Model Card: Nex-N2.5-mini

TL;DRNex-N2.5-mini (`nex-agi/Nex-N2.5-mini`) is a 35B MoE with about 3B active parameters, a 262,144-token context window, Apache-2.0 licensing, and a free OpenRouter tier. Nex reports large gains over N2-mini on DeepSWE and Terminal-Bench, with visual-feedback self-correction as its central design goal.

#ai-agent #model-release #daily #nex-agi #model-family-nex

Table of Contents

Model information
Highlights
Benchmark context
What it means for agent builders
Comparison with the prior model
Takeaway
References

Series: AI Model Tracker (20 / 24)

Model Card: DeepSeek-V4.1-Flash Model Card｜DeepSeek-V4.1-Flash

🌏 中文版

Model information

Item	Value
Model ID	`nex-agi/Nex-N2.5-mini`; OpenRouter free tier: `nex-agi/nex-n2.5-mini:free`
Parameters	35B total, roughly 3B active; a Qwen3.5-35B-A3B-Base post-training derivative
Context window	262,144 tokens; 235.9K maximum output
Price	$0.00 input and output on OpenRouter's free tier; Nex has not published standalone commercial API pricing
License	Apache-2.0
Released	2026-09-08

Highlights

Nex reports DeepSWE v1.1 improving from 8.0 for N2-mini to 36.1, and Terminal-Bench 2.1 improving from 60.7 to 73.4.
The model is designed for visual-feedback self-correction: it can operate a computer or browser, inspect the result, diagnose a mismatch, and retry.
Nex reports 82.9 on OSWorld-G, ahead of its reported Claude Opus 5 and GPT-5.6 Sol comparisons.

Benchmark context

Benchmark	Nex-N2.5-mini	Prior N2-mini	Reported leading competitor
Terminal-Bench 2.1	73.4	60.7	Claude Opus 5: 89.1
DeepSWE v1.1	36.1	8.0	Claude Opus 5: 73.7
BrowseComp	83.4	74.1	Claude Opus 5: 90.8
AutomationBench v1.0.6	32.3	N/A	Claude Opus 5: 50.3
SWE-Bench Pro	43.8	N/A	Claude Opus 5: 79.2

The release uses Nex AGI's own NexAU/NexCUA harness. Earlier N2-mini figures come from a mix of the project's GitHub material and third-party evaluators, so treat the table as directional rather than a controlled head-to-head comparison.

What it means for agent builders

Nex-N2.5-mini is aimed at computer-use and browser-use loops rather than general chat. Its visual-feedback design and free tier make it a practical model for testing whether an agent can observe a result, correct an action, and retry before committing to a paid provider.

It is less suitable for high-accuracy production coding work: the reported SWE-Bench Pro result is 43.8, well below Nex's Max model and Claude Opus 5. The lack of independently published commercial API terms is also a constraint for workloads that require a production SLA.

Comparison with the prior model

Nex's published gains are largest on longer software-engineering tasks, but the comparisons mix release-harness measurements with historical figures. Use them to form an evaluation shortlist, then run the same browser or computer-use tasks your agent must complete.

Takeaway

The free tier and computer-use orientation make Nex-N2.5-mini useful for prototypes. Its published scores and commercial support details do not yet justify treating it as a drop-in production coding default.