Open-weight multimodal and computer-use agents

Nex-AGI releases Nex-N2.6 FP8 weights for multimodal computer-use agents

The GLM-derived model targets longer tool-driven workflows; first-party benchmark scores are model-level claims, not independent tests of the published FP8 export.

Open Model Weights published12 Nov 2026
Primary sourceNex-AGI — official Nex-N2.6 checkpoint and model card
Source published2026-11-10
FOLLOW OUR COVERAGE

Follow Open Model Weights on Google See more of our open-weight AI reporting in Google Search.

Add as a Preferred Source on Google
Original cinematic technology poster with an illuminated workstation, agent observation-plan-action-verify circuitry and Nex-N2.6 FP8 weight block.
Original conceptual diagram of a computer-use agent workflow; not a real agent execution trace, product screenshot or observed benchmark. AI-generated editorial illustration; not a documentary photograph.

Nex-AGI has opened Nex-N2.6 model weights for a multimodal agent intended to work through computer interfaces, coding sessions and longer tool-based workflows. The public Hugging Face repository contains FP8-quantized weights for a roughly 321-billion-parameter model derived from GLM-5.3-Flash, and the repository is marked Apache-2.0. This is a concrete downloadable artifact, not merely access to a hosted API, but the FP8 export and its original model-level evaluations must be kept conceptually separate. Primary source ↗

The release focuses on agent reliability: choosing only tools that are actually available, avoiding repeated thinking loops and continuing after environmental feedback. Nex-AGI reports an OSWorld score of 83.0 compared with 82.2 for Nex-N2.5-Pro and 78.0 for the GLM-5.3-Flash baseline. Its table also gives Terminal-Bench 2.1 at 84.1 for Nex-N2.6. These are issuer-provided results with particular harnesses and settings, not a new Open Model Weights run or a verified quality ranking for the released FP8 quantization.

A second table is more operationally revealing. In Nex-AGI's internal known-issue replay suite, reasoning repetition is reported to drop from 16.0% on Nex-N2.5-Pro to 1.6% on the new model, while valid tool selection reaches 100% in the specified test set. That does not demonstrate a zero-error rate for live tools, prompt-injection resistance or safe execution against production accounts. Replayed next steps and real multi-hour agent trajectories impose different failure modes, including permissions, stale state and irreversible actions.

Nex-AGI recommends a GLM-5.3-Flash-compatible runtime such as SGLang and a configuration appropriate for FP8 weights. A checkpoint of this size is an infrastructure decision rather than an on-device model: memory, parallelism, processor compatibility and intended multimodal inputs must be verified against the actual file manifest. Teams should inspect the specific source SHA, licensing of included components and hosted-versus-self-run trade-offs. Open weights do not automatically confer independent evidence of reliability or justify a Leading designation in our curated registry.

OMW REGISTRY WATCH

What this changes for the evidence layer

Source verification candidate for nex-agi/Nex-N2.6; check public original revision SHA, FP8 weight shards and tokenizer, declared Apache-2.0 file versus base GLM-5.3-Flash notices, runtime compatibility and the gap between model-level reported scores and FP8 export. No automatic curation or independent benchmark.

PRIMARY SOURCE

Nex-AGI — official Nex-N2.6 checkpoint and model card

This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.

Open source article ↗