<!-- agent-source-envelope -->
# Artifact: VibeThinker-3B MLX Conversion | PTLLM

> A public Apple-Silicon-friendly conversion of WeiboAI/VibeThinker-3B, preserved as a runtime artifact rather than claimed as a PostTrainLLM-trained model.

Canonical: https://posttrainllm.com/artifacts/vibethinker-3b-mlx
<!-- /agent-source-envelope -->

Conversion case study

Report artifact

2026-07-03

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

Evidence: observed

**Local GSM8K screen:** 40/40

small historical slice; reasoning sanity check

Evidence: historical

**Training delta:** None

format conversion, not post-training

Evidence: derived

**Native tool calling:** No

not a drop-in agent

Evidence: 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.
