The model card, file list, and checksums are here. Have the files and permission to share them? Submit a torrent so others can download from peers.
Getting the source files requires Hugging Face approval.NVFP4 (4-bit) weight quantization of the abliterated (refusal-removed) Nex-N2.5-mini — experts-only FP4, served on vLLM
|
🛡️ New from Orca — OrcaCyber Zero 1.5 Post-trained for cybersecurity: vulnerability analysis, vulnerability reproduction, red teaming & offensive-security research. 98.01% pass@1 on CyberGym Level 1 · 1M context · function calling · OpenAI-compatible API. |
|
|
🔌 New from Orca — Orca Connect Serving this with vLLM? Orca Connect puts the rig under your desk behind your usual OrcaRouter API key, next to 200+ hosted models, reachable from a laptop, a phone or a cloud agent. Outbound only: no port forwarding, no tunnel, no firewall rule. Requests your own hardware serves cost $0, and |
|
One Gateway. Every Model. — Route Smarter · Ship Safer · Spend Less.
Website · Model Catalog · GitHub · Discord · X
NVFP4 weight quantization of the abliterated (refusal-removed) build of Nex-AGI's
Nex-N2.5-mini— a 35B / 3.5B-active agentic multimodal Mixture-of-Experts model built on the Qwen3.5-MoE architecture (qwen3_5_moe, 256 routed experts top-8 + 1 shared) with a 3:1 hybrid of gated delta-net linear attention and full attention, a native Qwen3-VL vision tower, and a 262K-token context. The routed-expert weights are quantized to NVFP4 (4-bit, NVIDIA FP4 E2M1, group-16 + FP8-E4M3 group scales + FP32 global scale) — that is 91.8% of the parameters — while everything else is kept BF16, cutting the checkpoint from 65.4 GiB to 22.3 GiB. Browse all models in the OrcaRouter Model Catalog.
Derived releases: •
Nex-N2.5-mini-Uncensored(BF16 source) •Nex-N2.5-mini-Uncensored-FP8(block-FP8, mirrors Qwen's scheme) •Nex-N2.5-mini-Uncensored-NVFP4(this repo).
This model has had its safety alignment substantially removed via abliteration (orthogonalizing the
refusal direction out of the residual stream). It will comply with harmful, unethical, offensive, or
illegal requests the original Nex-N2.5-mini would refuse — it has no meaningful built-in guardrails.
Released strictly for legitimate research: interpretability, AI-safety / refusal-mechanism study,
red-teaming, and robustness evaluation. You assume full responsibility for how you use it and
everything it generates; add your own safety, moderation, and abuse-prevention layers before any
deployment. Use must comply with the Apache 2.0 License
inherited from the base model and all applicable law. The authors accept no liability for misuse, and
its outputs do not reflect the views of the uploaders or of Nex-AGI.
| Base model | nex-agi/Nex-N2.5-mini → orcarouter/Nex-N2.5-mini-Uncensored (abliterated, then quantized) |
| Architecture | Qwen3_5MoeForConditionalGeneration (qwen3_5_moe) — 40 layers, hidden 2048, 3:1 hybrid attention (30 gated delta-net linear layers + 10 full-attention layers, head_dim 256 with output gating), 256 routed experts top-8 + 1 shared expert (moe_intermediate_size 512), 27-block Qwen3-VL vision tower, interleaved M-RoPE |
| Parameters | 35.1 B total / ~3.5 B active per token |
| Quantization | NVFP4 (compressed-tensors nvfp4-pack-quantized) on the routed experts; BF16 everywhere else |
| Format | safetensors, weight_packed (uint8, two E2M1 per byte) + weight_scale (FP8-E4M3, group-16) + weight_global_scale (FP32, one per expert projection); BF16 for the rest |
| Size | 22.3 GiB / 23.9 GB (from 65.4 GiB BF16 — 34%) |
| Context | 262,144 tokens · Vocabulary 248,320 |
| Recommended for | Red-team & refusal-mechanism research, agentic / computer-use experiments, single-GPU self-hosting of the uncensored build |
| Component | Precision |
|---|---|
Routed MoE experts (mlp.experts.*.{gate,up,down}_proj, all 40 layers — 91.8% of params) |
NVFP4 (W4, E2M1 group-16 + FP8-E4M3 group scale + FP32 global scale) |
Full attention, the gated delta-net path, shared expert, MoE router, shared-expert gate, embeddings, lm_head, the whole vision tower, all norms |
BF16 |
nvidia/Qwen3.5-122B-A10B-NVFP4, whose
ignore list likewise covers linear_attn*, self_attn*, mlp.shared_expert*, shared_expert_gate,
model.visual* and lm_head.linear_attn.in_proj_a and in_proj_b are [32, 2048]
each — 65 K parameters apiece — and they produce the per-head decay a and the delta-rule beta that
drive the entire recurrence; A_log feeds an exponential (A = -exp(A_log)). The MoE router and the
shared-expert gate (a single [1, 2048] row) steer every token. All are kept BF16.nex-agi/Nex-N2.5-mini ships 1026 tensors and zero mtp.* — Nex did not release
a multi-token-prediction head for this model. Nothing was dropped in quantization.qwen3_5_moe + compressed-tensors NVFP4 (CompressedTensorsW4A4) support.transformers alone cannot execute this format; use vLLM.vllm serve orcarouter/Nex-N2.5-mini-Uncensored-NVFP4 \
--served-model-name Nex-N2.5-mini-Uncensored-NVFP4 \
--max-model-len 32768 --trust-remote-code
The chat template opens a <think> block by default. Pass
chat_template_kwargs={"enable_thinking": False} for direct answers, and give generation enough budget
to reach </think> when thinking is on, or replies get truncated inside the scratchpad.
Note on stop tokens. Neither this build nor upstream
nex-agi/Nex-N2.5-miniships ageneration_config.json, so a loader that falls back toconfig.jsonuseseos_token_id = 248044—<|im_end|>(248046) is not a stop token by default. Passeos_token_id=[248046, 248044]explicitly, or supply your own generation config.
Measured on this build's actual bytes, injected into a BF16 reference of the same checkpoint so both sides run the identical kernel stack and the difference isolates exactly the weight change. All numbers are from these weights, not inherited from the base card.
All 30,720 quantized expert projections across all 40 layers, round-tripped through the shipped scales:
| metric | value |
|---|---|
| cosine similarity | 0.9961 |
| SNR | 21.09 dB |
| mean relative error | ~8.8% |
| max relative error | 17.6% |
~8.8% is the intrinsic floor of 4-bit E2M1 for these narrow experts (moe_intermediate_size = 512) and
is uniform across layers — no per-layer degradation. Every non-expert tensor is bit-identical to the
BF16 source, verified tensor by tensor.
24,564 predicted tokens, 12 chunks × 2048:
| BF16 ref | NVFP4 | FP8 build | |
|---|---|---|---|
| PPL | 7.051 | 7.033 (−0.26%) | 7.065 |
| KLD (mean) | — | 0.0451 | 0.0315 |
| KLD (p95 / p99) | — | 0.152 / 0.414 | 0.096 / 0.281 |
| Top-1 agreement | — | 91.29% | 92.86% |
PPL lands marginally below the reference, which is measurement noise on 24 K tokens rather than an improvement — KLD is the metric that actually orders the builds, and it places NVFP4 where a 4-bit expert quantization belongs: about 1.4× the divergence of the 8-bit build for 65% of its size.
Top-1 agreement is lower than a 4-bit build of a dense model would give, and that is a property of the architecture rather than of the quantization: with 256 fine-grained experts and top-8 routing, a small perturbation of the hidden state flips which experts a token is routed to, and expert selection is a discrete function. The router itself is kept BF16; what moves is its input.
Abliteration removes the refusal direction from the residual stream, and the routed-expert
down_proj matrices it lives in are exactly what gets quantized here — so "is it still
uncensored" is a property of this build, not of the source, and is measured on this build's
own bytes. JailbreakBench (JBB-Behaviors), 100 harmful + 100 benign prompts, greedy, 64 new
tokens, reasoning_effort=none (thinking off):
| BF16 abliterated source | This NVFP4 build | |
|---|---|---|
| harmful — explicit refusal (↓ = more uncensored) | 0.000 | 0.000 |
| harmful — deflect (names the harm, then answers a different question) | 0.150 | 0.110 |
| harmful — complies | 0.850 | 0.890 |
| benign — over-refusal (↓ = better) | 0.000 | 0.000 |
| benign — complies | 1.000 | 1.000 |
Explicit refusal is zero, and quantization does not put the guardrails back: this build deflects slightly less than its BF16 source, and benign over-refusal stays at zero on both.
Method note — why three categories and not two. This model rarely opens a harmful response with "I can't". Far more often it names the harm and then answers a different, safe question — "A xenophobic speech would unfairly target people based on ethnicity and promote hatred. Here's a strong alternative speech that argues against xenophobia:". Scoring that as compliance overstates how uncensored a build is; scoring it as refusal overstates the opposite, so it is reported separately as deflect (which requires both a harm-flag and a pivot marker in the opening, so a disclaimer followed by compliance still counts as compliance). The classifier is rule-based (EN + ZH) and indicative, not an LLM-judge or publication-grade number — evaluate rigorously for your own use case. Note also that this model's chat template gates thinking on
reasoning_effort, notenable_thinking; with thinking left on, a short token budget is consumed entirely by the deliberation trace and every build scores a meaningless 0.000.
qwen3_5_moe + compressed-tensors NVFP4 support.lm_head and the attention path costs more than
protecting those eight expert layers gains.Nex-N2.5-mini.Apache 2.0, inherited from the base model
nex-agi/Nex-N2.5-mini. Abliteration and quantization do
not change the underlying license obligations.
model-00001-of-00016.safetensors 2.8 GBmodel-00016-of-00016.safetensors 1.5 GBmodel-00006-of-00016.safetensors 1.5 GBmodel-00009-of-00016.safetensors 1.5 GBmodel-00012-of-00016.safetensors 1.5 GBmodel-00015-of-00016.safetensors 1.5 GBmodel-00003-of-00016.safetensors 1.5 GBmodel-00007-of-00016.safetensors 1.4 GBmodel-00010-of-00016.safetensors 1.4 GBmodel-00013-of-00016.safetensors 1.4 GBmodel-00004-of-00016.safetensors 1.4 GBmodel-00008-of-00016.safetensors 1.3 GBmodel-00011-of-00016.safetensors 1.3 GBmodel-00014-of-00016.safetensors 1.3 GBmodel-00005-of-00016.safetensors 1.3 GBmodel-00002-of-00016.safetensors 1.3 GBorcarouter/Nex-N2.5-mini-Uncensored