Overview
Nex-N2.5-Pro is the middle tier of the Nex-N2.5 family Nex AGI published on 8 September 2026, and the one built for driving a computer. Where the larger Nex-N2.5-Max is text-only, Pro is multimodal: it reads the screen, the browser window or the application it just launched, compares what it sees against the goal, and revises when the two do not match.
That focus shows up in the numbers Nex AGI published. Pro takes the top OSWorld-G score in the company's whole comparison table at 87.4, ahead of Qwen3.8-Max on 84.9, GLM-5.3-Flash on 83.3 and Claude Opus 5 on 76.8 — the only benchmark in the release where a Nex model beats every named competitor. It also posts 82.2 on OSWorld-Verified and 67.6 on WebArena-Verified.
Pro is the one tier of the three whose weights had not landed at launch. The Hugging Face repository carries the model card and figures but no safetensors files, and the card itself says the weights are coming soon. Until they arrive, Pro is reachable through hosted access — free on OpenRouter under the model id nex-agi/nex-n2.5-pro:free, and on ModelScope.
On text work Pro sits between its siblings rather than near the top of the field: 82.7 on Terminal-Bench 2.1 and 61.2 on SWE-Bench Pro, against 86.1 and 65.7 for Max. Nex AGI documents serving it on a single node of 8 H100s once the weights are available, using the same SGLang fork and Docker image as the rest of the family.
| Released | 2026-09-08 |
|---|---|
| License | Apache-2.0 |
| Weights | API only |
| Context | 262,144 tokens |
| Max output | 262,144 tokens |
| Architecture | Multimodal mixture-of-experts, built on the Nex-N2 foundation |
| Modalities | Text, Vision |
| Status | Generally available through hosted access (free on OpenRouter and on ModelScope); the model card lists downloadable weights as coming soon |
Benchmarks
Nex-N2.5-Pro on the multimodal benchmarks Nex AGI published
| Benchmark | Nex-N2.5-Pro | Nex-N2.5-mini | Claude Opus 5 | GPT-5.6 Sol | Qwen3.8-Max |
|---|---|---|---|---|---|
| OSWorld-G | 87.4 | 82.9 | 76.8 | 77.7 | 84.9 |
| OSWorld-Verified | 82.2 | 71.2 | 83.4 | 83.2 | 86.1 |
| OSWorld-2 | 56.4 | 30.5 | 68.3 | 62.7 | 46.7 |
| SWE-MM | 38.2 | 25.5 | 59.4 | 40.2 | 39.2 |
This model's scores
- OmniDoc92.2%
- BrowseComp89.7%
- OSWorld-G87.4%
- Terminal-Bench 2.182.7%
- OSWorld-Verified82.2%
- WebArena-Verified67.6%
- SWE-Bench Pro61.2%
- OSWorld-256.4%
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Strengths
- Top OSWorld-G score in Nex AGI's published table at 87.4 — the family's only outright win over every named competitor
- Multimodal by design: the screen is a feedback channel the model checks its own work against, not just an input
- Free hosted access on OpenRouter as nex-agi/nex-n2.5-pro:free, so it can be tried without any GPU
- 262,144-token context on both input and output, the same window as the 1.6T Max tier
- Apache-2.0 licensed, and a single-node 8×H100 serving command is already documented for when the weights land
Best for
- GUI and desktop automation, the task OSWorld-Verified and OSWorld-G measure
- Browser agents that need to see a page rather than parse its HTML
- Front-end and visual QA work, where the check is whether the rendered result looks right
- Trying an agent model at no cost before committing to self-hosted infrastructure
Nex-N2.5 — every version
The full lineage of the Nex-N2.5 line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| Nex-N2.5-Maxcurrent | 2026-09-08 | 262K | Apache-2.0 |
| Nex-N2.5-Pro | 2026-09-08 | 262K | Apache-2.0 |
| Nex-N2.5-mini | 2026-09-08 | 262K | Apache-2.0 |
FAQ
Can I download the Nex-N2.5-Pro weights?
Not yet. The Nex-N2.5-Pro repository on Hugging Face holds the model card and figures but no safetensors files, and the card describes the weights as coming soon. Its two siblings, Nex-N2.5-mini and Nex-N2.5-Max, do have published weight files, so only the middle tier is waiting.
How do I use Nex-N2.5-Pro right now?
Nex AGI offers Nex-N2.5-Pro through hosted access while the weights are pending. OpenRouter serves it free under the model id nex-agi/nex-n2.5-pro:free with the full 262,144-token context, and ModelScope carries it as well. Both accept OpenAI-compatible Chat Completions requests, including the reasoning_effort setting.
Should I pick Nex-N2.5-Pro or Nex-N2.5-Max?
Pick Nex-N2.5-Pro for anything visual. Pro reads screens and leads Nex AGI's table on OSWorld-G at 87.4, while Nex-N2.5-Max is text-only and cannot do GUI work at all. Max is the better choice for terminal agents, coding and agentic search, where it scores 86.1 on Terminal-Bench 2.1 against Pro's 82.7.
What hardware will Nex-N2.5-Pro need once weights ship?
The Nex-N2.5 README documents Nex-N2.5-Pro on a single node with 8 H100 GPUs, tensor parallelism set to 8, using the qwen3 reasoning parser and the qwen3_coder tool-call parser. That is a far smaller footprint than Nex-N2.5-Max, which the same README spreads across two nodes of 16 H200s.