CLM-v0.1-8B

Last updated: October 17, 2026

Open—$0/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

Maturity5.9/10
Public Apache-2.0 head weights, pip package, local server, API docs, and TypeSafe-compatible endpoint; it still requires an external Qwen3-8B encoder and has no hosted SLA.
Capability5.7/10
Official boards (17 Oct): Decision Index 0.3.2 Full 6.24 (#162); JevBench Intelligence 0.1, calibration 40.0, p50 0.021 s.
Adoption28/100
812 GitHub stars, 65 forks, 1.1K X likes, and 93.8K views show strong launch interest; production adoption is not yet reported.

Vendor claims

Up to 9× lower latency than Jev on computer-use, gaming, and tool-calling tasks[Vendor claim — not independently verified]
81.6% on DeepSWE and 87.6% on Terminal-Bench 2.1 after lightweight fine-tuning[Vendor claim — not independently verified]
4.1–5.7× faster than Jev in the reported agentic verifier setup[Vendor claim — not independently verified]

publicado (OK do Fabricio 17/10 BRT). Board audit of 17 Oct 2026: see the “Board audit” section below; where it conflicts with older board statements on this page, it wins.

CLM-v0.1-8B is an open System One model from Contrastive-LM. It connects a current state to candidate actions with a contrastive objective, then applies a softmax over the candidate scores. The project exposes a TypeSafe-compatible API through CLMClient.system_one().

The public artifact is not a self-contained 8B checkpoint. The Hugging Face repository ships a roughly 75 MB CLM_v0.1-8B.pt containing state/action projection heads. Inference also requires a compatible Qwen3-8B encoder and last-token pooling. The reference package provides contrastive-lm, clm-serve, and a local playground.

Typed interface

The API exposes all three catalog decision types:

Choice — option probabilities;
Noul — probability that a statement is true;
Score — expected position on an ordered rubric.

Probabilities are relative to the candidate set supplied in the request. This is useful for action ranking, routing, retrieval shortlisting, and verifier gates, but it is not the same as a globally calibrated probability over every possible action.

Claims and evaluation — UnverifiedClaim

The project reports zero-shot performance on par with Jev across computer-use, gaming, and tool-calling tasks, with up to 9× lower latency. The README reports 81.6% on DeepSWE and 87.6% on Terminal-Bench 2.1 after lightweight fine-tuning, with 4.1–5.7× lower verifier latency than Jev on an H100 setup.

These are author-reported results. The agentic numbers use fine-tuned heads, not only the released zero-shot checkpoint. The evaluation is also narrow: 38 held-out DeepSWE tasks and 30 held-out Terminal-Bench tasks, according to the repository. Treat the claims as UnverifiedClaim until independently reproduced.

Fit / anti-fit

Fit when you need: local action ranking, tool routing, best-of-N selection, retrieval reranking, or a typed verifier for agent trajectories.

Anti-fit when you need: free-text generation, a standalone checkpoint, long states beyond the configured encoder limit, or globally calibrated confidence independent of the candidate list.

Board audit (17 Oct 2026)

publicado (OK do Fabricio 17/10 BRT). Checked against the official boards on 17 Oct 2026: Decision Index 0.3.2 (data of 16 Oct, 19:00 UTC), JevBench and ImageJevBench live. The boards are separate sources and are never averaged. Older board statements and rankings further down this page are superseded where they conflict with this section.

Board	Row	Result
Decision Index 0.3.2 (16 Oct 19:00 UTC)	CLM-v0.1-8B Qwen3-8B heads	Full 6.24 (#162, tie group 48); public 7.41 (#165); same-skills 5.96; new-domains 11.84
JevBench open weights (live 17 Oct)	CLM-8B	Capability 20.1; Intelligence 0.1; calibration 40.0; p50 0.021 s; $0.045/1k

Score update (audit, 17 Oct 2026). Capability 6.5 → 5.7, first full working (the earlier score had no stored breakdown) = types 10 × 0.30 + calibration 3 × 0.25 + latency 7 × 0.25 + accuracy 1 × 0.20 = 5.70 → 5.7. Types from decisionTypes; calibration from the JevBench axis 40.0 (≥90 → 7, 85–89.9 → 6, 75–84.9 → 5, 60–74.9 → 4, below 60 → 3; no author ECE counted); latency from the board’s self-hosted p50 0.021 s without p99 (≤0.25 s → 7, ≤0.5 s → 6, ≤1 s → 5, above → 3); accuracy Decision Index Full 6.24 ÷ 60.11 × 8 = 0.8, half-up. Maturity and adoption unchanged.

Hugging Face on 17 Oct (Contrastive-LM/CLM-v0.1-8B): 4,915 downloads, 779 likes, licence tag apache-2.0, last modified 2026-09-24.