Halogen

halogen™ is local inference that behaves like a frontier API. Strix Halo + NPU, OpenAI & Anthropic APIs on one port.

[Halogen @halogenengine]

Halogen 0.20.4 is out for Strix Halo
• Short follow-up messages read 15–50% faster in cached conversations
• ~10–20% more combined throughput at 16 requests at once (>8 slots)
• Fix: tool_choice “none” respected after earlier tool calls
https://t.co/gTstxfEfJr

[Halogen @halogenengine]

Halogen 0.20.2 and 0.20.3 are out for Strix Halo
• @vllm_project��s d3-lite 14–18% faster on the @AIatAMD NPU, d3-nano 7–9%, both using less energy
• Up to ~4% sooner first token on long conversations
• NPU decisions on long prompts start sooner
https://t.co/gTstxfENyZ

[Halogen @halogenengine]

Halogen 0.20.1 is out for Strix Halo
• Faster follow-up turns: first token in 0.19s vs 0.60s on a code-edit session
• An answer even when the model ends its turn mid-reasoning
• Quieter logs on long sessions
https://t.co/gTstxfENyZ

[Halogen @halogenengine]

Halogen 0.20.0 is out for Strix Halo
• OvisOCR3 document parsing on the @AIatAMD NPU (your own weights)
• OCRBench 85.7 vs 85.3 upstream
• ROCm 10.1: ~9–15% less energy, same speed
• 2–6% faster prompt reading on Qwen3.8-Flash-Next
https://t.co/gTstxfENyZ

[Halogen @halogenengine]

Halogen 0.19.1 for Strix Halo:

• XRT now in the image: NPU needs only the host driver and firmware
• Malformed tool schemas get a 400 instead of locking up generation slots
• Faster first token when a /v1/completions prompt extends a recent one

https://t.co/gTstxfEfJr

[Halogen @halogenengine]

Halogen 0.19.0 for @AIatAMD Strix Halo:

• Preview: Halogen Spotter, an opt-in NPU agent guard that checks tool results
• Whole-file rewrites by coding agents ~30% faster
• NPU stability fixes

Spotter: https://t.co/R4xoQ77y2K
Changelog: https://t.co/gTstxfEfJr

[Halogen @halogenengine]

Halogen Flash is now a backend in @lemonade_server from @AIatAMD, just merged for Strix Halo:

• Runs Qwen3.8-Flash-Next with v2 quants
• Pinned to the Halogen 0.18.22 release
• Tested in CI on the full model

Thanks, Lemonade!

https://t.co/uNNv4ck7i2

[Halogen @halogenengine]

Halogen 0.18.23: the LFM3 family from @liquidai now runs on the Strix Halo NPU

• LFM3-VL-2B: pictures and video
• LFM3-Audio-2B: speech in and speech out
• LFM3-Omni-2B: every medium on one model
• Fix: replies no longer invent a user turn

https://t.co/gTstxfENyZ

[Halogen @halogenengine]

Another excellent video from @dcapitella
Thanks for including Halogen!

Strix Halo vs. two Intel Arc Pro B70s on Qwen3.8-Flash-Next:

• ~64 tok/s at 64K context
• 1,217 tok/s prompt processing
• Close in generation, faster at prompt processing

https://t.co/SdNRNVNWqg

[Halogen @halogenengine]

Wendell’s @Level1Techs review of the new @gmktecdotcom EVO-X5 Pro: with Halogen, Qwen3.8-Flash-Next on the Ryzen AI Max+ 495 runs at parity with DGX Spark.

Thanks Wendell!
https://t.co/9QHg0t6q9u

[Halogen @halogenengine]

Run your own @liquidai d1-2B on the @AIatAMD Strix Halo NPU. Halogen 0.18.22 is out.

• Up to 16 questions on text, images or video in one pass
• Finds objects in photos with bounding boxes
• ~0.35 s per image question

https://t.co/gTstxfENyZ

[Halogen @halogenengine]

Run your own @liquidai LFM3-1.7B on the @AIatAMD Strix Halo NPU. Halogen 0.18.21 is out.

Your checkpoint or fine-tune, up to ~38 tok/s
Full 262k context
Long NPU requests no longer stall behind a reply

https://t.co/gTstxfENyZ

[Halogen @halogenengine]

Halogen 0.18.20 for Strix Halo is out.

• 2 requests at once ~20% faster, 3 at once ~9%
• JSON schemas now work on requests with images
• Strict tools enforced with images too
• Clear NPU error on old XRT

https://t.co/gTstxfENyZ

[Halogen @halogenengine]

Running Qwen3.8-Flash-Next on Strix Halo? Halogen 0.18.19 is out.

• ~3% faster with 3 requests at once
• No more 64-image cap for screenshot agents
• Claude Code auto-compacts long prompts
• Tool-call fixes

#StrixHalo #LocalLLM #qwen
https://t.co/gTstxfEfJr

[Halogen @halogenengine]

Halogen 0.18.18 is out.

• Claude Code sessions with MCP tools keep their prompt cache
• ggml-org's official IQ4_NL GGUF now loads
• Short requests start during long prompts
• Tool-call args with </tool_call> arrive whole

Same outputs, same speed.
https://t.co/gTstxfEfJr

[Halogen @halogenengine]

Halogen 0.18.17 is out.

Agents sharing a prompt no longer swap turns on a full KV pool
Claude Code MCP tool search works on /v1/messages
Tool-call args with markup arrive intact
/v1/models reports image support

Same outputs, same speed.
https://t.co/gTstxfEfJr