sojufx/Nex-N2.5-mini-NVFP4

Nex-N2.5-mini NVFP4

A native NVIDIA ModelOpt NVFP4 conversion of nex-agi/Nex-N2.5-mini, prepared for NVIDIA Blackwell GPUs.

This is a fresh conversion from the original BF16 checkpoint, not a requantization of another quantized release. The precision layout keeps the sensitive and multimodal parts of the model at higher precision while reducing the large routed-expert weights:

• NVFP4 W4A16: routed MoE experts and shared-expert MLP projections
• FP8: full-attention projections and hybrid linear-attention projections
• FP8: KV-cache cast metadata
• BF16: lm_head, complete vision encoder/projector, and native one-layer MTP head

Keeping the LM head in BF16 is intentional. It directly produces next-token logits, so retaining it is a quality-first choice for instruction following, structured output, and tool use. The vision stack and MTP head are also preserved rather than being folded into the quantization pass.

Included

• NVIDIA ModelOpt mixed-precision safetensors checkpoint
• Original tokenizer, chat template, processor, and generation configuration
• Original model configuration and multimodal preprocessing settings
• hf_quant_config.json describing the precision map

Serving With vLLM

Use a current Blackwell-capable vLLM build with support for the Qwen 3.5 MoE multimodal architecture and ModelOpt mixed-precision checkpoints. The exported configuration allows vLLM to detect the ModelOpt format automatically.

pip install -U vllm

vllm serve sojufx/Nex-N2.5-mini-NVFP4 \
  --trust-remote-code \
  --max-model-len 262144 \
  --gpu-memory-utilization 0.80 \
  --kv-cache-dtype fp8 \
  --limit-mm-per-prompt '{"image": 1}'

Start with a shorter context window or lower GPU-memory utilization on smaller cards. This model is multimodal; retain the bundled chat_template.jinja and processor files so image tokens are rendered correctly.

For an OpenAI-compatible request with an image:

curl http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "sojufx/Nex-N2.5-mini-NVFP4",
    "messages": [{
      "role": "user",
      "content": [
        {"type": "text", "text": "Describe this image in one sentence."},
        {"type": "image_url", "image_url": {"url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}}
      ]
    }]
  }'

Conversion Details

• Source: nex-agi/Nex-N2.5-mini
• Source revision: 87420286149d9cce9bd46cd335ef9bda33c37c1b
• Quantizer: NVIDIA ModelOpt
• Recipe: huggingface/qwen3_5_moe/ptq/qwopus_w4a16_nvfp4_bf16_lm_head
• Calibration: 256 public WikiText samples
• Format: safetensors

The conversion completed with coherent pre- and post-quantization generation smoke tests. This repository contains converted weights; please follow the base model's Apache-2.0 license and usage terms.

Provenance

Nex-N2.5-mini was created and released by Nex-AGI. This repository repackages no training data and makes no model changes beyond the documented precision conversion.