posttrainllm HF repos
6all have first-class case studiesobservedHugging Face Specialist Model Archive v1
This archive makes artifact storage explicit: unique posttrainllm model outputs live on Hugging Face, while plain upstream base-model caches are removed locally instead of being mirrored under posttrainllm.
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.
Local model cache
clearedafter upload/remote-size verificationobservedStorage policy
HF firstR2 remains optional private cache or legacy mirrorderivedCompetitive Context
| System | Metric | Score | Size / Class | Comparable? | Readout |
|---|---|---|---|---|---|
| Hugging Face artifact storage | public model distribution | active | model repos | Direct | Current target for public weights, adapters, and large specialist artifacts. |
| Local Mac cache | durable artifact storage | rejected | ~30GB cleaned | Direct | Useful during training, but not the system of record after a model is uploaded or known re-downloadable. |
| Cloudflare R2 | artifact storage role | optional | private cache / legacy mirror | Directional | No longer the default public artifact store for model weights. |
Direct rows share this artifact's eval setup. Directional rows are useful market context but should not be read as leaderboard claims.
Uploaded posttrainllm artifacts
| Artifact | HF repo | Status | Readout |
|---|---|---|---|
| pace-intent-router-v8 | pace-intent-router-v8 | release-ready / rejected winner | 57.1% sealed accuracy at 3.8ms mean; source-matched 95.5% did not generalize |
| mt4b_fused | qwen3-4b-file-ops-distilled | release-ready | File-ops hard gate 58% -> 100%; breadth regression disclosed |
| mt4b_rest_fused | qwen3-4b-rest-fused | routed specialist | Fresh 12/12 vs 9/12 depth win; 25/45 vs 30/45 breadth reject |
| mt4b_mb_fused | qwen3-4b-multibackend-distilled | archive / failed attempt | Negative-transfer comparison artifact: depth 100%, breadth 31% |
| vibethinker-3b-mlx | vibethinker-3b-mlx | archive | Local MLX conversion of the VibeThinker 3B reasoning specialist |
| vibe_distill_fused | vibethinker-3b-agentic-distilled | archive | posttrainllm distilled VibeThinker variant; needs eval promotion before product use |
Deleted upstream caches
| Local cache | Upstream repo | Reason |
|---|---|---|
| mxbai-embed-large-v1 | mixedbread-ai/mxbai-embed-large-v1 | Public upstream cache; no posttrainllm delta |
| qwen3-embedding-0.6b | Qwen/Qwen3-Embedding-0.6B | Public upstream cache; no posttrainllm delta |
| qwen3-vl-2b-instruct | Qwen/Qwen3-VL-2B-Instruct | Public upstream cache; no posttrainllm delta |
Release Blockers
Archive entries are not ship decisions
Public weights can be useful as evidence without being the selected model for Pace or any product lane.
Unblock: Promote only candidates with a current factory run, eval report, package metadata, and routed-use decision.
VibeThinker distilled eval needs promotion
The weights are preserved, but the public artifact should not imply a measured win until the eval evidence is attached.
Unblock: Run the factory eval gate and publish a before/after report before using it as a specialist package.