feat(renderers): add Inkling renderer (text + image + audio)

README.md

@@ -40,7 +40,7 @@ next_prompt_ids = r.bridge_to_next_turn(
 )
 ```
 
-Hand-coded renderers ship for `qwen3`, `qwen3-vl`, `qwen3.5`, `qwen3.6`, `glm-5`, `glm-5.1`, `glm-4.5`, `minimax-m2`, `deepseek-v3`, `deepseek-r1`, `kimi-k2`, `kimi-k2.5` / `kimi-k2.6`, `nemotron-3`, `nemotron-3-ultra`, `llama-3`, `gpt-oss`, `hy3`, and `prime-qwen3`. Anything else falls back to `DefaultRenderer`, a generic `apply_chat_template` wrapper.
+Hand-coded renderers ship for `qwen3`, `qwen3-vl`, `qwen3.5`, `qwen3.6`, `glm-5`, `glm-5.1`, `glm-4.5`, `minimax-m2`, `deepseek-v3`, `deepseek-r1`, `kimi-k2`, `kimi-k2.5` / `kimi-k2.6`, `nemotron-3`, `nemotron-3-ultra`, `llama-3`, `gpt-oss`, `hy3`, `inkling`, and `prime-qwen3`. Anything else falls back to `DefaultRenderer`, a generic `apply_chat_template` wrapper. `qwen3-vl`, `qwen3.5`, `qwen3.6`, `kimi-k2.5` / `kimi-k2.6`, and `inkling` are multimodal (`inkling` handles both image **and** audio).
 
 ## API
 

docs/renderer-config.md

@@ -36,6 +36,7 @@ chat-template kwargs. Those fields are covered by parity tests against
 | Hy3 | `Hy3RendererConfig` | `reasoning_effort`, `preserved_thinking`, `is_training`, `raw_last_assistant`, `fallback_strategy` | - |
 | Kimi K2 | `KimiK2RendererConfig` | - | `enable_thinking` |
 | Kimi K2.5 / 2.6 | `KimiK25RendererConfig` | `thinking` | `image_cache_max` |
+| Inkling | `InklingRendererConfig` | `reasoning_effort` | `image_cache_max` |
 | Laguna XS.2 | `LagunaXS2RendererConfig` | `enable_thinking`, `render_assistant_messages_raw` | - |
 | Laguna XS-2.1 | `LagunaXS21RendererConfig` | `enable_thinking` | - |
 | Llama 3 | `Llama3RendererConfig` | `date_string`, `tools_in_user_message` | - |
@@ -136,7 +137,7 @@ the knobs its template actually exposes:
 | Nemotron-3 | `truncate_history_thinking=False -> all`; else `enable_thinking=False -> all`; else `tool_cycle` |
 | DeepSeek R1 | `template` |
 | MiniMax M2 | `tool_cycle` |
-| DeepSeek V3, Qwen3-VL, Kimi K2, Laguna XS.2 / XS-2.1, Llama 3 | `all` |
+| DeepSeek V3, Qwen3-VL, Kimi K2, Laguna XS.2 / XS-2.1, Llama 3, Inkling | `all` |
 | PrimeIntellect Qwen3 | `all` |
 
 Config construction raises when an explicit template knob directly contradicts

pyproject.toml

@@ -19,7 +19,8 @@ dependencies = [
     "openai>=1.108.1",
     "tiktoken",
     "jinja2",
-    "transformers>=4.50.0",
+    # >=5.14.0: first release with native Inkling processor support.
+    "transformers>=5.14.0",
     # Used by GptOssRenderer to render and parse harmony tokens. Vendoring
     # OpenAI's reference implementation keeps us byte-identical with vLLM
     # (which also uses it) and saves us mirroring a 330-line Jinja template.
@@ -101,6 +102,9 @@ exclude-newer = "7 days"
 # fast-track so first-party releases can land same-day. Only packages that
 # appear in `uv tree` are listed.
 prime-pydantic-config = false
+# transformers >=5.14 (native Inkling processor) lands inside the cooldown;
+# fast-track it, and drop this once 5.14.x clears the 7-day window.
+transformers = false
 
 [tool.ty.environment]
 python-version = "3.13"

renderers/__init__.py

@@ -52,6 +52,7 @@
     GLM5RendererConfig,
     GptOssRendererConfig,
     Hy3RendererConfig,
+    InklingRendererConfig,
     KimiK25RendererConfig,
     KimiK2RendererConfig,
     LagunaXS2RendererConfig,
@@ -88,6 +89,7 @@
     "GLM5Renderer": "renderers.glm5",
     "GptOssRenderer": "renderers.gpt_oss",
     "Hy3Renderer": "renderers.hy3",
+    "InklingRenderer": "renderers.inkling",
     "KimiK25Renderer": "renderers.kimi_k25",
     "KimiK2Renderer": "renderers.kimi_k2",
     "LagunaXS21Renderer": "renderers.laguna_xs2",
@@ -141,6 +143,8 @@ def __dir__() -> list[str]:
     "Hy3Renderer",
     "Hy3RendererConfig",
     "ImagePart",
+    "InklingRenderer",
+    "InklingRendererConfig",
     "KimiK25Renderer",
     "KimiK25RendererConfig",
     "KimiK2Renderer",

renderers/base.py

@@ -1071,6 +1071,8 @@ def bridge_to_next_turn(self, *args: Any, **kwargs: Any) -> "RenderedTokens | No
     # GPT-OSS.
     "openai/gpt-oss-20b": "gpt-oss",
     "openai/gpt-oss-120b": "gpt-oss",
+    # Thinking Machines Inkling (vision + audio; native processor in transformers >= 5.14).
+    "thinkingmachines/Inkling": "inkling",
     # Tencent Hunyuan Hy3 (295B-A21B MoE). The FP8 checkpoint shares the same
     # tokenizer and chat template. Hy3-preview is deliberately unmapped: it
     # ships an older, incompatible template (un-suffixed special tokens,
@@ -1115,6 +1117,7 @@ def bridge_to_next_turn(self, *args: Any, **kwargs: Any) -> "RenderedTokens | No
     # ``grid_thws``.
     "moonshotai/Kimi-K2.5": {"image"},
     "moonshotai/Kimi-K2.6": {"image"},
+    "thinkingmachines/Inkling": {"image", "audio"},
 }
 
 
@@ -1313,6 +1316,7 @@ def _populate_registry():
     from renderers.glm45 import GLM45Renderer
     from renderers.gpt_oss import GptOssRenderer
     from renderers.hy3 import Hy3Renderer
+    from renderers.inkling import InklingRenderer
     from renderers.kimi_k2 import KimiK2Renderer
     from renderers.kimi_k25 import KimiK25Renderer
     from renderers.laguna_xs2 import LagunaXS2Renderer, LagunaXS21Renderer
@@ -1340,6 +1344,7 @@ def _populate_registry():
             "deepseek-v3": DeepSeekV3Renderer,
             "deepseek-r1": DeepSeekR1Renderer,
             "hy3": Hy3Renderer,
+            "inkling": InklingRenderer,
             "kimi-k2": KimiK2Renderer,
             "kimi-k2.5": KimiK25Renderer,
             "laguna-xs.2": LagunaXS2Renderer,

renderers/configs.py

@@ -380,6 +380,64 @@ def _check_thinking_retention(self):
         return self
 
 
+# Inkling ``reasoning_effort`` label → float (the template's own effort_map).
+INKLING_EFFORT_MAP: dict[str, float] = {
+    "none": 0.0,
+    "minimal": 0.1,
+    "low": 0.2,
+    "medium": 0.7,
+    "high": 0.9,
+    "max": 0.99,
+}
+
+
+class InklingRendererConfig(BaseRendererConfig):
+    """Inkling renderer config (``thinkingmachines/Inkling``).
+
+    Inkling gates reasoning depth via a ``reasoning_effort`` knob rather
+    than a boolean ``enable_thinking``. It accepts either a string label
+    from :data:`INKLING_EFFORT_MAP` (``none`` … ``max``) or a raw float in
+    ``[0.0, 0.99]``; the renderer emits ``Thinking effort level: {N}`` in a
+    leading system message exactly as the chat template does (label mapped
+    to its float, ``0.0`` printed as ``"0"``). The template default —
+    applied when no ``reasoning_effort`` is passed — is ``0.9``, which this
+    config mirrors.
+
+    Reasoning is preserved on every historical assistant turn (the template
+    has no history-dropping knob), so the effective bridge policy is
+    ``"all"``.
+    """
+
+    name: Literal["inkling"] = "inkling"
+
+    reasoning_effort: str | float = 0.9
+    """Reasoning-effort gate. Mirrors the chat template's ``reasoning_effort``
+    kwarg: a label in :data:`INKLING_EFFORT_MAP` or a float in ``[0.0, 0.99]``.
+    Default ``0.9`` matches the template's own default (equivalent to
+    ``"high"``)."""
+
+    image_cache_max: int = 256
+    """FIFO bound on the per-renderer image-processor cache. Renderer-
+    internal — not a Jinja chat-template kwarg."""
+
+    _internal_fields = frozenset({"image_cache_max"})
+
+    @model_validator(mode="after")
+    def _check_reasoning_effort(self):
+        eff = self.reasoning_effort
+        if isinstance(eff, str):
+            if eff.strip() not in INKLING_EFFORT_MAP:
+                raise ValueError(
+                    f"reasoning_effort={eff!r} is not a known label. "
+                    f"Use one of {sorted(INKLING_EFFORT_MAP)} or a float in [0.0, 0.99]."
+                )
+        else:
+            num = float(eff)
+            if num < 0.0 or num > 0.99:
+                raise ValueError(f"reasoning_effort={eff!r} must be in [0.0, 0.99].")
+        return self
+
+
 class GptOssRendererConfig(BaseRendererConfig):
     """OpenAI gpt-oss (harmony) renderer config.
 
@@ -664,6 +722,7 @@ class DeepSeekR1RendererConfig(BaseRendererConfig):
         GLM45RendererConfig,
         GptOssRendererConfig,
         Hy3RendererConfig,
+        InklingRendererConfig,
         KimiK2RendererConfig,
         KimiK25RendererConfig,
         LagunaXS2RendererConfig,
@@ -704,6 +763,7 @@ class DeepSeekR1RendererConfig(BaseRendererConfig):
     "glm-4.5": GLM45RendererConfig,
     "gpt-oss": GptOssRendererConfig,
     "hy3": Hy3RendererConfig,
+    "inkling": InklingRendererConfig,
     "kimi-k2": KimiK2RendererConfig,
     "kimi-k2.5": KimiK25RendererConfig,
     "laguna-xs.2": LagunaXS2RendererConfig,
@@ -752,6 +812,8 @@ def config_from_name(name: str) -> BaseRendererConfig | None:
     "GLM5RendererConfig",
     "GptOssRendererConfig",
     "Hy3RendererConfig",
+    "INKLING_EFFORT_MAP",
+    "InklingRendererConfig",
     "KimiK25RendererConfig",
     "KimiK2RendererConfig",
     "LagunaXS2RendererConfig",