ZAYA1-8B-NVFP4

ZAYA1-8B

ZAYA1-8B is a small mixture of experts language model with 760M active parameters and 8.4B total parameters trained end-to-end by Zyphra. ZAYA1-8B sets a new standard of intelligence efficiency for its parameter count through a combination of novel architecture and innovations in pretraining and post-training.

ZAYA1-8B excels at detailed long-form reasoning especially for mathematical and coding task. It punches heavily above its weight in these regimes and due to its inference efficiency and small size can be highly effective in test-time compute harnesses.

Due to its small total parameter count, ZAYA1-8B can also be deployed on-device for local LLM applications.

Learn more in our technical report and blog.

This is the post-trained reasoning version of ZAYA1-8B. The pretraining base can be found here.

Performance

ZAYA1-8B performs extremely strongly, especially in challenging mathematical, reasoning, and coding benchmarks. ZAYA1-8B is competitive with models several times its own size including frontier-scale reasoning models at mathematical benchmarks.

First we compare ZAYA1-8B to the SOTA Qwen3 and Qwen3.5 model series of approximately the same parameter count as well as the recently released Gemma4 models and secondly to a variety of larger open-weights models.

In-class comparison against open-source reasoning models

Scaling comparison against larger open-source reasoning models

All numbers are run on the Zyphra evaluation harness. Models are ordered by total parameter count.

Quickstart

Prerequisites

We recommend installing the following libraries in a fresh python environment (tested with python 3.12).

To use ZAYA1-8B, install zaya1-pr branch from our fork of vllm library (the command will trigger a full build of vLLM from source):

pip install "vllm @ git+https://github.com/Zyphra/vllm.git@zaya1-pr"

If you want to run in transformers, install zaya1 branch from our fork of transformers library as well:

pip install "transformers @ git+https://github.com/Zyphra/transformers.git@zaya1"

Deployment

To start vLLM server, run the following command:

vllm serve Zyphra/ZAYA1-8B --port 8010 \
   --mamba-cache-dtype float32 --dtype bfloat16 \
   --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser zaya_xml

For parallel deployment we recommend using DP with EP as TP for CCA is not supported in the branch above. If running on 8 GPUs, set extra flags -dp 8 -ep to run with DP=EP=8.

For our evaluations and for general use, we recommend temperature 1.0, top-p 0.95, top-k -1. For agent and code use cases, we recommend temperature 0.6, top-p 0.95, top-k -1.

Once the server is up, you can query a model with curl like in the following example:

curl http://localhost:8010/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "Zyphra/ZAYA1-8B",
        "messages": [
            {"role": "system", "content": "You are a helpful assistant."},
            {"role": "user", "content": "Hello. How is it going?"}
        ]
    }'

NVFP4 quantization

This repository is a Blackwell-oriented NVFP4 quantization of Zyphra/ZAYA1-8B.

• Source revision: 970cfc9f5e7e5a4f5f6f0645955928a9b6a98415
• Quantizer: NVIDIA TensorRT Model Optimizer / modelopt.torch.quantization.NVFP4_DEFAULT_CFG
• Export format: ModelOpt unified Hugging Face checkpoint for vLLM modelopt_fp4
• Calibration: 128 representative chat/code/math/multilingual samples, truncated to 2048 tokens
• High-precision exclusions: lm_head, token embeddings, router layers, and CCA convolution layers
• Quantized on: NVIDIA RTX PRO 6000 Blackwell Workstation Edition / compute capability (12, 0)

Recommended vLLM command:

vllm serve switzerchees/ZAYA1-8B-NVFP4 \
  --trust-remote-code \
  --quantization modelopt_fp4 \
  --dtype auto \
  --max-model-len 8192 \
  --no-enable-prefix-caching

ZAYA1 currently needs Zyphra's vLLM branch or another vLLM build that contains ZayaForCausalLM support. For this NVFP4 export, the tested runtime also needs loader handling for ModelOpt expert scale tensors and unquantized ZAYA router layers. The patch used for testing is included in this repository as runtime/zaya-vllm-modelopt-nvfp4.patch.

On first Blackwell startup, FlashInfer compiles SM120 NVFP4/MoE kernels. The tested environment used CUDA 13.0 nvcc, nvrtc, and cccl Python packages and MAX_JOBS=2 to avoid RAM pressure during the one-time JIT build.

Smoke test:

• GPU: NVIDIA RTX PRO 6000 Blackwell Workstation Edition
• Runtime: Zyphra vLLM zaya1-pr commit 6c3178819e623ab915dd17e0b965269c3b745abd
• vLLM args: --quantization modelopt_fp4, max_model_len=1024, max_num_seqs=1, gpu_memory_utilization=0.30, eager mode
• Model weight load reported by vLLM: 5.42 GiB
• Prompt check with the model chat template produced plausible answers, including 17 * 23 = 391.

Docker runtime image:

docker run --gpus all --platform linux/amd64 -p 8000:8000 \
  switzerchees/vllm:zaya1-8b \
  --model switzerchees/ZAYA1-8B-NVFP4 \
  --trust-remote-code \
  --quantization modelopt_fp4 \
  --dtype auto \
  --max-model-len 8192 \
  --no-enable-prefix-caching

The image is published for linux/amd64 as switzerchees/vllm:zaya1-8b.