MN Uncensored v0.3.0

Released 2026-07-16.

Added

• mn/god: Huihui Qwen3.6 35B A3B Abliterated on one H200.
• mn/code: YuYu Ornith 35B Abliterated on one H200.
• mn/fast: Huihui Qwythos 9B Abliterated on one L40S.
• Independent lifecycle records and Modal applications for every route.
• Model selection in the CLI, Hermes, Pi, OpenCode, and endpoint tests.
• Strict public model routing and legacy alias migration.
• 131,072-token context and 16,384-token output ceiling.
• Persistent Hugging Face, vLLM, and FlashInfer caches.
• Exact model revision, attribution, compatibility, and license record.
• Apache-2.0 repository license.

Changed

• Replaced the impractical single 397B prototype direction with an affordable
  35B/35B/9B evaluation catalog.
• Disabled Qwen thinking by default so standard clients receive normal
  content; direct clients can opt in.
• Aligned the vLLM tool parser with each pinned chat template.
• Persisted the selected catalog profile inside each Modal image.
• Added one retry for transient Modal recreate-rollover timeouts.

Verified

• /v1/models returns exactly the three public MN IDs.
• Every model passed an ordinary streaming completion.
• Every model passed a forced function/tool call.
• Hard-stop isolation was tested per model.