fix(minimax): correct context lengths, model catalog, thinking guard, aux model, and config base_url by kshitijk4poor · Pull Request #6046 · NousResearch/hermes-agent

kshitijk4poor · 2026-04-08T05:40:16Z

What does this PR do?

Hardens the MiniMax provider implementation by fixing 5 bugs identified through comparison with OpenCode's provider patterns and MiniMax's official API documentation.

Changes

1. Context lengths corrected (critical)
MiniMax models were mapped to 204,800 tokens. Official API docs:

M1 family: 1,000,000 context window
M2.5/M2.7: 1,048,576 context window

This caused premature context compression, wasting user tokens. Added per-model context entries and a get_minimax_max_output() helper for per-model output token limits.

2. Model catalog updated

Added M1 family: MiniMax-M1, M1-40k, M1-80k, M1-128k, M1-256k
Removed deprecated MiniMax-M2.1 (not in official docs)
Removed undocumented highspeed variants (same model at 2x price)
Updated _PROVIDER_MODELS, _DEFAULT_PROVIDER_MODELS, and test fixtures

3. Thinking parameter guard (fixes crash)
build_anthropic_kwargs() was sending thinking to MiniMax's Anthropic-compatible endpoint. MiniMax doesn't support it, causing TypeError: Messages.create() got an unexpected keyword argument 'thinking'. Added 'minimax' not in model.lower() guard alongside the existing Haiku guard.

4. Auxiliary model pricing
Default aux model was MiniMax-M2.7-highspeed (2x price, same quality). Changed to standard MiniMax-M2.7.

5. Config base_url ignored for API-key providers
model.base_url from config.yaml was ignored for API-key providers (MiniMax, ZAI, Kimi, etc.). The code always used the hardcoded inference_base_url from the provider registry. China MiniMax users setting model.base_url: https://api.minimaxi.com/anthropic still got routed to api.minimax.io, causing 401 errors.

Fixed by checking model.base_url from config when the configured provider matches, in both the credential pool path and the direct resolution path. Env var MINIMAX_BASE_URL takes priority over config.

Research methodology

Compared against OpenCode's provider system (per-model context windows, per-model max output tokens, per-provider error handling) and verified all model specs against MiniMax's official documentation at minimaxi.com (extracted from JS bundles since it's a SPA).

Related Issues

Fixes #5777 (thinking parameter crash)
Fixes #4082 (highspeed aux model pricing)
Fixes #6039 (config base_url ignored)
Closes #3895 (model catalog update — superset of that PR)
Related: #5781 (aux URL routing — fixed in #5983)

Type of Change

🐛 Bug fix (non-breaking change that fixes an issue)

How to Test

python -m pytest tests/agent/test_minimax_provider.py -v
python -m pytest tests/hermes_cli/test_setup_model_selection.py -k minimax -v
python -m pytest tests/hermes_cli/test_runtime_provider_resolution.py -k minimax -v
python -m pytest tests/agent/test_anthropic_adapter.py -k thinking -v

24 new test assertions covering context lengths, max output lookup, thinking guard, aux model, catalog correctness, and config base_url override.

Checklist

I've read the Contributing Guide
Commit messages follow Conventional Commits
Searched for existing PRs — this supersedes feat(models): update minimax/minimax-cn model menus #3895 and fix(agent): use standard MiniMax-M2.7 for auxiliary model instead of highspeed #4115
Only changes related to this fix
pytest tests/hermes_cli/ tests/agent/ tests/run_agent/ — 414 targeted tests pass
Added tests: tests/agent/test_minimax_provider.py, runtime resolution tests
Tested on macOS (Apple Silicon)

… aux model, and config base_url Context lengths were 204K — actual MiniMax API docs show 1M (M1) and 1,048,576 (M2.5/M2.7). Premature context compression was wasting tokens. Model catalog listed deprecated M2.1 and undocumented highspeed variants. Updated to match official MiniMax API: M1 family (M1, M1-40k, M1-80k, M1-128k, M1-256k), M2.5, M2.7. build_anthropic_kwargs() was sending the 'thinking' parameter to MiniMax's Anthropic-compatible endpoint — MiniMax doesn't support it, causing TypeError crashes (Messages.create() got an unexpected keyword argument). Default auxiliary model was M2.7-highspeed (2x price, same quality). Changed to standard M2.7. API-key providers (MiniMax, ZAI, Kimi, etc.) ignored model.base_url from config.yaml, always using the hardcoded inference_base_url. China MiniMax users setting model.base_url to api.minimaxi.com/anthropic still got routed to api.minimax.io, causing 401. Now honours config.yaml base_url when the configured provider matches, with env var taking priority. Added get_minimax_max_output() for per-model output token limits and comprehensive test coverage (24 new tests). Fixes NousResearch#5777, fixes NousResearch#4082, fixes NousResearch#6039. Closes NousResearch#3895.

… aux model, and config base_url Cherry-picked from PR #6046 by kshitijk4poor with dead code stripped. - Context lengths: 204800 → 1M (M1) / 1048576 (M2.5/M2.7) per official docs - Model catalog: add M1 family, remove deprecated M2.1 and highspeed variants - Thinking guard: skip extended thinking for MiniMax (Anthropic-compat endpoint) - Aux model: MiniMax-M2.7-highspeed → MiniMax-M2.7 (same model, half price) - Config base_url: honour model.base_url for API-key providers (fixes China users) - Stripped unused get_minimax_max_output() / _MINIMAX_MAX_OUTPUT (no consumer) Fixes #5777, #4082, #6039. Closes #3895.

teknium1 · 2026-04-08T09:20:55Z

Merged via PR #6082. Your commit was cherry-picked onto current main with authorship preserved in git log. Stripped the unused get_minimax_max_output() helper (no consumer yet), but kept all 5 substantive fixes. Thanks for the thorough research and clean implementation!

… aux model, and config base_url Cherry-picked from PR NousResearch#6046 by kshitijk4poor with dead code stripped. - Context lengths: 204800 → 1M (M1) / 1048576 (M2.5/M2.7) per official docs - Model catalog: add M1 family, remove deprecated M2.1 and highspeed variants - Thinking guard: skip extended thinking for MiniMax (Anthropic-compat endpoint) - Aux model: MiniMax-M2.7-highspeed → MiniMax-M2.7 (same model, half price) - Config base_url: honour model.base_url for API-key providers (fixes China users) - Stripped unused get_minimax_max_output() / _MINIMAX_MAX_OUTPUT (no consumer) Fixes NousResearch#5777, NousResearch#4082, NousResearch#6039. Closes NousResearch#3895.

lghi mentioned this pull request Apr 8, 2026

minimax provider ignores custom base_url, hardcodes api.minimax.io causing 401 #6039

Closed

kshitijk4poor force-pushed the fix/minimax-provider-hardening branch from 9f9a197 to b2dcf8e Compare April 8, 2026 06:15

kshitijk4poor changed the title ~~fix(minimax): correct context lengths, model catalog, thinking guard, and aux model~~ fix(minimax): correct context lengths, model catalog, thinking guard, aux model, and config base_url Apr 8, 2026

kshitijk4poor force-pushed the fix/minimax-provider-hardening branch from b2dcf8e to 87dfaff Compare April 8, 2026 06:51

teknium1 mentioned this pull request Apr 8, 2026

fix(minimax): correct context lengths, model catalog, thinking guard, aux model, and config base_url #6082

Merged

teknium1 closed this Apr 8, 2026

teknium1 mentioned this pull request May 3, 2026

[Bug]: Minimax Default Auxiallry Model MiniMax-M2.7-highspeed is twice the price at API rates #4082

Closed

1 task

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix(minimax): correct context lengths, model catalog, thinking guard, aux model, and config base_url#6046

fix(minimax): correct context lengths, model catalog, thinking guard, aux model, and config base_url#6046
kshitijk4poor wants to merge 1 commit into
NousResearch:mainfrom
kshitijk4poor:fix/minimax-provider-hardening

kshitijk4poor commented Apr 8, 2026 •

edited

Loading

Uh oh!

teknium1 commented Apr 8, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

kshitijk4poor commented Apr 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What does this PR do?

Changes

Research methodology

Related Issues

Type of Change

How to Test

Checklist

Uh oh!

teknium1 commented Apr 8, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

kshitijk4poor commented Apr 8, 2026 •

edited

Loading