Stanford and NVIDIA Research researchers open-source CLM-8B, saying it can reduce decision latency by up to approximately 9 times

Article

Stanford and NVIDIA Research open-source Jev-like model CLM-8B, up to 9 times faster

Dongcha Beating AI news flash: Researchers from Stanford University and NVIDIA Research have open-sourced the Jev-like System One model CLM-8B, designed specifically to help Agents make judgments and choices quickly.

CLM and Jev perform the same type of task: both allow an Agent to skip long-form text generation and choose directly from several candidate actions. CLM's main difference lies in its underlying architecture: it computes the current state and candidate actions separately. Fixed actions can therefore be precomputed and reused repeatedly, making it faster in scenarios requiring continuous decision-making. In official tests, CLM-8B performed overall close to Jev, with latency reduced by up to approximately 9 times.

The team also used CLM to select answers for Coding Agents. Opus 5 and Fable 5 first generate multiple candidate solutions, after which a fine-tuned CLM selects the best one. The success rate reached 81.6% on 38 held-out DeepSWE tasks and 87.6% on 30 held-out Terminal-Bench 2.1 tasks.

CLM-8B is based on a frozen Qwen3-8B and trains only additional state and action projection heads. The training data includes approximately 60 million question-answer pairs, 30 million hard negative samples, and approximately 1 million Agent trajectories. Both the code and model weights are openly available under the Apache 2.0 license.

Original link https://m.theblockbeats.info/flash/368853

Key points

Researchers from Stanford University and NVIDIA Research have open-sourced CLM-8B.
CLM-8B and Jev are System One models of the same type, used to select from candidate actions.
CLM computes the current state and candidate actions separately, allowing fixed actions to be precomputed and reused.
Official tests report that CLM-8B performs overall close to Jev and reduces latency by up to approximately 9 times.
When CLM was used to select Coding Agent answers, the success rate was 81.6% on 38 held-out DeepSWE tasks.
The success rate was 87.6% on 30 held-out Terminal-Bench 2.1 tasks.
CLM-8B is based on a frozen Qwen3-8B and trains only additional state and action projection heads.
The code and model weights are openly available under the Apache 2.0 license.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

CLM-8B does not generate long answers; it selects one from multiple candidate actions. It completes reusable action computations in advance, so it may be faster during continuous decision-making.

Why it matters to readers

If the official test results can be reproduced, CLM-8B could help Coding Agents screen candidate solutions more quickly.

It can be understood as a fast “selector” for an Agent: other models first propose multiple solutions, and CLM then selects one.

Risks and unknowns

The speed and success rates in the text come mainly from official tests; no independent verification results are provided.
The evidence does not specify the exact official repository addresses or availability details for the code and model weights.
The evidence does not describe CLM-8B's performance on other tasks or in actual products.

Stanford and NVIDIA Research researchers open-source CLM-8B, saying it can reduce decision latency by up to approximately 9 times

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center