feat(models): upgrade MiniMax default model to M3 - #6473
Conversation
Add MiniMax-M2.7-highspeed to the minimax and minimax-cn provider model lists. MiniMax-M2.7-highspeed delivers the same performance as M2.7 but with faster inference and lower latency. API reference: https://platform.minimax.io/docs/api-reference/text-openai-api
|
Addresses #11367 (request to add MiniMax highspeed model variants). |
- Add MiniMax-M3 to model list and set as default for minimax and minimax-cn - Keep MiniMax-M2.7 and MiniMax-M2.7-highspeed as alternatives - Remove older models (M1 variants, M2.5) from catalogs and metadata - Update auxiliary model defaults to MiniMax-M3 - Update hindsight default model and docs examples - Update related unit tests
|
Thanks for the MiniMax catalog update. Main has since landed most of the M3 runtime work: Problems
Suggested changes
Automated hermes-sweeper review. |
GottZ
left a comment
There was a problem hiding this comment.
This was generated by AI during triage.
Summary
Six PRs address successive MiniMax catalog upgrades: #1882, #1900, #1932, and #4197 moved native and OpenRouter references from M2.5 to M2.7, while #6473 and #36191 move remaining defaults and catalogs to M3. The M2.7 work is already represented on main through #1900 and the #4197 salvage in #4208; for M3, #36191 covers OpenRouter/Nous lists, while #6473 contains several remaining updates but also conflicts with current runtime architecture and metadata.
Related pull requests
- #1882 [closed]
related— (+21/-13) — superseded by #1900: upgraded MiniMax defaults, provider catalogs, metadata, fallback tests, RL examples, and documentation from M2.5 to M2.7; the contributor confirmed these changes were manually carried into merged PR #1900 after refactor conflicts. - #1900 [merged]
related— (+25/-14) — merged reference implementation: landed the substantive #1882 M2.7 upgrade across native defaults, setup, metadata, tests, RL tooling, and docs, with additional OpenRouter catalog entries. - #1932 [closed]
related— (+9/-3) — already satisfied on main: repeated the M2.7 catalog, metadata, and setup-selection changes after #1900 had landed; the contributor confirmed that M2.7 and its highspeed variant were already present and ordered first. - #4197 [closed]
related— (+195/-4) — partially salvaged via #4208: overlapped #1932 and existing main for most M2.7 registrations, but identified remaining gaps in the CLI provider flow and opencode-go catalog that the contributor reports were cherry-picked into #4208. - #6473
related— (+48/-56) — keep open for a narrowed revision, not merge as-is: it adds M3/M2.7-highspeed to setup and updates the Hindsight default and direct-provider examples, but its 524,288 context value would regress main's 1M-class M3 handling, its auxiliary edit targets a removed legacy mapping, and its mutable catalog assertions conflict with repository testing policy. This follows the visible keep_open review on #6473: retain current profile routing and metadata, then salvage only the still-stale setup, Hindsight, and documentation changes. - #36191 [merged]
related— (+2/-2) — merged complementary implementation: updated the OpenRouter and Nous curated entries from minimax-m2.7 to minimax-m3; it does not cover #6473's remaining native setup fallback, Hindsight default, or direct-provider documentation.
Duplicates
#1882 and #1900 implement substantially the same M2.7 upgrade, with #1900 being the merged superseding version. #1932 and most of #4197 duplicate M2.7 work already on main, while #4197's remaining gaps were salvaged through #4208. #36191 is complementary to, not a duplicate of, the narrow salvageable portion of #6473.
Suggested consolidation
Revise and then merge #6473 as the consolidation target for the remaining M3 work: keep only the current-main-compatible setup fallback, Hindsight default, and direct-provider documentation updates, while dropping the 524,288 metadata regression, legacy auxiliary mapping edit, already-landed catalogs, and mutable catalog-list tests. Keep #1882, #1932, and #4197 closed as superseded/already-salvaged duplicates; #1900 and #36191 remain the merged reference implementations for the M2.7 baseline and OpenRouter/Nous M3 catalog update, respectively.
Cross-PR triage: Reviewed 6 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 45 kB of PR diffs, 7 kB of issue/PR text, 3 kB of discussion (6 comments), 0 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.
Summary
Upgrade the MiniMax model catalogs to use MiniMax-M3 as the default while keeping MiniMax-M2.7 and MiniMax-M2.7-highspeed as alternatives. Older models (M1 family and M2.5) are removed from the catalogs and metadata.
Changes
MiniMax-M3to the catalog for bothminimaxandminimax-cnproviders and place it first (default position)MiniMax-M2.7andMiniMax-M2.7-highspeedas alternativessetup.pydefaults:MiniMax-M1,MiniMax-M1-40k,MiniMax-M1-80k,MiniMax-M1-128k,MiniMax-M1-256k,MiniMax-M2.5minimaxandminimax-cntoMiniMax-M3inagent/auxiliary_client.pyagent/model_metadata.pyDEFAULT_CONTEXT_LENGTHS: addminimax-m3(524288 / 512K), drop M1 and M2.5 entries, set genericminimaxprefix fallback to 524288MiniMax-M2.7toMiniMax-M3website/docs/integrations/providers.mdto useMiniMax-M3tests/agent/test_minimax_provider.pyandtests/hermes_cli/test_setup_model_selection.pyWhy
MiniMax-M3is the latest model, with a 512K context window, up to 128K output, and image input support (OpenAI- and Anthropic-compatible). M2.7 and M2.7-highspeed remain available for callers that still prefer them.Testing
pytest tests/agent/test_minimax_provider.py— 12 passedpytest tests/hermes_cli/test_setup_model_selection.py— 9 passedpytest tests/hermes_cli/test_setup_model_provider.py tests/hermes_cli/test_model_normalize.py tests/hermes_cli/test_api_key_providers.py tests/agent/test_auxiliary_client.py— 256 passedpytest tests/run_agent/test_fallback_model.py tests/agent/test_minimax_auxiliary_url.py— 36 passed