litigerking/Feyospace-v1-small-NInfer

Feyospace-v1-small — NInfer artifact

This repository contains a converted NInfer artifact for the Feyospace-v1-small model. Feyospace-v1-small is a post-trained Qwen3.8-27B family model for long-context reasoning, cyber-agent research, and controlled security experimentation.

Artifact

• File: feyospace-v1-small-q4.v2.ninfer
• Size: approximately 18.2 GB
• Architecture: Qwen3_5ForConditionalGeneration
• Context length: 262,144 tokens
• Quantization: NInfer groupwise integer conversion with Q4/Q5/Q6 and W8 components; BF16/FP32 tensors are retained where required by the runtime
• MTP tensors: preserved in the artifact
• Conversion device: NVIDIA A40, CUDA

Running with NInfer

git clone https://github.com/tctiger/ninfer-A40.git
cd ninfer-A40

ninfer-serve /models/feyospace-v1-small-q4.v2.ninfer \
  --host 0.0.0.0 \
  --port 8080 \
  --max-context 262144 \
  --kv-dtype bf16 \
  --spec mtp \
  --draft-tokens 3 \
  --lm-head-draft

NInfer provides an OpenAI-compatible HTTP API. Performance depends on the GPU, KV-cache configuration, prompt length, and concurrency.

Intended use and safety

Use this model only for authorized research, defensive security work, education, evaluation, and other lawful purposes. Do not use it to access, modify, or disrupt systems without explicit authorization.

Provenance

The source model is the Feyospace-v1-small checkpoint. The original model card and project information are available from the upstream Feyospace release and the associated paper. This repository contains an inference-format conversion, not the original BF16 Transformers weights.

Limitations and license

This artifact requires an NInfer runtime with Qwen3.8 support and cannot be loaded directly by transformers or llama.cpp. Review the upstream Feyospace and Qwen licenses before redistribution or commercial use.