Open-weight agent models and the NexAU framework, from Shanghai
Nex AGI posted its Nex-N1 paper on 4 December 2025 with 64 listed authors and shipped Nex-N2.5 in three sizes in September 2026. It suits teams that want open-weight agent models to self-host, but it discloses no funding, leadership or headcount, and it builds on other labs' base models.
| Model | Lab | Context | In / out per 1M | Reads and writes |
|---|---|---|---|---|
| Nex AGI | 262K | $0.12 / $0.40 | text, image → text |
| Product | What it is |
|---|---|
| Nex-N2.5 model family | Open weight agentic models |
| Nex AGI on Hugging Face | Open weight model repositories |
| Nex-N2.6 | Hosted model |
Every model Hokai holds for this company that carries a release date, newest first.
| Company | Type | Based in | Founded | On Hokai |
|---|---|---|---|---|
| Research lab | Shanghai, China | -- | 1 | |
| Research lab | Hangzhou, China | 2023 | 8 | |
| Research lab | Beijing, China | 2023 | 6 |
Hokai found no funding amount, investor or valuation for Nex AGI on its website, its Hugging Face organisation or its arXiv paper, all read on 11 November 2026. The only backer named anywhere is the Shanghai Innovation Institute, which the GitHub organisation says initiated the alliance.
Nex AGI began as an alliance started by the Shanghai Innovation Institute, per its GitHub organisation, and the website footer names Shanghai Qiji Zhifeng Intelligent Technology Co., Ltd. as the legal entity. Hokai found no stock ticker or listing, and no named chief executive.
Mostly yes: Hugging Face's API, read on 11 November 2026, shows an apache-2.0 tag on the N2.5 Mini, Pro and Max checkpoints, on N2 Pro and on N1. The newest model, Nex-N2.6, has no repository in the Nex AGI Hugging Face organisation, so its weights were not available when Hokai checked.
Nex AGI's own Nex-N2.5 page says Nex-N2.5 Max is built on DeepSeek-V4-Pro-Base, and the Nex-N1 repositories are named after DeepSeek V3.1, Qwen3 and InternLM3. OpenRouter's listing says Nex-N2.6 is built on GLM-5.3-Flash. Hokai found no Nex AGI model trained from scratch.
OpenRouter lists Nex-N2.5 Mini, Nex-N2.5 Pro and Nex-N2.6 under the nex-agi namespace, and the Nex-N2.5 page links to the Pro and Mini listings through its Try Now button. For weights, the Hugging Face organisation hosts the N1, N2 and N2.5 repositories.
Nex AGI is an open-source AI group that, per its GitHub organisation, the Shanghai Innovation Institute started as an alliance. It publishes the Nex-N1, Nex-N2 and Nex-N2.5 agentic model families under Apache 2.0, and its GitHub organisation was created in November 2025.
Nex AGI, written Nex-AGI on its own pages, is an open-source group that builds agentic language models and the tooling to train them. Its GitHub organisation, read through GitHub's API on 11 November 2026, says the group was initiated by the Shanghai Innovation Institute as an alliance with a stated aim of an open-source, closed-loop system driven by agents, and dates the organisation to 2 November 2025. The website footer names the legal entity as Shanghai Qiji Zhifeng Intelligent Technology Co., Ltd. Hokai found no founding date, chief executive, headcount or funding disclosed on the website, the Hugging Face organisation or the arXiv paper.
The lineup is a run of post-trained models. Nex-N1 appeared on Hugging Face in November 2025 on top of other labs' bases (DeepSeek V3.1, Qwen3 and InternLM3 variants), and a paper from the "Nex-AGI Team" with 64 listed authors followed on arXiv on 4 December 2025, describing the NexAU agent framework and two data pipelines named NexA4A and NexGAP. Nex-N2 repositories date from 3 June 2026. Nex-N2.5 came in September 2026 in three sizes: Mini and Pro are multimodal and aimed at computer use and browsing, and Max is a 1.6-trillion-parameter text-only mixture-of-experts model built on DeepSeek-V4-Pro-Base, according to the Nex-N2.5 page. The Hugging Face cards for Nex-N2.5 Mini, Pro and Max, Nex-N2 Pro and Nex-N1 list the Apache 2.0 licence.
The latest move is Nex-N2.6, listed on OpenRouter on 10 November 2026. OpenRouter's listing says it is built on GLM-5.3-Flash with about 321 billion parameters. The Hugging Face API, read on 11 November 2026, lists no Nex-N2.6 repository in the Nex AGI organisation, and Hokai found no Nex AGI statement about weights for it.
Nex AGI's weakness is dependence. Every model Hokai could trace starts from another lab's base: DeepSeek for N1 and N2.5 Max, and Z.ai for N2.6. On its own Nex-N2.5 chart, which is vendor-reported, Anthropic's Claude Opus 5 is shown at 89.1 on Terminal-Bench 2.1 against 86.1 and 82.7 for the two Nex-N2.5 text models, and at 79.2 on SWE-Bench Pro, so a buyer who needs the highest score should look at Anthropic first. Teams that want the same open-weight agent focus from a lab with its own pretraining should compare Moonshot AI and its Kimi K3.
What the company has not published is as useful to a buyer as what it has: no trust centre, certifications, data-region list, privacy contact or safety policy turned up on the pages Hokai read, so any commercial use should run through a hosted provider's own terms or a self-hosted deployment.
Everything on this page (funding, leadership, compliance posture, the product list) is read from Nex AGI's own pages and filings, or from the Hokai records for the products it ships. A field the company does not publish is left off the page rather than guessed at.