[DIYA73]

What

Hard-coded Python renderer mirroring the Mistral-3 [INST]/[/INST] v3 chat template, following the same pattern as Llama3Renderer.

Covers

• mistralai/Mistral-Small-3.1-24B-Instruct-2503
• mistralai/Mistral-Small-3.2-24B-Instruct-2506
• mistralai/Mistral-Nemo-Instruct-2407

Changes

• renderers/mistral3.py — Mistral3Renderer with render, parse_response, bridge_to_next_turn
• renderers/parsing.py — parse_mistral3 (detects [TOOL_CALLS]-anchored tool-call arrays)
• renderers/configs.py — Mistral3RendererConfig
• renderers/base.py — MODEL_RENDERER_MAP entries for all three checkpoints
• renderers/__init__.py — lazy-load registration and __all__
• tests/parity.py — catalog entry for Mistral-Small-3.1

Validation

uv run pytest tests/ -x -q: 9588 passed, 146 skipped

Note

Add Mistral3Renderer for Mistral-3.x Instruct chat rendering

• Adds mistral3.py with Mistral3Renderer, which renders system, user, assistant, tool, and ipython messages into the Mistral-3 control-token format and emits assistant tool calls as a JSON array.
• Adds Mistral3RendererConfig to the RendererConfig union and registers mistral-3 in the renderer and config registries; the Mistral-Small 3.1 Instruct and Mistral-Nemo Instruct model IDs map to it automatically.
• Adds parse_mistral3 in parsing.py to parse completion text and tool-call JSON after the tool-call marker, with EOS as the only stop token.
• bridge_to_next_turn reuses a prior prompt through its EOS close and appends new user/tool turns; render accepts tools but does not emit tool definitions, and unsupported roles raise errors.
• Excludes mistral-3 from several reasoning-boundary test parameterizations in test_reasoning_boundaries.py and adds the Mistral-Small 3.1 checkpoint to the parity catalog with roundtrip and some tool-call suites disabled.
• Risk: mistral-3 constructs raise errors when required special tokens are missing from the tokenizer vocabulary; check _token_id in mistral3.py if tokenizer load changes.

^{Macroscope summarized ddecce7.}

[awesome-pro]

@DIYA73, tried to run the suite locally against your branch.

Two things actually:

1. The renderer is never registered, so nothing can construct it. _populate_registry() in renderers/base.py (~L1353) is missing both the from renderers.mistral3 import Mistral3Renderer import and the "mistral-3": Mistral3Renderer entry. Currently:

ValueError: Unknown renderer 'mistral-3'. Available: deepseek-r1, ..., qwen3.8

Adding those two lines moves test_bridge.py from an error to 23 passed.

2. After that fix the full suite is 61 failed / 11195 passed (main is 11116 passed / 126 skipped). The bulk - 31 - is test_reasoning_boundaries.py[...-mistral-3]: that file builds its case list from every renderer in MODEL_CATALOG and only excludes {"llama-3","default"} (L77 / L199). Since Mistral-3 has no reasoning channel (as your own config docstring says), it should be added to those exclusion sets; llama-3 is the precedent.

Two more things:

• mistralai/Mistral-Small-3.1-24B-Instruct-2503 has no chat_template on the Hub (tokenizer.chat_template is None), so the parity oracle can't produce a reference. the three test_parity.py failures are "Cannot use chat template functions because tokenizer.chat_template is not set". Either exclude these models from the parity suite with a note, or point the oracle at the Mistral reference.

• mistralai/Mistral-Small-3.2-24B-Instruct-2506's tokenizer currently fails to load at all (Couldn't instantiate the backend tokenizer, with and without fix_mistral_regex=True), so that mapping can't be verified in CI right now.

[DIYA73]

Thanks for the thorough review. Fixed in 30e004e:

Added Mistral3Renderer import + "mistral-3" entry to _populate_registry()
Added "mistral-3" to all no-reasoning exclusion sets in test_reasoning_boundaries.py (same pattern as llama-3)
Set roundtrip=False for Mistral-Small-3.1 in parity catalog — no chat_template on Hub
Dropped Mistral-Small-3.2 from MODEL_RENDERER_MAP — tokenizer fails to load

Full suite now: 989 passed in test_reasoning_boundaries.py, 0 failed.