Skip to content

v0.3.2: new model support, reasoning fixes, fresh pricing

Latest
Latest

Choose a tag to compare

View all tags
@hashangit hashangit released this 12 Sep 20:25
· 4 commits to main since this release
v0.3.2
07bda33

OpenFusion 0.3.2

New models, fewer reasoning glitches, current pricing. This is a dependency refresh, so there is nothing to reconfigure: your config, keys, and activity history carry over untouched.

Released September 13, 2026.

New model support

OpenFusion routes every candidate and judge call through its inference library, and this release moves that library forward three months (pi-ai 0.79.4 to 0.85.1). If you pick models in Settings, here is what you gain:

  • GPT-6 Astra now resolves for OpenAI API keys and Codex subscriptions.
  • Grok 4.6 is xAI's new default model, and xAI calls use the Responses API with encrypted reasoning replay. Note that xAI retired Grok 3, Grok 3 Fast, Grok 4.20, and Grok Code Fast 1 in this window, so switch any candidate pinned to those.
  • Qwen Token Plan Individual is a new built-in provider. Provide your QWEN_TOKEN_PLAN_API_KEY and its subscription catalog, including Qwen3.8 Flash, appears in the picker.
  • ZAI Coding Plan models for China, including GLM-4.6V with vision, plus usage cost estimates for models with published pay-as-you-go prices.
  • An experimental, vision-capable DeepSeek V4 Flash model.

Fixes you will notice in fusion runs

  • Reasoning no longer breaks mid-run. OpenRouter reasoning effort now comes from each model's own metadata, so models that require reasoning never receive effort: none. Anthropic thinking blocks that have empty text but a valid signature replay correctly on newer Claude models, including through OpenRouter. Gemini 3 tool-call ids survive history conversion, which repairs signed multi-turn replay.
  • Cost numbers are more honest. Kimi's cached tokens now count as cache reads instead of full-price input, and DeepSeek pricing data is current. Fusions using either provider report closer-to-invoice costs.
  • OpenAI-compatible endpoints replay assistant reasoning details verbatim, and long prompt-cache requests use the correct cache TTL.

Under the hood

  • MCP SDK 1.30.0, better-sqlite3 13, and zod 4.6, all re-verified against the full suite: 217 tests covering the MCP tool, the Tasks contract, _resume_from retrieval, and the Playground route, plus a clean production build.
  • Source builds now run on TypeScript 7, vitest 5, and tsx 4.23.
  • One library detail worth knowing: the inference library moved its global API to a compat entrypoint in this range, and OpenFusion now imports from it with unchanged behavior. Migrating to the new Models runtime is planned for when upstream retires compat.

Upgrade

npm install -g openfusion-mcp

Or pull and pnpm install if you run from source. No config migration, no database changes, no new environment variables.

Assets 2
Loading