CLM-8B: Contrastive Language Model Released as Fast Open-Source Jev Competitor

X postnetrunnerAnnouncement

Jacky Kwok and a Stanford/NVIDIA team have introduced CLM-8B, an Apache 2.0 open-weights model claiming up to 9x faster inference than Jev with comparable task performance.

Key points

  1. CLM-8B (Contrastive Language Model) is a System One model trained with a contrastive learning objective linking states and actions.
  2. The authors claim up to 9x faster inference than Jev across computer use, gaming, and tool calling.
  3. The architecture disaggregates states and actions, enabling independent embedding caching when action sets remain static.
  4. With lightweight fine-tuning, the team reports new benchmarks on DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%).
  5. Model weights and source code are openly available under Apache 2.0 via Hugging Face and GitHub.

Why it matters

Latency is a critical bottleneck in interactive AI agents and real-time computer use. Open-weights models designed specifically for rapid action selection provide an accessible, self-hostable alternative to proprietary inference engines.

Analysis

CLM represents a shift from pure autoregressive generation toward contrastive scoring for agent actions. Decoupling states and actions allows embedding caching that significantly reduces computational overhead. However, claims of parity with models like Jev will require rigorous independent verification regarding accuracy in ambiguous, long-horizon environments.

Original post

netrunner

@plotarmordev · 24 September 2026

Another Jev competitor just dropped: CLM-8B from @jackyk02 and the Stanford/NVIDIA team, with open weights under Apache 2.0. They claim it's comparable to Jev with up to 9x faster inference We went from zero options to a new one almost every day! t.co/i74jjdBoeU

561 likes11 replies634 bookmarks86.2k views

View on X
More posts in the thread (2)
  1. @FreeBTC21SN @jackyk02 looking to use any of these open source models?
  2. @thybead @jackyk02 feels like you do make the same question every time there are improvements on models haha
Selected replies (5)
  • @FreeBTC21SN @plotarmordev @jackyk02 Nice..👍 t.co/zJ0yA84Gwm
  • @thybead @plotarmordev @jackyk02 9x is a pretty big jump. how much of that holds once the prompts get messy?
  • @FreeBTC21SN @plotarmordev @jackyk02 I'll have to take a closer look at it. This could be interesting. Things are moving in the right direction.
  • @natejtom @plotarmordev @jackyk02 Under the hood it works differently from Jev. It turns the state and each answer option into separate vectors using a frozen Qwen3-8B plus a small 20M-parameter trained layer, then picks the option whose vector lines up best with the state. That’s why it’s so fast when the
  • @natejtom @plotarmordev @jackyk02 | | Jev | CLM | | ------------------------- | ----- | ------ | | Survived | 5/5 | 5/5 | | Picked the correct action | 98.7% | 65.8% | | Safety-net overrides | 28 | 4,883 | | Median latency | 150ms | 16.5ms

Mentioned links

Summary generated by AI (Gemini 3.8 Flash, 27 September 2026). It may contain errors – the original source is authoritative.

Related entries

Similar in content, found by the AI search.

  • Article:Nikita Pavlichenko

    JetBrains Introduces Mellum2: A 12B Mixture-of-Experts Model for Code and Text

    JetBrains has released Mellum2, a 12B-parameter Mixture-of-Experts model trained from scratch on natural language and code. Activating only 2.5B parameters per token and published under the Apache 2.0 license, it is designed for low-latency, high-throughput developer and AI workloads.

    AI & LLMs· Announcement

  • Article:Chris Alexiuk and Chintan Patel

    NVIDIA Nemotron 3.5 Lightning: Compact MoE Model for Agentic Execution

    NVIDIA has announced Nemotron 3.5 Lightning, an open 30B mixture-of-experts (MoE) model featuring 3B active parameters. It is tailored for the high-volume execution layer of long-running AI agents, targeting routine tasks such as tool calling and result validation with low latency.

    AI & LLMs· Announcement

  • Article:Moo Jin Kim et al.

    OpenVLA: An Open-Source Vision-Language-Action Model for Robot Control

    Researchers from Stanford, UC Berkeley, Google DeepMind, and partner institutions introduce OpenVLA, an open-source 7-billion-parameter vision-language-action model. Trained on 970k real-world robot demonstrations, it surpasses closed models such as RT-2-X while enabling parameter-efficient fine-tuning on consumer-grade hardware.

    AI & LLMs· Research

Back to the knowledge base

Let's talk about your project

Get in touch

Locations

  • Mattersburg
    Johann Nepomuk Bergerstraße 7/2/14
    7210 Mattersburg, Austria
  • Vienna
    Ungargasse 64-66/3/404
    1030 Wien, Austria

Parts of this content were created with the assistance of AI.