Nex-N2.5 Mini: Open Weights Landed at Launch — Nex-AGI's Smallest Computer-Use Agent Is the On-Ramp

The way to find out whether open computer-use agents are ready is now a model you can pull the same day it was announced. Nex-AGI introduced its Nex-N2.5 family on September 8, 2026 — three agentic tiers, Nex-N2.5 Mini, Nex-N2.5 Pro and Nex-N2.5 Max — and the tier with the download-and-run story is the smallest one. Nex-N2.5 Mini has its Apache-2.0 weights live on Hugging Face in sixteen BF16 shards, roughly 35 billion parameters with a reported ~3 billion active per token, and a two-H100 reference deployment. Its larger multimodal sibling Nex-N2.5 Pro is still marked "coming soon" as of this writing, while the text-only 1.6-trillion-parameter Nex-N2.5 Max is also up but wants sixteen H200s across two nodes before it will run. For anyone testing the thesis behind the family — that open agentic models can hold up in long-horizon computer use rather than only answer questions — the Mini is the on-ramp.

Who is Nex-AGI? The name is new, and the launch has the flavor of a first public release. Coverage of the announcement describes the effort as a collaboration across several Shanghai organizations — the Shanghai Innovation Institute, Shanghai Qiji Zhifeng, Mosi Intelligence and Kuafu Technology — with the family positioned as open-source and the team operating under the @NexEcosystem handle. The model card credits Nex-N2.5-mini and Nex-N2.5-Pro as continuing "the multimodal foundations of Nex-N2," the earlier Nex release whose weights are already open. What is genuinely new here is the framing: Nex-AGI says it built these models for long-horizon tasks in real-world environments — operating a computer, driving a browser, executing and testing code, and correcting course from what it sees on screen — and it published the weights to make that claim testable.

Three tiers, one announcement

The family splits by scale and by modality, and the differences matter more than the shared name:

• Nex-N2.5 Mini — roughly 35B total / ~3B active, a multimodal model accepting images and text, aimed at visual computer use, browsing and tasks that need feedback from an interface. Reference deployment is a single two-H100 node.

• Nex-N2.5 Pro — reported 397B total / ~17B active, also multimodal, for heavier computer-use workloads. Reference deployment is eight H100s. Its weights were not downloadable at launch.

• Nex-N2.5 Max — 1.6T total / ~49B active, text-only, and by Nex-AGI's account its "first complete post-training effort at trillion-parameter scale." The card says it is built on a DeepSeek-V4-Pro-Base foundation. Reference deployment is 16×H200 across two nodes.

All three are mixture-of-experts. The Mini and Pro sit on the Qwen3.5 model family — Nex-AGI's launch materials describe both as post-trained from Qwen3.5 variants, and the Mini's own config carries a qwen3_5_moe architecture tag — while Max is the DeepSeek-based outlier. So the family is not one recipe at three sizes; it is two recipes, with the multimodal "do things on a screen" tiers sharing one lineage and the text reasoning giant on another.

What the Nex-N2.5 Mini that shipped actually is

The Hugging Face repo for nex-agi/Nex-N2.5-mini is a real, downloadable checkpoint, not a placeholder — 16 safetensors shards at about 70 GB total, BF16, license Apache-2.0, first created September 8. The config file is specific enough to design against:

• Architecture — Qwen3_5MoeForConditionalGeneration, the qwen3_5_moe family, with a vision stack: the card tags it image-text-to-text, and the repo ships preprocessor_config.json alongside the text config.

• Shape — 40 layers, hidden size 2048, grouped-query attention with 16 query heads and 2 KV heads, a vocabulary of 248,320.

• MoE — 256 routed experts with 8 active per token plus a shared expert, intermediate size 512 — the sparsely-activated profile that makes a ~3B-active "35B-class" model runnable on two H100s.

• Context — max_position_embeddings of 262,144 tokens in the released config, the same 262K figure that shows up across the family's deployment recipes.

• Weights and license — BF16, Apache-2.0, not gated; the repo also points to a ModelScope mirror and to a GitHub org (nex-agi) that carries the family's deployment recipes.

Nex-AGI's deployment documentation is the part most open releases get wrong and this one gets unusually right. The Mini is served through a custom SGLang Docker image (nexagi/sglang:v0.5.18-nex-patch) on a single node with two H100s using tensor parallelism 2. Recommended sampling is temperature 0.7, top_p 0.95, top_k 40. Tool calling runs through SGLang's qwen3_coder parser, and the model exposes three reasoning modes via a reasoning_effort field — "none" for non-thinking, "medium" (the adaptive default) and "high" for always-on thinking. For a small agentic model, that thinking-mode control is the difference between a usable agent loop and a slow one, and it is baked into the serving recipe rather than left to the user.

Weights: what is actually downloadable

The launch-day status of the three tiers is worth stating precisely, because the announcement and the repositories do not quite line up. As of this writing the Mini is fully downloadable under Apache-2.0 — that is what this article is built on. Nex-N2.5 Max is also up, its Hugging Face repository carrying 644 safetensors shards, though reproducing its trillion-parameter footprint is a 16×H200 project. Nex-N2.5 Pro is the holdout: its Hugging Face repository exists but contains only the README and figures, no model shards, and the card still says the weights are coming. If the tier you care about is the big multimodal one, that is the gap to watch — the launch materials describe Pro as the strongest computer-use model of the family, and it is the one you cannot run locally yet.

The numbers Nex-AGI reported — and the caveat that travels with them

Nex-AGI's model card includes a full benchmark table for the Mini, and every figure in it is the vendor's own reporting. No independent run had been published as of this writing, and the card is explicit that scores come from official benchmark leaderboards and the latest evaluation reports where available, and Nex-AGI's own evaluations elsewhere. Coding results were produced with the team's NexAU harness; computer- and browser-use results used the NexCUA harness, which the card says will be open-sourced. Treat the headline rows as a vendor's best case until a neutral party reproduces them.

With that framing, the Mini's reported rows are: on text-and-coding work, Terminal-Bench 2.1 at 73.4, SWE-Bench Pro at 43.8, DeepSWE v1.1 at 36.1, AutomationBench v1.0.6 at 32.3, Toolathlon Verified at 54.6, BrowseComp at 83.4, and a GDPval-AA v2 score of 1446. On the multimodal/computer-use side, the attention-getters are OSWorld-Verified at 71.2, WebArena-Verified at 63.4, WebTest at 48.6 (run in oracle mode against ground-truth checklists), OSWorld-G at 82.9 and OmniDoc at 89.7, alongside more modest OSWorld-2 (30.5) and Vision2Web (52.9) results.

Two reading notes. First, OSWorld-Verified and OSWorld-2 are different suites — one is a verified subset graded from the resulting environment state, the other a newer and generally harsher run — so a 71.2 on one and a 30.5 on the other are not contradictory; they are different tests, and Nex's own table keeps them in separate columns. Second, in every row of the family table the Mini scores below Pro and Max, and the top of each column belongs to a frontier incumbent such as Claude Opus 5 or GPT-5.6 Sol. The Mini's story is not that it beats frontier models; it is that a ~3B-active open model reports competitive scores on computer-use benchmarks at a two-H100 footprint. Whether that holds up is precisely the question the open weights let you answer yourself.

Running Nex-N2.5 Mini today

The two-H100 recipe makes the Mini the cheapest way to get hands-on with the family's core claim. Because the architecture is standard Qwen3.5-MoE with a qwen3_coder tool-call parser, it loads in the Nex-AGI SGLang fork without custom inference code, and the recommended settings are published rather than left to reverse-engineering. The honest expectation to set before you start is that a day-old checkpoint with a brand-new harness is an experiment, not a dependency. Downloads on the Mini repo are still in single digits and no third party had shipped an independent evaluation when this was written — which is the normal state for a model released yesterday, and the reason the two-H100 footprint matters: you can run the experiment yourself this week instead of waiting for someone else to.

If you would rather compare the Mini against the frontier agents in its own benchmark table than hand-tune a single deployment, that is where a routing layer earns its keep. When a hosted provider lists Nex-N2.5 Mini — and the incumbents it is measured against are already reachable through routers — one API across 200+ models, with provider list prices passed through at 0% markup and automatic failover, is the cheap way to run the same task against several of them without a second integration. On OrcaRouter that is the standing setup for every model we route: the same key, the same call shape, the vendor's own price with nothing added. That matters most for an unproven model like this one, because failover is what lets you try it on a real path without betting the path on it.

Who should pull these weights

Pull them if your work is computer-use agents, browser automation, or visually-grounded tool use and you want a permissively-licensed, self-hostable model to benchmark against the hosted leaders. The Apache-2.0 license removes the usual friction for a Chinese-lab release, the two-H100 footprint is within reach of a serious agent team, and the reasoning-effort control plus a real tool-call parser make it a credible testbed rather than a toy. Hold off if you need the strongest computer-use model in the family — that is Pro's role, and Pro's weights are the ones still pending. And keep the vendor label on every benchmark row until an independent run confirms them; a 71.2 on OSWorld-Verified from a ~3B-active open model is a striking claim, which is exactly the kind of claim worth verifying in your own harness before you build on it.

What to watch next. The Pro weight drop is the single biggest signal — it turns the family from "Mini plus a trillion-parameter giant" into the full lineup the launch materials describe. Second is the open-sourcing of the NexCUA harness, without which the computer-use numbers cannot be independently reproduced under the same protocol. Third is the first neutral run, from Artificial Analysis, a university lab or a practitioner who publishes their OSWorld-Verified reproduction. When any of those three land, the question this launch poses — whether open agentic models have genuinely closed the gap to hosted computer-use stacks — stops being a matter of vendor tables.