- Who it helps
- Teams with server-scale inference capacity evaluating agent reliability
- Availability
- Apache-2.0 model card lists approximately 321 billion parameters in FP8 weights, based on GLM-5.3-Flash. Hosting requirements and API access were not tested. FP8 stores numbers in an 8-bit floating-point format.
A focused reliability change
The team reports reduced repetition and stronger adherence to a provided tool list. Its checkpoint-replay test fixes the preceding conversation and evaluates the next response.
Replay is only part of a workflow
The card says scores are model-level results rather than a separate evaluation of this FP8 export. A successful next response does not establish a reliable complete coding or computer-use session.
What you can do
- Review the model card’s memory, serving and evaluation caveats. Open reference ↗
Where the evidence stops
Model card and publisher activity checked November 11. No weights downloaded, inference served or reliability test run.
Sources read. Public source text checked November 11, 2026 UTC. No third-party project or complete workflow executed.
Sources & further reading
Hugging Face · official
Nex-N2.6 model card ↗
FP8 model card and internal replay caveats read. No weights loaded.
Nex AGI · page checked 2026-11-11
Coverage note: Model card and publisher activity checked November 11. No weights downloaded, inference served or reliability test run.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
Nex-N2.6 is an open-weight multimodal model intended for computer use, coding and long tasks. Its release focuses on reducing repeated reasoning and calls to tools that were never supplied.