> Canonical page: https://posttrainllm.com/artifacts/vibethinker-3b-mlx

Conversion case study Report artifact 2026-07-03

# VibeThinker-3B MLX Conversion

This release answers a packaging question, not a training question: can the reasoning model be preserved in a local MLX-compatible form? A small local GSM8K screen scored 40/40, but no conversion-parity suite was recorded and the base has no native tool-calling behavior. The case study therefore reports a useful conversion without inventing a model-quality delta.

- MLX
- conversion
- reasoning
- archive

## Headline Numbers

Evidence class applies to each value: measured = retained evaluation or runtime result; historical = recorded result not freshly reproduced; derived = calculated or decided from records; observed = current artifact state; not-measured = an explicit evidence gap.

Model class

 3B converted WeiboAI/VibeThinker-3B observed

Local GSM8K screen

 40/40 small historical slice; reasoning sanity check historical

Training delta

 None format conversion, not post-training derived

Native tool calling

 No not a drop-in agent historical

## Competitive Context

**System**

**Metric**

**Score**

**Size / Class**

**Comparable?**

**Readout**

 PostTrainLLM VibeThinker-3B MLX local GSM8K sanity slice 40/40 3B MLX conversion Direct Historical local verification of reasoning behavior on a small slice; not a broad benchmark claim.

 [WeiboAI/VibeThinker-3B](https://huggingface.co/WeiboAI/VibeThinker-3B) upstream reasoning model source weights 3B Not comparable The conversion derives from this public model, but no controlled pre/post conversion parity table was preserved.

 Agentic distilled descendant current tool-calling eval not recorded 3B Not comparable Public descendant weights exist, but they have no current eval promotion.

Direct rows share this artifact's eval setup. Directional rows are useful market context but should not be read as leaderboard claims.

## What this release proves

**Question**

**Evidence**

**Decision**

 Are unique converted weights public? Yes Preserve on Hugging Face

 Did a local reasoning sanity screen run? GSM8K 40/40 Useful, small historical slice

 Was conversion parity measured? Not recorded No parity claim

 Was the model post-trained by PostTrainLLM? No Conversion-only artifact

 Does it natively call tools? No Not an agentic specialist

## Release Blockers

### No controlled conversion-parity report

The local sanity result does not prove numerical or benchmark parity with the upstream runtime.

 Unblock: Run paired upstream-vs-MLX checks only if a consumer needs this conversion as an active dependency.

### No native agentic behavior

The reasoning base does not provide a validated tool-calling interface.

 Unblock: Treat it as a reasoning/runtime artifact; evaluate a separately adapted candidate before agent use.

## Evidence

- [Hugging Face model](https://huggingface.co/posttrainllm/vibethinker-3b-mlx)
- [Upstream VibeThinker-3B](https://huggingface.co/WeiboAI/VibeThinker-3B)
- [Local reasoning-method review](https://posttrainllm.com/docs/learn/diversity-driven-small-model-reasoning/)
- [Historical model inventory](https://posttrainllm.com/docs/sessions/06-17-stepback-inventory-roi/)
- [GSM8K evaluator](https://github.com/PostTrainLLM/posttrainllm/blob/main/scripts/gsm8k_eval.py)

## Next Release Action

Keep this as a conversion and runtime artifact. Add a pinned loader plus parity receipt only when a real Mac-local consumer justifies maintaining it.