Original title: nex-agi/Nex-N2.6
Nex-AGI is a multimodal agent model trained on GLM-5.3-Flash, designed for computer operation, programming, and long-horizon tasks requiring tools; this repository contains FP8 quantized weights with approximately 321B parameters. Compared to the previous version, Nex-N2.5-Pro, this version primarily reduces redundant reasoning and ensures the model only calls tools declared in the current environment. Officially, internal replay evaluations show that the reasoning repetition rate dropped from 16.0% to 1.6%, with tool selection being 100% compliant in responses involving tool calls; OSWorld scores reached 83.0, higher than Nex-N2.5-Pro's 82.2 and the base model's 78.0.
Why it matters Improvements target two common agent issues: redundant reasoning loops and calling non-existent tools. The FP8 weights are Apache licensed, allowing for self-deployment.
Rising score 62 · Likes 64 · Downloads 191
Reasons
- Improvements target real-world tasks: Internal evaluations were derived from issues encountered by the team and partners in actual tasks, fixing the context and tool outputs to test only the next response step.
- The official announcement also lists areas where it did not outperform others: Terminal-Bench 2.1 scored 84.1 (base model 84.3), and AutomationBench 45.7 scored lower than the base model's 48.8.
- In addition to self-deployment, an API is also available on OpenRouter.
Usage
- Self-deployment: Follow SGLang's deployment guide for GLM-5.3-Flash to select the FP8 configuration, replace the model path with Nex-N2.6, set --served-model-name to Nex-N2.6, and retain the tokenizer and chat template provided in the repository.
- When calling the OpenAI-compatible interface, you can set reasoning_effort to low / high / max (default is max); the tool list must be passed with each request, and tools are executed by the client or agent framework.
Risks
- The repetition rate and tool compliance rates come from internal team evaluations; the evaluation set is not public.
- The FP8 version's performance figures reuse model-level results; the official note states that this exported version was not evaluated separately.
- This entry has not been run; refer to the model card for actual performance.