Hugging Face's logo

nex-agi
/
Nex-N2.5-Pro

Text Generation
Transformers
Safetensors
qwen3_5_moe
image-text-to-text
conversational
Eval Results
compressed-tensors
Model card Files
xet
 Community
3
00index commited on
Commit
11f1e46
·
verified ·
1 Parent(s): 48e569b

Add model card with coming-soon notice (sync from github.com/nex-agi/Nex-N2.5 @ f6b5588)

Browse files
Files changed (5) hide show
  1. .gitattributes +2 -0
  2. README.md +260 -0
  3. figures/NEX_logo.svg +7 -0
  4. figures/Nex-N2-Benchmark-white.png +3 -0
  5. figures/Nex-N2.5-Benchmark-white.png +3 -0
.gitattributes CHANGED
@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ figures/Nex-N2-Benchmark-white.png filter=lfs diff=lfs merge=lfs -text
37
+ figures/Nex-N2.5-Benchmark-white.png filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,260 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ pipeline_tag: text-generation
4
+ library_name: transformers
5
+ ---
6
+ <div align="center">
7
+ <img src="./figures/NEX_logo.svg" width="20%"/>
8
+ </div>
9
+
10
+ ---
11
+
12
+ <div align="center">
13
+ <p>
14
+ 💻 <a href="https://github.com/nex-agi/Nex-N2.5">GitHub</a>&nbsp; · &nbsp;
15
+ 🤗 <a href="https://huggingface.co/nex-agi/Nex-N2.5-Pro">Hugging Face</a>&nbsp; · &nbsp;
16
+ 🌐 <a href="https://nex-agi.com/">Website</a>&nbsp; · &nbsp;
17
+ 🔀 <a href="https://openrouter.ai/nex-agi/nex-n2.5-pro">OpenRouter (Pro)</a>
18
+ </p>
19
+ </div>
20
+
21
+ <div align="center">
22
+ <table><tr><td align="center" style="padding: 12px 24px;">
23
+ 🚧 &nbsp;<b>Nex-N2.5-Pro weights are coming soon.</b><br/>
24
+ </td></tr></table>
25
+ </div>
26
+
27
+ # Nex-N2.5
28
+
29
+ **A next-generation family of agentic models built for long-horizon tasks in real-world environments.**
30
+
31
+ Today, Nex-AGI officially introduces **Nex-N2.5**, its next-generation family of agentic models.
32
+
33
+ Nex-N2.5 is available in three sizes: **mini**, **Pro**, and **Max**. Nex-N2.5-mini and Nex-N2.5-Pro continue to build on the multimodal foundations of Nex-N2, with focused improvements in computer use, web browsing, and visually grounded agentic capabilities. Nex-N2.5-Max is built on a 1.6-trillion-parameter, text-only Mixture-of-Experts (MoE) foundation model, marking our first complete post-training effort at trillion-parameter scale.
34
+
35
+ For long-horizon tasks in real-world environments, Nex-N2.5 further strengthens its ability to act continuously and self-correct through visual feedback. The models can operate computers and browsers, as well as autonomously execute and test programs. Vision is therefore no longer merely an input modality; it has become a critical interface through which an agent perceives its environment, verifies outcomes, and moves a task forward.
36
+
37
+ Building on this foundation, we have further expanded the range of agent training environments, task types, and productivity scenarios, while completing systematic post-training at trillion-parameter scale for the first time. Through broader task coverage and richer environmental feedback, Nex-N2.5 delivers further gains in scientific research, knowledge work, and complex productivity tasks. This work also provides valuable practical experience for training agentic capabilities in even larger models.
38
+
39
+ By jointly advancing model training, infrastructure, and real-world agent scenarios, Nex-AGI aims to continue driving progress in agentic intelligence.
40
+
41
+ ## Open Source
42
+
43
+ Model weights for the Nex-N2.5 family will be released as open source, alongside hosted online services.
44
+
45
+ - **Nex-N2.5-Max:** [Hugging Face](https://huggingface.co/nex-agi/Nex-N2.5-Max) | [ModelScope](https://modelscope.cn/models/nex-agi/Nex-N2.5-Max)
46
+ - **Nex-N2.5-Pro:** [Hugging Face](https://huggingface.co/nex-agi/Nex-N2.5-Pro) | [ModelScope](https://modelscope.cn/models/nex-agi/Nex-N2.5-Pro)
47
+ - **Nex-N2.5-mini:** [Hugging Face](https://huggingface.co/nex-agi/Nex-N2.5-mini) | [ModelScope](https://modelscope.cn/models/nex-agi/Nex-N2.5-mini)
48
+ - **Hosted Access:** [OpenRouter (Nex-N2.5-Pro)](https://openrouter.ai/nex-agi/nex-n2.5-pro)
49
+ - **Websites:** [Global](https://nex-agi.com/) | [China](https://nex-agi.cn/)
50
+
51
+ We welcome developers and enterprises to integrate and try Nex-N2.5 and share their feedback.
52
+
53
+ ## Performance
54
+
55
+ We evaluate Nex-N2.5 across coding, agentic workflows, computer use, and multimodal understanding.
56
+
57
+ ![Nex-N2.5 Benchmark Overview: Text and Multimodal](./figures/Nex-N2.5-Benchmark-white.png)
58
+
59
+ The tables below compare **Nex-N2.5-mini**, **Nex-N2.5-Pro**, and **Nex-N2.5-Max** with leading models across our evaluation suite. **Bold** marks the best result in each benchmark, including ties; — indicates unavailable data.
60
+
61
+ ### Text Benchmarks
62
+
63
+ <!-- Keep benchmark rows first in each tbody; only section labels receive GitHub's alternating row background. -->
64
+ <table>
65
+ <thead>
66
+ <tr>
67
+ <th align="left">Benchmark</th>
68
+ <th align="center">Nex-N2.5-mini</th>
69
+ <th align="center">Nex-N2.5-Pro</th>
70
+ <th align="center">Nex-N2.5-Max</th>
71
+ <th align="center">Claude Opus 5</th>
72
+ <th align="center">GPT-5.6 Sol</th>
73
+ <th align="center">Kimi-K3</th>
74
+ <th align="center">GLM-5.3</th>
75
+ <th align="center">DeepSeek-V4-Pro</th>
76
+ <th align="center">Qwen3.8-Max</th>
77
+ </tr>
78
+ <tr>
79
+ <th colspan="10" align="left">CODING</th>
80
+ </tr>
81
+ </thead>
82
+ <tbody>
83
+ <tr><td>Terminal-Bench 2.1</td><td align="center">73.4</td><td align="center">82.7</td><td align="center">86.1</td><td align="center"><b>89.1</b></td><td align="center">88.8</td><td align="center">88.3</td><td align="center">88.2</td><td align="center">87.9</td><td align="center">86.6</td></tr>
84
+ </tbody>
85
+ <tbody>
86
+ <tr><td>SWE-Bench Pro</td><td align="center">43.8</td><td align="center">61.2</td><td align="center">65.7</td><td align="center"><b>79.2</b></td><td align="center">64.6</td><td align="center">63.3</td><td align="center">64.6</td><td align="center">55.4</td><td align="center">67.7</td></tr>
87
+ </tbody>
88
+ <tbody>
89
+ <tr><td>DeepSWE v1.1</td><td align="center">36.1</td><td align="center">55.8</td><td align="center">65.6</td><td align="center"><b>73.7</b></td><td align="center">72.7</td><td align="center">67.5</td><td align="center">66.9</td><td align="center">62.8</td><td align="center">69.3</td></tr>
90
+ <tr><th colspan="10" align="left">AGENTIC</th></tr>
91
+ </tbody>
92
+ <tbody>
93
+ <tr><td>AutomationBench v1.0.6</td><td align="center">32.3</td><td align="center">44.2</td><td align="center">50.2</td><td align="center">50.3</td><td align="center">45.8</td><td align="center">46.7</td><td align="center">48.2</td><td align="center">43.2</td><td align="center"><b>50.8</b></td></tr>
94
+ </tbody>
95
+ <tbody>
96
+ <tr><td>Toolathlon Verified</td><td align="center">54.6</td><td align="center">68.5</td><td align="center">74.7</td><td align="center"><b>76.5</b></td><td align="center">74.9</td><td align="center"><b>76.5</b></td><td align="center">73.0</td><td align="center">74.1</td><td align="center">73.3</td></tr>
97
+ </tbody>
98
+ <tbody>
99
+ <tr><td>GDPval-AA v2</td><td align="center">1446</td><td align="center">1628</td><td align="center">1713</td><td align="center"><b>1861</b></td><td align="center">1748</td><td align="center">1668</td><td align="center">1769</td><td align="center">1554</td><td align="center">1630</td></tr>
100
+ </tbody>
101
+ <tbody>
102
+ <tr><td>Job Bench</td><td align="center">28.5</td><td align="center">41.4</td><td align="center">53.6</td><td align="center"><b>65.7</b></td><td align="center">45.4</td><td align="center">52.9</td><td align="center">58.2</td><td align="center">54.1</td><td align="center">64.0</td></tr>
103
+ </tbody>
104
+ <tbody>
105
+ <tr><td>BrowseComp</td><td align="center">83.4</td><td align="center">89.7</td><td align="center"><b>92.6</b></td><td align="center">90.8</td><td align="center">90.4</td><td align="center">91.2</td><td align="center">—</td><td align="center">—</td><td align="center">—</td></tr>
106
+ </tbody>
107
+ </table>
108
+
109
+ ### Multimodal Benchmarks
110
+
111
+ <table>
112
+ <thead>
113
+ <tr>
114
+ <th align="left">Benchmark</th>
115
+ <th align="center">Nex-N2.5-mini</th>
116
+ <th align="center">Nex-N2.5-Pro</th>
117
+ <th align="center">MiniMax-M3</th>
118
+ <th align="center">Claude Opus 5</th>
119
+ <th align="center">GPT-5.6 Sol</th>
120
+ <th align="center">Kimi-K3</th>
121
+ <th align="center">GLM-5.3-Flash</th>
122
+ <th align="center">DeepSeek-V4-Flash-Vision</th>
123
+ <th align="center">Qwen3.8-Max</th>
124
+ </tr>
125
+ </thead>
126
+ <tbody>
127
+ <tr><td>OSWorld-Verified</td><td align="center">71.2</td><td align="center">82.2</td><td align="center">75.2</td><td align="center">83.4</td><td align="center">83.2</td><td align="center">84.8</td><td align="center">62.3</td><td align="center">76.7</td><td align="center"><b>86.1</b></td></tr>
128
+ </tbody>
129
+ <tbody>
130
+ <tr><td>OSWorld-2</td><td align="center">30.5</td><td align="center">56.4</td><td align="center">22.3</td><td align="center"><b>68.3</b></td><td align="center">62.7</td><td align="center">58.3</td><td align="center">—</td><td align="center">—</td><td align="center">46.7</td></tr>
131
+ </tbody>
132
+ <tbody>
133
+ <tr><td>WebTest</td><td align="center">48.6</td><td align="center">52.8</td><td align="center">—</td><td align="center">—</td><td align="center"><b>54.0</b></td><td align="center">—</td><td align="center">—</td><td align="center">—</td><td align="center">52.3</td></tr>
134
+ </tbody>
135
+ <tbody>
136
+ <tr><td>WebArena-Verified</td><td align="center">63.4</td><td align="center">67.6</td><td align="center">—</td><td align="center">—</td><td align="center">69.7</td><td align="center"><b>71.6</b></td><td align="center">—</td><td align="center">62.3</td><td align="center">66.8</td></tr>
137
+ </tbody>
138
+ <tbody>
139
+ <tr><td>OSWorld-G</td><td align="center">82.9</td><td align="center"><b>87.4</b></td><td align="center">—</td><td align="center">76.8</td><td align="center">77.7</td><td align="center">79.6</td><td align="center">83.3</td><td align="center">59.4</td><td align="center">84.9</td></tr>
140
+ </tbody>
141
+ <tbody>
142
+ <tr><td>Vision2Web</td><td align="center">52.9</td><td align="center">68.2</td><td align="center">59.0</td><td align="center">—</td><td align="center"><b>79.8</b></td><td align="center">—</td><td align="center">—</td><td align="center">—</td><td align="center">75.1</td></tr>
143
+ </tbody>
144
+ <tbody>
145
+ <tr><td>SWE-MM</td><td align="center">25.5</td><td align="center">38.2</td><td align="center">—</td><td align="center"><b>59.4</b></td><td align="center">40.2</td><td align="center">37.3</td><td align="center">20.6</td><td align="center">39.2</td><td align="center">39.2</td></tr>
146
+ </tbody>
147
+ <tbody>
148
+ <tr><td>OmniDoc</td><td align="center">89.7</td><td align="center">92.2</td><td align="center">91.6</td><td align="center">—</td><td align="center"><b>92.9</b></td><td align="center">91.1</td><td align="center">—</td><td align="center">—</td><td align="center">92.1</td></tr>
149
+ </tbody>
150
+ </table>
151
+
152
+ ## Usage
153
+
154
+ ### Docker Deployment
155
+
156
+ We also provide a prebuilt Docker image with our customized `sglang` fork preinstalled: **`nexagi/sglang:v0.5.18-nex-patch`**. The launch command is the same as above.
157
+
158
+ #### Nex-N2.5-Max
159
+
160
+ ```bash
161
+ # Multi-node (2 nodes, 16 x H200). Run the same command on every node with:
162
+ # <node-rank> = 0 on the head node, 1 on the other node
163
+ # <node0-ip> = IP of the head node (reachable from all others)
164
+ docker run --gpus all --shm-size 32g --network host \
165
+ -v /path/to/your/model:/model \
166
+ nexagi/sglang:v0.5.18-nex-patch \
167
+ python3 -m sglang.launch_server \
168
+ --model-path /path/to/your/model \
169
+ --trust-remote-code \
170
+ --host 0.0.0.0 \
171
+ --port 8000 \
172
+ --nnodes 2 \
173
+ --node-rank "${NODE_RANK}" \
174
+ --dist-init-addr "${MASTER_ADDR}:5000" \
175
+ --tp 16 \
176
+ --pp-size 1 \
177
+ --dp 1 \
178
+ --ep-size 16 \
179
+ --attention-backend dsv4 \
180
+ --kv-cache-dtype fp8_e4m3 \
181
+ --page-size 256 \
182
+ --moe-a2a-backend deepep \
183
+ --moe-runner-backend deep_gemm \
184
+ --moe-dense-tp-size 1 \
185
+ --deepep-mode auto \
186
+ --context-length 262144 \
187
+ --mem-fraction-static 0.84 \
188
+ --chunked-prefill-size 8192 \
189
+ --enable-mixed-chunk \
190
+ --disable-overlap-schedule \
191
+ --max-running-requests 64 \
192
+ --cuda-graph-max-bs-decode 64 \
193
+ --cuda-graph-backend-decode full \
194
+ --cuda-graph-backend-prefill disabled \
195
+ --chat-template /path/to/nex-n2.5-max/chat_template.jinja \
196
+ --reasoning-parser deepseek-r1 \
197
+ --tool-call-parser qwen3_coder
198
+ ```
199
+
200
+ #### Nex-N2.5-Pro
201
+
202
+ Single node with 8 × H100:
203
+
204
+ ```bash
205
+ docker run --gpus all --shm-size 32g --ipc=host \
206
+ -p 30000:30000 \
207
+ -v /path/to/your/model:/model \
208
+ nexagi/sglang:v0.5.18-nex-patch \
209
+ python3 -m sglang.launch_server \
210
+ --model-path /model \
211
+ --tp 8 \
212
+ --host 0.0.0.0 --port 30000 \
213
+ --reasoning-parser qwen3 \
214
+ --tool-call-parser qwen3_coder \
215
+ --chat-template /path/to/nex-N2.5-Pro/chat-template.jinja \
216
+ --mamba-scheduler-strategy extra_buffer
217
+ ```
218
+
219
+ #### Nex-N2.5-mini
220
+
221
+ Single node with 2 × H100:
222
+
223
+ ```bash
224
+ docker run --gpus all --shm-size 32g --ipc=host \
225
+ -p 30000:30000 \
226
+ -v /path/to/your/model:/model \
227
+ nexagi/sglang:v0.5.18-nex-patch \
228
+ python3 -m sglang.launch_server \
229
+ --model-path /model \
230
+ --tp 2 \
231
+ --host 0.0.0.0 --port 30000 \
232
+ --reasoning-parser qwen3 \
233
+ --tool-call-parser qwen3_coder \
234
+ --chat-template /path/to/nex-N2.5-mini/chat-template.jinja \
235
+ --mamba-scheduler-strategy extra_buffer
236
+ ```
237
+
238
+ ### Recommended Sampling Parameters
239
+
240
+ For the best generation quality, we recommend the following sampling parameters:
241
+
242
+ - `temperature`: 0.7
243
+ - `top_p`: 0.95
244
+ - `top_k`: 40
245
+
246
+ ### Function Calling
247
+
248
+ Nex-series models support robust function-calling capabilities. To enable function calling, add the `--tool-call-parser qwen3_coder` flag when launching the server:
249
+
250
+ ```bash
251
+ python -m sglang.launch_server --model-path /path/to/your/model --tool-call-parser qwen3_coder
252
+ ```
253
+
254
+ ### Reasoning Parser
255
+
256
+ Nex-series models emit explicit reasoning traces. Add the `--reasoning-parser qwen3` flag to parse the reasoning content separately from the final response. It can be combined with the function-calling parser above:
257
+
258
+ ```bash
259
+ python -m sglang.launch_server --model-path /path/to/your/model --tool-call-parser qwen3_coder --reasoning-parser qwen3
260
+ ```
figures/NEX_logo.svg ADDED
figures/Nex-N2-Benchmark-white.png ADDED

Git LFS Details

  • SHA256: b17f8413b470576323264467193956f297576e0e2720c7f0687e9bdeb8acb84a
  • Pointer size: 131 Bytes
  • Size of remote file: 487 kB
figures/Nex-N2.5-Benchmark-white.png ADDED

Git LFS Details

  • SHA256: c042b1239bf27f167ba34f25438cb0f081b539de1931b2dcf6ee1a8e4779b0b8
  • Pointer size: 131 Bytes
  • Size of remote file: 895 kB