Furiosa-LLM 2026.5.0 beta3 has been released!

Furiosa LLM
1

Hi everyone! 2026.5.0 Beta3 is now available on the public Python package repository.

This public beta release 2026.5.0 Beta3 includes the following key improvements since 2026.5.0 Beta0:

Performance and functionality

  • Improved the performance of Qwen3.5/3.6/3.8 (dense/MoE) family models.
  • Added video input support for Qwen3 VL and Qwen 3.5/3.6/3.8 models.
  • Fixed a bug in DP support for 4PE models (e.g., Qwen 3.5 0.8B).
  • Added or improved support for the following sampling parameters: parallel_tool_calls, reasoning_effort, and repetition_detection.
  • Improved scheduling for multimodal and linear attention models.
  • Added speculative decoding support for multimodal models.
  • Enabled NPU sampling by default.
  • Added support for Qwen3.5 0.8B.

Bug fixes and security updates

  • Frontend bug fixes and improvements for Solar mini4 and Qwen3, including reasoning and tool calling.
  • Improved handling of Unicode U+FFFD.
  • Fixed the eMudhra CA security issue.

Installation

uv pip install \
  --prerelease=allow \
  --torch-backend auto \
  furiosa-llm==2026.5.0b3

You can test the video modality as follows:

apt-get install ffmpeg

uv pip install \
	--prerelease=allow \
	--torch-backend auto \
	 furiosa-llm[video]==2026.5.0b3

furiosa-llm serve furiosa-ai/Qwen3.5-0.8B --reasoning-parser qwen3   --enable-auto-tool-choice   --tool-call-parser qwen3_coder

curl --fail-with-body -sS http://localhost:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "furiosa-ai/Qwen3.5-0.8B",
    "messages": [{
      "role": "user",
      "content": [
        {
          "type": "video_url",
          "video_url": {
            "url": "https://huggingface.co/datasets/raushan-testing-hf/videos-test/resolve/main/sample_demo_1.mp4"
          }
        },
        {
          "type": "text",
          "text": "Describe what happens in this video."
        }
      ]
    }]
  }'