What is Nex-N2-mini?
Nex-N2-mini is a 35.1B-parameter open-weight agent model from Nex-AGI and the smaller of the two models in the Nex-N2 release. It is post-trained on Qwen3.5-35B-A3B-Base and built for real productivity work: agentic coding, deep research, tool calling and terminal execution. Nex-AGI published the weights on Hugging Face on June 4, 2026 under Apache 2.0, alongside the larger Nex-N2-Pro, which is post-trained on Qwen3.5-397B-A17B.
| Specification | Nex-N2-mini |
|---|---|
| Total parameters | 35.1B |
| Base model | Qwen3.5-35B-A3B-Base, post-trained |
| Context window | Not stated on the Nex-AGI model card |
| Modalities | Not stated by Nex-AGI; the GGUF repo ships mmproj projector files next to the weights |
| Larger sibling | Nex-N2-Pro, post-trained on Qwen3.5-397B-A17B |
| Training focus | Agentic Thinking: agentic coding, deep research, tool calling, terminal execution |
| Reasoning | Explicit reasoning traces, parsed with the qwen3 reasoning parser |
| Function calling | Supported, qwen3_coder tool-call parser |
| Recommended sampling | temperature 0.7, top_p 0.95, top_k 40 |
| Vendor serving stack | Nex-AGI SGLang fork; the launch example for the mini is one server with two H100s |
| Release date | June 4, 2026 |
| License | Apache 2.0 |
The idea behind the release is what Nex-AGI calls Agentic Thinking, a framework that connects requirement understanding, task planning, code implementation, environmental feedback, evaluation and debugging, and continuous iteration into a single closed loop. Adaptive Thinking lets the model decide when to think and how deeply, so it executes simple actions quickly and reasons thoroughly on critical decisions. Coherent Thinking keeps one reasoning paradigm across general reasoning and agentic tasks. In practice the model emits explicit reasoning traces and calls functions, and Nex-AGI's reference stack parses both with the qwen3 reasoning parser and the qwen3_coder tool-call parser in its customized SGLang fork.
Nex-N2-mini benchmarks
Nex-AGI published the launch numbers on the model card, scoring both N2 variants against GPT-5.5, Opus 4.7, Kimi-K2.6, GLM-5.1, MiniMax M3 and DeepSeek-V4-Pro. The rows below put the mini next to its Pro sibling and to GPT-5.5 and Opus 4.7.
| Benchmark | Nex-N2-mini | Nex-N2-Pro | GPT-5.5 | Opus 4.7 |
|---|---|---|---|---|
BrowseComp Web browsing | 74.1 | 83.7 | 84.4 | 79.8 |
Toolathlon Long-horizon tools | 33.3 | 51.9 | 55.6 | 52.8 |
SWE-Bench Verified Software engineering | 74.4 | 80.8 | 82.9 | 87.6 |
SWE-Bench Pro Harder engineering | 50.2 | 58.8 | 58.6 | 64.3 |
Terminal-Bench 2.1 Terminal agents | 60.7 | 75.3 | 83.4 | 69.7 |
DeepSWE Agentic coding | 8.0 | 33.6 | 70 | 54 |
GPQA Diamond Expert science | 82.6 | 90.7 | 93.6 | 94.2 |
The mini takes no row here, and the vendor table does not claim it does: these are frontier-scale comparisons and the Pro variant is the one that keeps pace. The widest gap in the rows above is DeepSWE, where the mini scores 8.0 against 70 for GPT-5.5 and 54 for Opus 4.7. On SWE-Bench Verified it scores 74.4 against 82.9 for GPT-5.5 and 87.6 for Opus 4.7, and on GPQA Diamond 82.6 against 93.6 and 94.2. Outside the rows above, the vendor table gives the mini 1402 on GDPval, 47.7 on WildClawBench, 31.5 on SWE Atlas QnA, 30.0 on SWE Atlas RF, 23.3 on SWE Atlas TW and 9.4 on Apex. On rows the vendor left blank for the frontier models, the mini still posts numbers: 62.0 on WideSearch, 65.9 on TAU3 and 89.1 on IFEval.
Nex-N2-mini hardware requirements
The system requirement to check is memory. Nex-AGI ships the original safetensors weights and recommends its own SGLang fork, with a launch example for the mini on one server with two H100s, so for a local machine the practical route is a GGUF build. The ladder below uses the real file sizes from the community repo bartowski/nex-agi_Nex-N2-mini-GGUF on Hugging Face, whose smallest weight file is IQ2_XXS at 9.78 GB.
| Memory | Build to pick | File size |
|---|---|---|
| 12 GB | IQ2_XXS | 9.78 GB |
| 16 GB | Q2_K_L | 13.11 GB |
| 24 GB | IQ4_XS | 18.81 GB |
| 32 GB | Q4_K_M | 21.39 GB |
| 48 GB | Q6_K | 30.05 GB |
| 64 GB and up | Q8_0 | 36.91 GB |
When two builds both fit, take the larger one: most neighbouring builds in this repo are less than a gigabyte apart, and a few sit almost on top of each other, such as IQ3_XS at 16.22 GB and Q3_K_M at 16.23 GB. The low end runs IQ2_XXS at 9.78 GB, IQ2_XS at 10.80 GB and IQ2_S at 11.01 GB, and above that the repo lists IQ3_XXS at 14.87 GB, IQ4_NL at 19.86 GB and Q5_K_M at 25.02 GB. It also carries two mmproj projector files of 0.90 GB each, in bf16 and f16, separate from the weight builds. If the format is new to you, start with what GGUF is.
How to run Nex-N2-mini in Atomic Chat
Atomic Chat is a free local app for macOS, Windows and Linux. It includes a Hugging Face model browser and a built-in chat, with no manual llama.cpp build required.
- Download Atomic Chat for your platform and open it.
- Search for Nex-N2-mini in the model browser and open Download Options.
- Pick the build that fits the memory you have, then start a chat.
For the bigger variant, see Nex-N2-Pro, or browse every Nex model you can run locally.
Nex-N2-mini license
Nex-N2-mini is released under Apache 2.0, the license Nex-AGI set on the model repository. That permits commercial use, modification and redistribution with no royalties, so you can fine-tune the model, ship it inside your own products, and run it on your own hardware without a usage fee.