CLM-v0.1-8B — MLX encoder

The encoder half of Contrastive-LM/CLM-v0.1-8B, unquantised bf16 for MLX on Apple Silicon. This repo is the reference every width below is measured against, not a quantisation of it.

⚠️ Not recommended for ranking: 4bit. agreement on decisive (>1 nat) decisions 0.889651 < 0.995; planner accuracy falls 31.95 points, beyond the 3.0-point budget. Published with the measured numbers so the result is reproducible, not because it is a drop-in replacement for the bf16 encoder.

Repositories

One repository per bit width, matching mlx-community. Each has the weights at the repo root, so the Hub file browser lists every file with its size and mlx_lm.load("<repo>") works.

A parallel outq2 line ships each width again with lm_head at 2-bit instead of the bulk width. Pooling never evaluates the output projection, so those are indistinguishable from re-running the repos above — measured over the full corpus against a matched null, since MLX is not run-to-run deterministic on the GPU — and smaller for free:

Pooling only — a 2-bit lm_head is a wrecked output projection and is not for generation.

Evaluation

Agreement against a bf16 MLX reference of the same encoder in the same runtime, over 23,926 scored System One questions, through the real head stack (argmax(scale · cos), scale = 100.0 — cosine error is amplified 100×).

Pass line. top-1 >= 1.0000 — the measured bf16-vs-bf16 noise floor of this corpus in this runtime — and top-1 (decisive) >= 0.995, where decisive means the reference's own top-1 led by more than 1 nat. A third condition applies: planner accuracy Δ, the change in accuracy against the T-Rex physics planner's label. A variant that holds every decisive decision but still loses measurable accuracy is marked usable rather than recommended.

Against the numbers published in the parent model card

How to use

Each per-width repo is a standard mlx-community layout. Load one directly:

mlx_lm.load("czl/CLM-v0.1-8B-MLX-8bit")

Or serve it for CLM:

python -m clm_mlx.server --model <dir> --port 8092
clm-serve --emb-url http://127.0.0.1:8092/v1/embeddings

Quantization details

Tool: mlx-lm 0.31.3, one mlx_lm.convert pass per width:

mlx_lm.convert --hf-path Qwen/Qwen3-8B --mlx-path <dir> -q --q-bits N --q-group-size 32

• affine mode, per-group bf16 scale and bias.
• group_size = 32 — measured best at every width. mlx-lm's default (64) and Qwen's mlx-community repos (128) are both worse here. See the per-width cards for the full group-size ladder.

Limitations

• The projection heads are encoder-locked to Qwen3-8B last-token-pooled 4096-d embeddings.
• This is a ranker, not a generator. lm_head is never evaluated under pooling.
• The head stack amplifies cosine error 100×. Read the confident-decision column.
• Probabilities are relative to the candidate set.

License

Apache-2.0, inherited from Qwen/Qwen3-8B and Contrastive-LM/CLM-v0.1-8B. Not gated.

Citation

@misc{kwok2026contrastivelanguagemodels,
  title={Contrastive Language Models: A System One Model for Fast and Generalizable Decision-Making},
  author={Jacky Kwok and Hangoo Kang and Tarun Suresh and Jon Saad-Falcon and Marco Pavone and Christopher Ré and Azalia Mirhoseini},
  year={2026},
  note={Notion Blog},
  url={https://contrastive-lm.notion.site}
}