Nex-N2.5-Mini: A 35B Apache-Licensed Agent Model on Two GPUs

Most open agent models worth talking about in 2026 need a rack to run. Nex-N2.5-Mini is interesting for the opposite reason: it is a multimodal model aimed at computer use, browsing and coding agents, it ships under Apache 2.0, and Nex AGI's reference deployment is a single node with two H100s.

That combination makes it a candidate for a job the big models are bad at: running agents somewhere you control, at a cost you can predict. Whether it is good enough for your agents is a separate question, and the vendor's benchmark table answers it less clearly than it appears to.

What Nex shipped

Nex AGI released the N2.5 family on 8 September 2026 in three sizes. Mini and Pro are multimodal and, according to the model card, build on the multimodal foundations of the earlier Nex-N2. Max is a separate, much larger text-only model built on a 1.6-trillion-parameter mixture-of-experts base.

Mini's Hugging Face repository lists 35B parameters and an architecture tag of qwen3_5_moe, so it is a mixture-of-experts model. The card does not state the active parameter count, and Nex does not name the base model in so many words; treat third-party claims about either as unconfirmed. The licence is Apache 2.0, which means fine-tuning, redistribution and commercial serving without a separate agreement. Open-weight model licences compared covers why that is rarer at this capability level than it should be.

OpenRouter lists a 262K context window and image input at the time of writing. The catalogue's maximum output figure sits almost on top of the full window, which tells you little about what you will actually get back; set your own max_tokens.

Running it yourself

The deployment notes are the most practical part of the card. Nex publishes a patched SGLang image and a reference command for two H100s with tensor parallelism of two. It uses the qwen3 reasoning parser and the qwen3_coder tool-call parser, and the card also mentions vLLM and community GGUF quantisations.

Two things follow. First, the custom SGLang fork matters: if you serve it with a stock build and tool calls come back as plain text, the parser configuration is the first thing to check. Second, two H100s is the reference, not the floor. Quantised builds bring it within reach of smaller hardware, at a quality cost you should measure rather than assume. Best model for self-hosting covers how to size that decision.

Recommended sampling is temperature 0.7, top-p 0.95 and top-k 40, which is also what Nex used for its own evaluations. Running at temperature 0 because that is your default for agents will put you outside the conditions the published numbers describe.

Reasoning effort: three modes, one model

Mini exposes a reasoning_effort parameter with three settings. none answers directly. medium, the default, lets the model decide whether to think. high always thinks before responding.

For agent loops this is more useful than it sounds. Many steps in a computer-use task are trivial — click the button that is plainly labelled — and spending reasoning tokens on each one adds latency to every turn. A reasonable starting point is none for clearly mechanical tool steps and medium elsewhere, with high reserved for planning or recovery after a failure. Measure it: if none fails a step type more than occasionally, that step type was never as mechanical as it looked.

Reading the benchmark table carefully

The card reports, for Mini, 73.4 on Terminal-Bench 2.1, 43.8 on SWE-Bench Pro, 83.4 on BrowseComp and 71.2 on OSWorld-Verified, alongside a long list of coding, browsing and computer-use evaluations.

Three caveats before you lean on any of it:

• Harnesses. Coding tasks were run in Nex's own NexAU harness and computer-use tasks in NexCUA. A vendor harness tuned for its own model usually scores above a neutral one.
• Mixed sources. The card says comparison scores come from leaderboards, provider reports or Nex's own runs. Columns assembled that way are not strictly comparable.
• No independent reproduction yet. At the time of writing these are vendor figures. Vendor-claimed vs independent benchmarks covers how much to discount.

The most useful reading of the table is relative rather than absolute. Within its own family, Mini trails Pro on every reported benchmark, by a small margin on some and a large one on others. That pattern is more trustworthy than any single number, because the harness bias applies to both.

Hosted price, and one capability to check

At the time of writing OpenRouter lists Mini at $0.025 per million input tokens and $0.10 per million output, with cached input at $0.0025. That is cheap enough that cost is unlikely to be the deciding factor for most workloads; latency and reliability will be.

One discrepancy to check before you build on it: the model card documents function calling, but OpenRouter's capability flags for Mini did not include tool calling when this was written, while Pro's did. Whether tools work on a hosted endpoint depends on the provider's configuration. Send a tool-call request and inspect the response before wiring it into a loop.

Mini or Pro

The Nex-N2.5-Pro guide covers the larger sibling in detail. The short version:

• Pick Mini when you are self-hosting on limited hardware, when steps are short and well defined, or when per-step latency matters more than recovering from hard failures.
• Pick Pro when tasks are long or open-ended, when you need reliable tool calling on a hosted endpoint, or when a failed run costs more than the price difference. Pro needs a full eight-GPU node to self-host.

The honest way to decide is to run both on twenty of your own agent tasks and compare completed tasks per pound, not price per token.