Abliteration orthogonalizes the model weights against the measured "refusal" direction, reducing refusals while keeping the base model's capabilities as intact as possible (low KL divergence). It does not add any new knowledge or capability — all credit for the underlying model belongs to Nex-AGI. The original model card is reproduced in full below.
This model has had its built-in refusal behavior deliberately reduced. As a result it may produce unexpected, offensive, inaccurate, or otherwise harmful output, and may comply with requests that the original model would have refused.
It is provided by the publisher, edougawa, "as is" and without warranty of any kind, express or implied. Use at your own risk.
You are solely responsible for how you use this model and for ensuring your use — and any generated output — complies with all applicable laws, regulations, and the terms of the base model's license.
To the maximum extent permitted by law, the publisher (edougawa), the base-model authors (Nex-AGI), and the Abliterix authors accept no liability for any claim, damages, or other consequences arising from the use of this model or its outputs.
Outputs do not reflect the views of the publisher (edougawa), the base-model authors (Nex-AGI), or the Abliterix authors. Apply your own safety filtering, human review, and guardrails before any production or user-facing use.
Steering parameters
Parameter
Value
vector_index
29.50
attn.k_proj.max_weight
3.43
attn.k_proj.max_weight_position
28.18
attn.k_proj.min_weight
1.88
attn.k_proj.min_weight_distance
11.27
attn.o_proj.max_weight
2.41
attn.o_proj.max_weight_position
37.58
attn.o_proj.min_weight
0.30
attn.o_proj.min_weight_distance
20.37
attn.q_proj.max_weight
2.29
attn.q_proj.max_weight_position
35.88
attn.q_proj.min_weight
0.52
attn.q_proj.min_weight_distance
9.47
attn.v_proj.max_weight
0.59
attn.v_proj.max_weight_position
29.14
attn.v_proj.min_weight
0.22
attn.v_proj.min_weight_distance
10.12
mlp.down_proj.max_weight
3.84
mlp.down_proj.max_weight_position
30.09
mlp.down_proj.min_weight
2.22
mlp.down_proj.min_weight_distance
13.71
Performance
Metric
This model (Nex-N2-mini-Abliterated)
Base model (Nex-N2-mini)
KL divergence
0.0180
0 (by definition)
Refusals
17/200
197/200
Deployment, sampling, function-calling, and reasoning-parser instructions from the base model (see the original card below) apply unchanged to this abliterated checkpoint.
Quickstart
Download this model (edougawa/Nex-N2-mini-Abliterated) and serve it with the Nex-AGI sglang fork:
bash
1# 1. Download edougawa/Nex-N2-mini-Abliterated2hf download edougawa/Nex-N2-mini-Abliterated --local-dir ./Nex-N2-mini-Abliterated
34# 2. Serve it (single node, 2× H100 — see the original card for multi-node / Docker)5python -m sglang.launch_server \6 --model-path ./Nex-N2-mini-Abliterated \7 --tp 2\8 --reasoning-parser qwen3 \9 --tool-call-parser qwen3_coder \10 --mamba-scheduler-strategy extra_buffer
python
1# Or load it directly with transformers2from transformers import AutoModelForCausalLM, AutoTokenizer
34model_id ="edougawa/Nex-N2-mini-Abliterated"5tokenizer = AutoTokenizer.from_pretrained(model_id)6model = AutoModelForCausalLM.from_pretrained(model_id, dtype="bfloat16", device_map="auto")
Original model card
The following is the original model card for the base model, nex-agi/Nex-N2-mini, reproduced here in full with credit to Nex-AGI. The serving instructions below apply to this abliterated checkpoint as well.
Today, we are officially releasing and open-sourcing our next-generation model, Nex-N2 — an agent model built for real-world productivity scenarios. With first-tier coding and agentic capabilities, Nex-N2 keeps driving complex, long-horizon tasks forward in real environments to deliver stable, end-to-end results.
Over the past year, a paradigm shift led by Vibe Coding and Harness Engineering has been redefining the limits of LLM agents. From dialogue, to reasoning, to agents that execute long-horizon tasks with environmental feedback, the tasks models must handle keep growing harder, the contexts longer, and the environments more realistic. The core of next-generation model competition is no longer whether a model can think, but whether it can reliably and efficiently turn thinking into actions that are executable, verifiable, and iterable.
Rather than treating reasoning, tool use, and environment execution as separate capabilities, Nex-N2 unifies them through an Agentic Thinking framework that connects requirement understanding, task planning, code implementation, environmental feedback, evaluation and debugging, and continuous iteration into a single closed loop. The framework has two parts:
Adaptive Thinking lets the model decide on its own when to think and how deeply — executing simple actions quickly while reasoning thoroughly on critical decisions.
Coherent Thinking carries one consistent reasoning paradigm across general reasoning and diverse agentic tasks, staying consistent across tasks and modalities to enable stable capability transfer.
Across real agentic workflows — agentic coding, deep research, tool calling, and terminal execution — Nex-N2 reaches first-tier performance, with substantial gains over the previous-generation Nex-N1 on multiple authoritative benchmarks. In real productivity scenarios such as OpenClaw one-person-company workflows, end-to-end game development, and web and multimodal generation, it likewise demonstrates outstanding usability, robustness, and stability.
Open Source
In keeping with our commitment to open source, we are releasing both Nex-N2-Pro and Nex-N2-mini as open-source models starting today.
We welcome developers and enterprises to integrate and try Nex-N2 and share their feedback.
Performance
We evaluate Nex-N2 in real agentic workflows along three directions — agentic tasks, coding tasks, and general tasks — covering benchmarks across tool calling, search-based decision-making, software engineering, and terminal execution. Nex-N2-Pro delivers strong performance that keeps pace with top-tier models such as GPT-5.5 and Opus 4.7: it excels at coding (e.g., 75.3 on Terminal-Bench 2.1) and long-horizon tasks (1585 on GDPval), and shows especially strong generalization and competitiveness on newer benchmarks like SWE-Atlas and DeepSWE. On general capability and core reasoning, it stands on par with leading frontier models.
Nex-N2 Benchmark Overview
Nex-N2 ships in two variants, both post-trained on the Qwen3.5 series: Nex-N2-Pro (built on Qwen3.5-397B-A17B) and Nex-N2-mini (built on Qwen3.5-35B-A3B-Base), covering different latency and quality trade-offs. The table below reports their scores alongside leading proprietary and open models across our full evaluation suite.
Benchmark
Nex-N2-mini
Nex-N2-Pro
GPT-5.5
Opus 4.7
Kimi-K2.6
GLM-5.1
MiniMax M3
DeepSeek-V4-Pro
Agent
BrowseComp
74.1
83.7
84.4
79.8
83.2
79.3
83.5
83.4
GDPval
1402
1585
1769
1753
1481
1535
-
1554
Toolathlon
33.3
51.9
55.6
52.8
50.0
40.7
-
51.8
WildClawBench
47.7
53.5
58.2
62.2
-
48.2
-
43.7
WideSearch
62.0
75.6
-
-
80.8
-
-
-
TAU3
65.9
71.1
-
-
-
70.6
-
-
Coding & SWE
SWE-Bench Pro
50.2
58.8
58.6
64.3
58.6
58.4
59.0
55.4
Terminal-Bench 2.1
60.7
75.3
83.4
69.7
-
58.7
66.0
72.0
DeepSWE
8.0
33.6
70
54
24
18
-
8
SWE-Bench Verified
74.4
80.8
82.9
87.6
80.2
-
80.5
80.6
SWE Atlas QnA
31.5
37.9
45.4
45.2
-
-
37.9
-
SWE Atlas RF
30.0
32.9
44.8
48.6
-
-
-
-
SWE Atlas TW
23.3
40.0
42.6
38.2
-
-
30.8
-
General & Reasoning
GPQA Diamond
82.6
90.7
93.6
94.2
90.5
86.2
-
90.1
IFEval
89.1
94.0
-
-
94.5
94.5
-
91.9
Apex
9.4
36.5
-
-
24.0
11.5
-
38.3
Usage
Local Deployment
Note: For the best performance with Nex-series models, we recommend serving them with our customized sglang fork.
First, install our sglang fork:
bash
1# Use the customized `sglang` fork2git clone https://github.com/nex-agi/sglang.git
3cd sglang
45# Install the python packages6pip install --upgrade pip
7pip install -e "python"
Nex-N2-Pro
Launch the server (example on two 8× H100 servers with CUDA 13.0):
bash
1# Multi-node (2 nodes). Run the same command on every node with:2# <node-rank> = 0 on the head node, 1 on the other node3# <node0-ip> = IP of the head node (reachable from all others)4python -m sglang.launch_server \5 --model-path /path/to/your/model \6 --tp 16\7 --nnodes 2\8 --node-rank <node-rank>\9 --dist-init-addr <node0-ip>:20000 \10 --reasoning-parser qwen3 \11 --tool-call-parser qwen3_coder \12 --mamba-scheduler-strategy extra_buffer
Nex-N2-mini
Launch the server (example on one 2× H100 server with CUDA 13.0):
We also provide a prebuilt Docker image with our customized sglang fork preinstalled: nexagi/sglang:v0.5.12. The launch command is the same as above.
Nex-N2-Pro
bash
1# Multi-node (2 nodes). Run the same command on every node with:2# <node-rank> = 0 on the head node, 1 on the other node3# <node0-ip> = IP of the head node (reachable from all others)4docker run --gpus all --shm-size 32g --network host\5 -v /path/to/your/model:/model \6 nexagi/sglang:v0.5.12 \7 python3 -m sglang.launch_server \8 --model-path /model \9 --tp 16\10 --nnodes 2\11 --node-rank <node-rank>\12 --dist-init-addr <node0-ip>:20000 \13 --host 0.0.0.0 --port 30000\14 --reasoning-parser qwen3 \15 --tool-call-parser qwen3_coder \16 --mamba-scheduler-strategy extra_buffer
For the best generation quality, we recommend the following sampling parameters:
temperature: 0.7
top_p: 0.95
top_k: 40
Function Calling
Nex-series models support robust function-calling capabilities. To enable function calling, add the --tool-call-parser qwen3_coder flag when launching the server:
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:
