",
+        "high"
+      ],
+      "inference_parameters": {
+        "temperature": 0.7,
+        "top_p": 1.0,
+        "repeat_penalty": 1.0
+      },
+      "patterns": [
+        "tencent/hy-memory",
+        "/hy-memory",
+        "hy-memory-30b",
+        "hy-memory-step",
+        "hy3a3b",
+        "hy3a3b-memrl1009-step250"
+      ]
+    },
     "longcat": {
       "description": "Meituan LongCat 2.0 large-scale MoE with LongCat Sparse Attention (July 2026; dedicated <longcat_*> role/thinking/tool special tokens — serve via OpenAI-compatible stacks)",
       "message_format": "openai_chat",

{abstractcore-2.26.3 → abstractcore-2.31.0}/abstractcore/assets/model_capabilities.json

@@ -1086,6 +1086,44 @@
         ]
       }
     },
+    "claude-haiku-5-5": {
+      "max_output_tokens": 128000,
+      "tool_support": "native",
+      "structured_output": "native",
+      "parallel_tools": true,
+      "max_tools": -1,
+      "vision_support": true,
+      "audio_support": false,
+      "thinking_support": true,
+      "thinking_budget": true,
+      "thinking_control_mode": "adaptive",
+      "max_effort_supported": true,
+      "reasoning_levels": [
+        "low",
+        "medium",
+        "high",
+        "xhigh"
+      ],
+      "notes": "Anthropic Haiku tier, 5th generation (released 2026-10-07), for high-volume, latency-sensitive work (classification, routing, extraction, subagents). Official docs: text and image input, text output, 1M context, 128K max output (300K on the Message Batches API with the output-300k-2026-03-24 beta header), adaptive thinking on by default, effort parameter with all five levels low|medium|high|xhigh|max (default medium; max is not in the unified thinking vocabulary, see max_effort_supported). thinking:{type:disabled} is accepted at effort high or below and returns 400 at xhigh/max. Omit temperature, top_p and top_k. Newer tokenizer (about 30% more tokens than Haiku 4.5 for the same text); 512-token prompt-cache minimum. Observed via OpenAI-compatible proxy 2026-10-25: /v1/chat/completions rejects reasoning_effort 'none' and accepts low|medium|high|xhigh|max.",
+      "source": "Anthropic docs: platform.claude.com/docs/en/models/haiku-5-5/overview, /docs/en/about-claude/models/overview and /docs/en/build-with-claude/effort (accessed 2026-10-25); reasoning_effort acceptance observed via OpenAI-compatible proxy 2026-10-25",
+      "canonical_name": "claude-haiku-5-5",
+      "aliases": [
+        "anthropic/claude-haiku-5-5"
+      ],
+      "max_tokens": 1000000,
+      "video_support": false,
+      "video_input_mode": "frames",
+      "capability_routes": {
+        "input": [
+          "text",
+          "image",
+          "video"
+        ],
+        "output": [
+          "text"
+        ]
+      }
+    },
     "cogito:3b": {
       "max_output_tokens": 4096,
       "tool_support": "native",
@@ -5643,6 +5681,41 @@
         ]
       }
     },
+    "gpt-6-luna": {
+      "max_output_tokens": 128000,
+      "tool_support": "native",
+      "structured_output": "native",
+      "parallel_tools": true,
+      "max_tools": -1,
+      "vision_support": true,
+      "audio_support": false,
+      "thinking_support": true,
+      "token_param_name": "max_completion_tokens",
+      "reasoning_levels": [
+        "none",
+        "low",
+        "medium",
+        "high",
+        "xhigh"
+      ],
+      "max_effort_supported": true,
+      "notes": "GPT-6 family efficient model for focused, high-volume tasks (successor of gpt-5.6-luna). Official model page: 1,050,000-token context window, 922,000 max input, 128,000 max output; reasoning token support; reasoning.effort none|low|medium|high|xhigh|max, default medium (max is not in the unified thinking vocabulary, see max_effort_supported); knowledge cutoff 2026-05-18; text+image input, text output; streaming, structured outputs, function calling, prompt caching; Chat Completions, Responses and Batch endpoints. The page notes Chat Completions supports function calling only with reasoning_effort none and recommends the Responses API for tools. max_tokens records the published 1,050,000 window: the 400,000-token Codex-backend session ceiling wire-verified for gpt-5.6-sol is NOT verified for GPT-6. Observed via OpenAI-compatible proxy 2026-10-25: accepts reasoning_effort none|low|medium|high and rejects 'minimal'.",
+      "source": "OpenAI model page developers.openai.com/api/docs/models/gpt-6-luna (accessed 2026-10-25); reasoning_effort acceptance observed via OpenAI-compatible proxy 2026-10-25",
+      "canonical_name": "gpt-6-luna",
+      "aliases": [],
+      "max_tokens": 1050000,
+      "video_support": false,
+      "video_input_mode": "frames",
+      "capability_routes": {
+        "input": [
+          "text",
+          "image"
+        ],
+        "output": [
+          "text"
+        ]
+      }
+    },
     "text-embedding-ada-002": {
       "max_output_tokens": 0,
       "tool_support": "none",
@@ -9549,6 +9622,71 @@
         ]
       }
     },
+    "hy-memory-30b-a3b": {
+      "max_output_tokens": 16384,
+      "tool_support": "native",
+      "structured_output": "prompted",
+      "parallel_tools": true,
+      "vision_support": false,
+      "audio_support": false,
+      "video_support": false,
+      "thinking_support": true,
+      "thinking_tags": [
+        "<think>",
+        "</think>"
+      ],
+      "thinking_output_field": "reasoning_content",
+      "thinking_control": {
+        "effort_template_kwarg": "reasoning_effort",
+        "effort_disable_value": "no_think",
+        "request_param": "reasoning_effort"
+      },
+      "reasoning_levels": [
+        "low",
+        "high"
+      ],
+      "architecture": "hy_v3",
+      "total_parameters": "30B",
+      "active_parameters": "3B",
+      "experts": 128,
+      "shared_experts": 1,
+      "experts_activated": 8,
+      "license": "Apache-2.0",
+      "release_date": "2026-10-15",
+      "notes": "Specialty seems to be a long-term memory writer: Tencent Hunyuan trained it to execute the WRITE path inside external memory frameworks (drop-in writer LLM for Mem0, MemOS and Hy-Memory; transfers to EverMemOS) — deciding what to retain, how precisely to represent it and how new evidence updates the memory bank. HYV3ForCausalLM (model_type 'hy_v3'): 48 layers, hidden 2048, GQA 32 heads / 4 KV heads (head_dim 128, QK-norm), layer 0 dense (intermediate 6912), layers 1-47 MoE with 128 routed experts (8 activated) + 1 shared expert, sigmoid router with selection-only expert bias, route_norm and router_scaling_factor 2.826; no MTP layers; 120,832-token vocab; default RoPE (theta 11,158,840); config max_position_embeddings 262,144. CHECKPOINT QUIRK: routed and shared expert tensors are 768 wide (config 'expert_hidden_dim': 768) while config 'moe_intermediate_size' says 1536, which is what transformers' hy_v3 (5.8) reads — loaders must size experts from expert_hidden_dim. Chat template: BOS then bare system text, <｜hy_User｜>/<｜hy_Assistant｜>, eos <eos:6124c78e>; ONE 'reasoning_effort' template variable (high | low | no_think; unset or unknown renders no_think, so thinking is OFF by default; no_think pre-closes an empty <think></think> after the assistant token, low/high open <think>). Tool calls: <tool_calls><tool_call>name<tool_sep><arg_key>k</arg_key><arg_value>v</arg_value></tool_call></tool_calls> (non-string values JSON-encoded); tool results in <tool_responses><tool_response>. Card quickstart: vLLM (tested v0.31.0), chat_template_kwargs {\"reasoning_effort\": \"no_think\"}, temperature 0.0 and max_tokens 512 for memory extraction; the GGUF card's llama-server defaults are --temp 0.7 --top-k 0 --top-p 1.0 --repeat-penalty 1.0 --jinja with reasoning_effort no_think (inference_parameters mirror temp/top_p; top_k 0 = disabled is left to the backend). No output cap is published: max_output_tokens 16,384 leaves room for a reasoning_effort=high trace plus the memory payload while bounding runaway local generations — raise per deployment. Artifacts: tencent BF16 safetensors (51 shards, 60.1 GB), tencent FP8 (F8_E4M3, vLLM/SGLang only — no Apple-silicon path), tencent GGUF (Q4_K_M-imat 18.2 GB, Q6_K-imat 24.7 GB, Q8_0 32 GB; llama.cpp arch hy_v3), and mlx-community 4-bit / 8-bit MLX conversions that bundle the model class (hy_v3.py via config 'model_file') because mlx-lm ships no hy_v3 module. Paper arXiv 2610.22196; code github.com/Tencent-Hunyuan/Hy-Memory.",
+      "source": "Hugging Face model cards, config.json, tokenizer.json, chat_template.jinja and safetensors index (tencent/Hy-Memory-30B-A3B @64a460a9, tencent/Hy-Memory-30B-A3B-FP8, tencent/Hy-Memory-30B-A3B-GGUF; HF API createdAt 2026-10-15), accessed 2026-10-21",
+      "canonical_name": "hy-memory-30b-a3b",
+      "aliases": [
+        "tencent/Hy-Memory-30B-A3B",
+        "tencent/Hy-Memory-30B-A3B-FP8",
+        "tencent/Hy-Memory-30B-A3B-GGUF",
+        "mlx-community/Hy-Memory-30B-A3B-4bit",
+        "mlx-community/Hy-Memory-30B-A3B-8bit",
+        "hy3a3b-memrl1009-step250-Q4_K_M-imat.gguf",
+        "hy3a3b-memrl1009-step250-Q6_K-imat.gguf",
+        "Hy-Memory-Step250-Q8_0.gguf",
+        "hy3a3b-memrl1009-step250",
+        "hy-memory-step250",
+        "hy-memory-30b-a3b-fp8",
+        "hy-memory-30b-a3b-gguf"
+      ],
+      "max_tokens": 262144,
+      "inference_parameters": {
+        "temperature": 0.7,
+        "top_p": 1.0,
+        "repeat_penalty": 1.0
+      },
+      "max_tools": -1,
+      "video_input_mode": "none",
+      "capability_routes": {
+        "input": [
+          "text"
+        ],
+        "output": [
+          "text"
+        ]
+      }
+    },
     "minimax-m3": {
       "max_output_tokens": 8192,
       "tool_support": "native",

{abstractcore-2.26.3 → abstractcore-2.31.0}/abstractcore/assets/model_downloads_catalog.json

@@ -3162,6 +3162,71 @@
     }
    ]
   },
+  {
+   "id": "hy-memory-30b-a3b",
+   "family": "hy-memory",
+   "display_name": "Hy-Memory 30B-A3B",
+   "vendor": "Tencent",
+   "params_total": 30064725888,
+   "params_active": 3000000000,
+   "license": "apache-2.0",
+   "capabilities_key": "hy-memory-30b-a3b",
+   "tags": [
+    "chat",
+    "thinking",
+    "tools",
+    "moe"
+   ],
+   "starter": false,
+   "kv_geometry": {
+    "n_layers": 48,
+    "n_kv_heads": 4,
+    "head_dim": 128,
+    "source": "config.json of tencent/Hy-Memory-30B-A3B@64a460a94d54: 48 full-attention layers, GQA 32 query / 4 KV heads"
+   },
+   "artifacts": [
+    {
+     "provider": "mlx",
+     "artifact": "mlx-community/Hy-Memory-30B-A3B-4bit",
+     "quant": "4bit",
+     "download_bytes": 16929648349,
+     "size_source": "catalog",
+     "verified": true,
+     "upstream": {
+      "method": "hf_api",
+      "checked": "2026-10-21",
+      "revision": "65eb23e40d0d61fab54c887f432718491cd51858"
+     }
+    },
+    {
+     "provider": "mlx",
+     "artifact": "mlx-community/Hy-Memory-30B-A3B-8bit",
+     "quant": "8bit",
+     "download_bytes": 31955742589,
+     "size_source": "catalog",
+     "verified": true,
+     "upstream": {
+      "method": "hf_api",
+      "checked": "2026-10-21",
+      "revision": "fe36f0ba3190ae05283429fe69e29b13eedec98f"
+     }
+    },
+    {
+     "provider": "huggingface",
+     "artifact": "tencent/Hy-Memory-30B-A3B-GGUF:Q4_K_M",
+     "quant": "q4_k_m",
+     "download_bytes": 18236702464,
+     "size_source": "catalog",
+     "verified": true,
+     "upstream": {
+      "method": "hf_api",
+      "checked": "2026-10-21",
+      "revision": "7778309f5109ec6e86820c687e638b9589b541a9"
+     }
+    }
+   ],
+   "notes": "Long-term memory writer (the write path of Mem0, MemOS and Hy-Memory). Thinking is off by default (reasoning_effort no_think). The MLX repos bundle the hy_v3 model class (config model_file) because mlx-lm ships none; AbstractCore uses its own copy."
+  },
   {
    "id": "nomic-embed-text-v1.5",
    "family": "nomic-embed",