Overview
Nex-N2.5-Max is the top tier of the Nex-N2.5 family that Nex AGI published on 8 September 2026, and the company describes it as its first complete post-training run at trillion-parameter scale. The Hugging Face weight index adds up to 1,600,787,478,430 parameters across 644 safetensors shards, with the bulk of them stored in FP8 and a further 53 billion in BF16.
Unlike its two smaller siblings, Max is text-only. Nex AGI built the mini and Pro tiers as multimodal models that read a screen, and kept Max on text so the post-training budget went into reasoning, tool use and long-horizon planning instead. In practice that means Max is the tier to reach for on terminal work, agentic search and knowledge tasks, and the wrong tier for anything that involves looking at a GUI.
The published config declares a `DeepseekV4ForCausalLM` architecture with 384 experts and 6 active per token — a sparse mixture-of-experts layout, so only a small slice of that 1.6T runs on any given token. The config allows positions up to 1,048,576, but the serving command Nex AGI documents sets a 262,144-token context, which is the figure to plan against.
Nex AGI publishes a comparison table putting Max against Claude Opus 5, GPT-5.6 Sol, Kimi-K3, GLM-5.3, DeepSeek-V4-Pro-0813 and Qwen3.8-Max. Max takes the top BrowseComp score in that table at 92.6 and lands close to Claude Opus 5 on AutomationBench v1.0.6, but trails on coding: 65.7 against 79.2 on SWE-Bench Pro.
| Released | 2026-09-08 |
|---|---|
| License | Apache-2.0 |
| Weights | Open weights |
| Parameters | 1,600,787,478,430 total · 6 of 384 experts active per token |
| Context | 262,144 tokens |
| Architecture | Mixture-of-experts (DeepseekV4ForCausalLM), 384 experts, 6 active per token, FP8 weights |
| Modalities | Text |
| Status | Generally available — open weights on Hugging Face |
Benchmarks
Nex-N2.5-Max against the models Nex AGI published it with
| Benchmark | Nex-N2.5-Max | Claude Opus 5 | GPT-5.6 Sol | Kimi-K3 | Qwen3.8-Max |
|---|---|---|---|---|---|
| Terminal-Bench 2.1 | 86.1 | 89.1 | 88.8 | 88.3 | 86.6 |
| SWE-Bench Pro | 65.7 | 79.2 | 64.6 | 63.3 | 67.7 |
| DeepSWE v1.1 | 65.6 | 73.7 | 72.7 | 67.5 | 69.3 |
| AutomationBench v1.0.6 | 50.2 | 50.3 | 45.8 | 46.7 | 39.8 |
| Toolathlon Verified | 74.7 | 76.5 | 74.9 | 76.5 | 72.5 |
| Job Bench | 53.6 | 65.7 | 45.4 | 52.9 | 53.4 |
This model's scores
- BrowseComp92.6%
- Terminal-Bench 2.186.1%
- Toolathlon Verified74.7%
- SWE-Bench Pro65.7%
- DeepSWE v1.165.6%
- Job Bench53.6%
- AutomationBench v1.0.650.2%
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Strengths
- Best BrowseComp score in Nex AGI's published table at 92.6, ahead of GLM-5.3 (91.2) and Claude Opus 5 (90.8)
- Apache-2.0 weights at trillion-parameter scale, so a tier normally locked behind a closed API can be self-hosted
- Sparse mixture-of-experts — 6 of 384 experts fire per token, so serving cost tracks the active slice, not the 1.6T total
- FP8 weights out of the box, which is what makes a 1.6T model fit on two nodes of H200s
- 262,144-token served context, with an SGLang launch command published rather than left to the reader
Best for
- Long-horizon terminal and shell agents where a run spans many steps
- Agentic web research, the task BrowseComp measures and Max leads on
- Self-hosted deployments that need frontier-scale behaviour without sending data to a hosted API
- Evaluation work comparing open weights against closed flagship models on the same harness
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
Why is Nex-N2.5-Max text-only when the smaller models are not?
Nex AGI built Nex-N2.5-mini and Nex-N2.5-Pro on a multimodal base so they can read a screen, and kept Nex-N2.5-Max on a text-only mixture-of-experts foundation. Max is the tier for terminal work, agentic search and knowledge tasks; anything that needs the model to look at a GUI belongs on Pro.
What hardware does Nex-N2.5-Max need?
The Nex-N2.5 README documents Nex-N2.5-Max running across two nodes of 16 H200 GPUs, with tensor parallelism set to 16, expert parallelism at 16 and FP8 KV cache. Nex AGI ships a prebuilt Docker image, nexagi/sglang:v0.5.18-nex-patch, containing the customised SGLang fork the launch command expects.
How does Nex-N2.5-Max compare with Claude Opus 5?
Nex-N2.5-Max leads Claude Opus 5 on BrowseComp, 92.6 to 90.8, and effectively ties on AutomationBench v1.0.6 at 50.2 against 50.3. Opus 5 stays clearly ahead on coding and knowledge work: 89.1 to 86.1 on Terminal-Bench 2.1, 79.2 to 65.7 on SWE-Bench Pro, and 65.7 to 53.6 on Job Bench.
Is Nex-N2.5-Max free to use commercially?
The Nex-N2.5-Max weights are published on Hugging Face under Apache-2.0, so commercial use needs no separate licence agreement. The cost is compute rather than licensing: unlike the mini and Pro tiers, Max is not offered on OpenRouter's free endpoints, so running it means providing the GPUs yourself.