feat: add MiniMax as a supported LLM provider with M2.7 models#5666
feat: add MiniMax as a supported LLM provider with M2.7 models#5666octo-patch wants to merge 2 commits intocomet-ml:mainfrom
Conversation
Add MiniMax integration documentation, temperature constraint handling, and tests to support MiniMax models (MiniMax-M2.5, MiniMax-M2.5-highspeed) as a first-class provider in Opik. Changes: - Add MiniMax integration documentation page (minimax.mdx) with usage examples for OpenAI SDK, LiteLLM, and evaluation metrics - Add MiniMax to the Model Providers section in overview and README - Add MiniMax navigation entry and redirect in docs.yml - Add MiniMax-specific temperature filter in util.py (MiniMax requires temperature > 0, so zero values are clamped to 0.01) - Add unit tests for MiniMax temperature constraint handling
github-actions
bot
added
documentation
![baz-reviewer[bot]](https://avatars.githubusercontent.com/in/933528?s=60&v=4){kind=small}
| ## Important Notes | ||
|
|
||
| - **Temperature**: MiniMax models require temperature to be strictly greater than 0. If you set `temperature=0`, it will be automatically adjusted when using LiteLLM with Opik. | ||
| - **API Compatibility**: MiniMax's API is fully compatible with the OpenAI SDK, so any OpenAI-compatible tool or framework will work with MiniMax. | ||
|
|
There was a problem hiding this comment.
Docs claim MiniMax temperature=0 is auto-adjusted when using LiteLLM with Opik, but the clamp is only applied in LiteLLMChatModel (sdks/python/src/opik/evaluation/models/litellm/util.py) and the Opik monitoring path (sdks/python/src/opik/evaluation/models/litellm/opik_monitor.py lines 20–94) never calls _apply_minimax_filters. Calling litellm.completion with temperature=0 and only Opik callbacks still sends temperature=0 to MiniMax and fails; can we either narrow the doc note to the evaluation-model use case or add the clamp to the LiteLLM monitoring integration?
Finding type: Logical Bugs | Severity: 🔴 High
Want Baz to fix this for you? Activate Fixer
Other fix methods
Prompt for AI Agents:
In apps/opik-documentation/documentation/fern/docs/tracing/integrations/minimax.mdx
around lines 208-212, the note incorrectly states that MiniMax temperature=0 is
automatically adjusted when using LiteLLM with Opik. Change the wording to accurately
reflect the current behavior: state that the automatic clamp only occurs when using the
LiteLLMChatModel evaluation integration, not when using the generic OpikLogger
callbacks; alternatively, implement the clamp in the LiteLLM monitoring integration
(sdks/python/src/opik/evaluation/models/litellm/opik_monitor.py) so OpikLogger also
applies _apply_minimax_filters before sending requests. Make the doc update concise and
unambiguous, or if you implement the code change, add a brief comment in opik_monitor.py
indicating why the clamp is needed for MiniMax compatibility.
| def _apply_minimax_filters( | ||
| params: Dict[str, Any], | ||
| already_warned: Set[str], | ||
| warn: Callable[[str, Any], None], | ||
| ) -> None: | ||
| """Apply MiniMax specific parameter filters. | ||
|
|
||
| MiniMax requires temperature to be strictly greater than 0. | ||
| A temperature of 0 is rejected by the API, so we clamp it to a small | ||
| positive value to avoid errors. | ||
| """ | ||
|
|
There was a problem hiding this comment.
params['temperature'] is treated as optional, but _apply_minimax_filters only clamps values that parse to float and are <= 0.0; float(value) raises for None/non-numeric inputs, so params['temperature'] remains None. This lets MiniMax requests send a nullish temperature and the API rejects them. Can we normalize invalid/null temperatures before returning (e.g., set to 0.01 or drop/log them) so callers can rely on a positive number?
Finding type: Validate nullable inputs | Severity: 🔴 High
Want Baz to fix this for you? Activate Fixer
Other fix methods
Prompt for AI Agents:
In sdks/python/src/opik/evaluation/models/litellm/util.py around lines 109-128, the
_apply_minimax_filters function currently only clamps numeric temperatures <= 0 but
leaves None or non-numeric values unchanged. Change the logic so that if params contains
"temperature" but parsing to float fails or the parsed value is None, set
params["temperature"] = 0.01 (the same clamp used for non-positive numbers) and call
warn(...) to indicate an invalid/null temperature was replaced; also keep the existing
behavior of clamping numeric <= 0 to 0.01. This ensures downstream callers never receive
a null/invalid temperature.
- Add MiniMax-M2.7 and MiniMax-M2.7-highspeed to model lists (frontend, backend, OpenRouter enum, pricing JSON) - Update default model references from M2.5 to M2.7 in docs and tests - Add M2.7 pricing data (cache_read updated to $0.06/M tokens) - Keep all previous models as available alternatives
github-actions
bot
added
java
octo-patch
changed the title
| from opik.evaluation.models import LiteLLMChatModel | ||
|
|
||
| # Create a MiniMax model for evaluation | ||
| minimax_model = LiteLLMChatModel(model_name="minimax/MiniMax-M2.7") |
There was a problem hiding this comment.
Should we replace the hand‑pasted minimax_model example with the generated example link or expand it to include the checklist-required intent line, a minimal runnable context with an inline behavior comment, and a maintenance note linking to the canonical/generated example and owner?
Finding type: Keep docs accurate | Severity: 🟢 Low
Want Baz to fix this for you? Activate Fixer
Other fix methods
Prompt for AI Agents:
In apps/opik-documentation/documentation/fern/docs/tracing/integrations/minimax.mdx
around line 190, the single-line example minimax_model =
LiteLLMChatModel(model_name="minimax/MiniMax-M2.7") is hand-pasted and missing required
metadata. First search the autogenerated SDK/docs and examples for a canonical/generated
Minimax example and, if found, replace this line with a link to that canonical example
(include the file path/URL and a brief note). If no canonical example exists, expand
this snippet by adding: (1) a one-line intent/trigger sentence immediately above it
describing when to use the snippet, (2) a minimal runnable context (import and variable
assignment) with an inline comment on the same line explaining the observable behavior,
and (3) a maintenance note below pointing to the canonical/generated artifact (or state
“none”), include the owning team, and an update cadence. Ensure the maintenance note
references a repo path or URL rather than vague text.
1 participant
Summary
Add MiniMax as a supported LLM provider for Opik, with the latest M2.7 flagship model as default.
Changes
Why
MiniMax-M2.7 is the latest flagship model with enhanced reasoning and coding capabilities, offering a 1M token context window.
Testing