Stanford and Nvidia researchers have released CLM-8B, an open-weight model that ranks actions supplied to an AI agent instead of generating open-ended text. In project-reported tests, CLM-8B had lower latency than Jev across four tasks, while Jev scored higher on tool calling and WikiRacing.
CLM-8B is built to rank agent actions
A surrounding agent system gives CLM-8B a representation of the current state and a set of possible actions. The model scores those candidates; another part of the system can then act on the ranking. That makes CLM-8B a decision component for tasks such as routing, triage, and ranking—not a general-purpose text generator.
Its design starts with a frozen Qwen3-8B encoder and adds separate projection heads for states and actions. Contrastive learning trains those representations to match compatible state-action pairs more closely than mismatched pairs. The project says training used about 60 million question-answer pairs, 30 million synthetic hard negatives, and roughly one million agent trajectories. CLM-8B’s weights are listed under the Apache 2.0 license.
How CLM-8B scores and reuses actions
CLM-8B maps the supplied state and candidate actions into representations, then scores how well they match. If an application keeps using the same action set, it can cache the action representations and reuse them as the state changes. That avoids encoding the same options again for each decision.
What the reported zero-shot tests show
In the project-reported comparisons, CLM-8B returned lower latency than Jev on all four listed tasks. The task results were not uniform: the two models tied on T-Rex and Super Mario, while Jev scored higher on BFCL v4 and WikiRacing.
| Task | CLM-8B latency | Jev latency | CLM-8B task result | Jev task result |
| T-Rex | 16.5 ms | 149.8 ms | 5/5 | 5/5 |
| BFCL v4 tool calling | 76.8 ms | 125.5 ms | 95.2% | 99.2% |
| WikiRacing | 79.8 ms | 225 ms | 26/30 | 30/30 |
| Super Mario | 33.5 ms | 132.6 ms | 5/5 | 5/5 |
The project’s “up to 9× faster” claim describes reported test conditions, not every workload. The largest reported gains were associated with reusable actions or larger candidate sets, where caching can reduce repeated work.
Coding results come from verifier experiments
Separate coding experiments used fine-tuned CLM heads to rank candidate solutions generated by larger models. Opus 5 supplied candidates for DeepSWE, and Fable 5 supplied them for Terminal-Bench 2.1; CLM ranked the candidates rather than solving the tasks from scratch.
| Benchmark and held-out subset | CLM verifier result | Jev result | CLM verifier latency | Jev latency |
| DeepSWE, 38 tasks | 81.6% | 71.1% | 79 ms | 449 ms |
| Terminal-Bench 2.1, 30 tasks | 87.6% | 83.1% | 32 ms | 131 ms |
These subset results use a different setup from the zero-shot task tests: the CLM heads were fine-tuned, and the candidates came from other models.
Where a candidate scorer fits in an agent workflow
CLM-8B’s probability scores are relative to the candidates supplied in a request. If every option is poor, it still ranks that set; the surrounding workflow must account for the possibility that none is suitable. The model also depends on an external agent system to provide the state and available actions.
That bounded role suits decisions with a defined menu of choices. It does not replace a general-purpose model for open-ended reasoning, long-form writing, or high-level planning.
A screen-recorded support-ticket walkthrough illustrates how candidate scoring can sort urgency and route a request.
The project’s next stated step
On September 25, CLM project lead Jacky Kwok said the team was training a multimodal CLM-35B-A3B and planned to release it in early October 2026.
