Nex-AGI, an agentic AI team based in Shanghai, released Nex-N2.6 on November 10, 2026: an open-weight multimodal model aimed at computer use, coding and long tasks that chain many tool calls together. The weights appeared on Hugging Face under the Apache-2.0 licence that day, alongside the GitHub repository sii-nex/Nex-N2.6 and a paid endpoint on OpenRouter. The published checkpoint is FP8-quantized and holds roughly 321 billion parameters.
Nex-N2.6 was not trained from scratch. Its model card says it is built on GLM-5.3-Flash, the open model from Z.ai, and it succeeds Nex-N2.5-Pro, the team's previous release. The goals of this version are narrow and practical: stop the model from circling in its own reasoning, and keep it from calling tools the current environment does not offer. It accepts text and image input, and its chat template exposes three reasoning_effort levels, low, high and max, with max as the default.
The headline numbers come from Nex-AGI itself. The team reports 83.0 on OSWorld, the computer-use benchmark, up from 82.2 for Nex-N2.5-Pro and 78.0 for GLM-5.3-Flash. On Terminal-Bench 2.1 it lists 84.1, a fraction below the base model's 84.3, and on AutomationBench 45.7, behind both GLM-5.3-Flash (48.8) and DeepSeek-V4.1-Flash (54.8). In internal replays built from failures that the team and its partners hit in real work, the reasoning repetition rate fell from 16.0% to 1.6%, and every response that made a tool call picked a tool declared in the request. The card notes that scores depend on harness and settings and were measured at model level, not on this FP8 export.
The first outside test arrived a day later. Plotly, which builds agentic analytics products, ran Nex-N2.6 through its private data-analysis exam via OpenRouter, where Nex-AGI is the only provider. The model passed 38 of 43 questions, exactly the score of GLM 5.3 Flash. It was perfect on the 20 routine analytics questions, the ambiguous prompts, the impossible requests and the statistics traps, but found only one of five hidden patterns planted in the data, against four for its base model.
That run also exposed a failure mode that sits awkwardly next to the release's main claim. On two hidden-pattern questions, Nex-N2.6 kept issuing small variations of the same query: about 1,670 calls over 2 hours and 20 minutes in one case, about 650 calls over 70 minutes in the other. Each attempt ended only when the conversation overflowed the 262,144-token context window. Those loops cost roughly $9.80 of a $10.19 run at list price, against $0.69 for GLM 5.3 Flash. Plotly observed that Nex-AGI measured repetition inside the model's reasoning, whereas what its harness saw was repetition in tool use, and it cautioned that a single run is one sample.
For teams that want to host it, the model card points to SGLang's GLM-5.3-Flash deployment guide with an FP8-compatible configuration, exposing an OpenAI-compatible API. Tool execution stays with the client or agent framework, which feeds results back to the model. On OpenRouter, launch pricing is $0.12 per million input tokens and $0.40 per million output tokens, a 20% discount on list prices of $0.15 and $0.50.
Nex-N2.6 adds another large open-weight option tuned specifically for agent work on top of an existing base model. Whether its reliability gains hold outside Nex-AGI's own replays is still an open question: so far, the only independent evaluation shows accuracy level with the base model and an expensive tendency to loop during open-ended exploration.