Skip to content

fix: default to qwen-3-235b (free tier) instead of gated gpt-oss-120b - #5

Open
sugi-chan wants to merge 1 commit into
mainfrom
fix/default-model
Open

sugi-chan wants to merge 1 commit into
mainfrom
fix/default-model

Conversation

@sugi-chan

@sugi-chan sugi-chan commented Apr 20, 2026 •

Copy link
Copy Markdown
Contributor

Summary

Fixes #3. Changes the default CEREBRAS_MODEL from gpt-oss-120b (gated on most tiers) to qwen-3-235b-a22b-instruct-2507 (available on free tier).

The Bug

The previous default gpt-oss-120b returns 404 from /chat/completions on Cerebras free and lower paid tiers, even though it appears in /models. Users would install the plugin, see the MCP connect successfully, then have every generation call fail — a confusing failure mode.

Free tier model availability (tested 2026-04):

Model Free tier
qwen-3-235b-a22b-instruct-2507 ✅
llama3.1-8b ✅
gpt-oss-120b ❌ gated
zai-glm-4.7 ❌ gated

What Changed

  • mcp/src/cerebras-client.ts — default changed to qwen-3-235b-a22b-instruct-2507, with inline comment explaining tier gating
  • CLAUDE.md — config table updated with new default, added "Model selection" section documenting tier availability
  • CLAUDE.md — also fixed GENERATION_TIMEOUT doc entry (was 30000 misleadingly, code treats value as seconds not ms, so accurate default is 120)
  • .claude-plugin/plugin.json — bumped to 1.2.0 (minor version — default behavior change, not a bugfix)
  • mcp/dist/index.js — rebuilt bundle with new default

Test plan

  • Install plugin fresh on Cerebras free tier, verify mcp__speed-run__generate works without setting CEREBRAS_MODEL
  • Users with tier access to gpt-oss-120b can still set CEREBRAS_MODEL=gpt-oss-120b explicitly

Fixes #3

🤖 Generated with Claude Code

Summary by CodeRabbit

  • Chores

    • Updated plugin version to 1.2.0.
  • Documentation

    • Updated configuration defaults for Cerebras integration.
    • Added "Model selection" section documenting free-tier compatibility and tier-gated models, with a table of tested models.

The previous default `gpt-oss-120b` is tier-gated — /models lists it
but /chat/completions returns 404 on free/lower tiers, causing the
plugin to fail silently for most new users despite the MCP server
being connected.

Tested against Cerebras free tier (2026-04):
- qwen-3-235b-a22b-instruct-2507: works
- llama3.1-8b: works
- gpt-oss-120b: 404 gated
- zai-glm-4.7: 404 gated

qwen-3-235b is the strongest of the free-tier-available models for
codegen. Users on higher tiers can still override via CEREBRAS_MODEL.

Also fixes CLAUDE.md config table — GENERATION_TIMEOUT is parsed
as seconds (default 120) not milliseconds. Previous docs said "30000"
which would be 8+ hours if treated as seconds. Added a model tier
compatibility section.

Bumps plugin.json to 1.2.0 — this is a behavior change (new default
model) not a bugfix, so minor bump is appropriate.

Fixes #3

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
@coderabbitai

coderabbitai Bot commented Apr 20, 2026 •

Copy link
Copy Markdown

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info
⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: d09fc3b5-4ef8-4294-98d0-db5f50322b8d

📥 Commits

Reviewing files that changed from the base of the PR and between 3baa3d9 and 952ee4a.

⛔ Files ignored due to path filters (1)
  • mcp/dist/index.js is excluded by !**/dist/**
📒 Files selected for processing (3)
  • .claude-plugin/plugin.json
  • CLAUDE.md
  • mcp/src/cerebras-client.ts

Walkthrough

The plugin manifest version was incremented from 1.1.1 to 1.2.0. The default Cerebras model was changed from gpt-oss-120b to qwen-3-235b-a22b-instruct-2507 due to tier-gating restrictions, with corresponding documentation updates explaining model availability across tiers.

Changes

Cohort / File(s) Summary
Version and Configuration Updates
.claude-plugin/plugin.json, CLAUDE.md, mcp/src/cerebras-client.ts
Version bump to 1.2.0; default Cerebras model changed from gpt-oss-120b to qwen-3-235b-a22b-instruct-2507 to address tier-gating restrictions; documentation added for model selection and tier compatibility guidance.

Estimated code review effort

🎯 1 (Trivial) | ⏱️ ~3 minutes

Possibly related PRs

Poem

🐰 A model swap brings cheer so bright,
From gated doors to freedom's light,
Qwen now guides our quest with grace,
Free tiers smile across the space! ✨

🚥 Pre-merge checks | ✅ 5
✅ Passed checks (5 passed)
Check name Status Explanation
Description Check ✅ Passed Check skipped - CodeRabbit’s high-level summary is enabled.
Title check ✅ Passed The PR title clearly describes the main change: switching the default model from gpt-oss-120b (gated) to qwen-3-235b (free tier), which directly addresses the core issue.
Linked Issues check ✅ Passed All coding requirements from issue #3 are met: the default model in cerebras-client.ts changed to qwen-3-235b-a22b-instruct-2507, documentation updated with model availability notes, and version bumped.
Out of Scope Changes check ✅ Passed All changes are directly related to the linked issue #3 objective: updating default model, documentation, version bump, and rebuilt bundle. No unrelated modifications detected.
Docstring Coverage ✅ Passed No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

✏️ Tip: You can configure your own custom pre-merge checks in the settings.

✨ Finishing Touches
📝 Generate docstrings
  • Create stacked PR
  • Commit on current branch
🧪 Generate unit tests (beta)
  • Create PR with unit tests
  • Commit unit tests in branch fix/default-model

Comment @coderabbitai help to get the list of available commands and usage tips.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Change default CEREBRAS_MODEL — gpt-oss-120b is gated on most tiers

1 participant