The best Qwen3.8 27B model for AMD Strix Halo is here. In collaboration with Pete Hopton and the Kairic.ai team, I bring you: Qwen3.8-27B-IU4-KAIRIC-EDGE. To our knowledge, this is the world’s first public 27B LLM release with an accelerated native-IU4 inference lane, purpose-built for AMD Strix Halo (gfx1151). Why does IU4 matter? Most 4-bit models save memory but still convert weights into wider formats during execution. Our IU4 lane keeps four-bit integer data in the active compute path and maps it directly onto RDNA 3.5 integer hardware. That reduces memory traffic and turns quantization from a storage format into a genuine hardware execution strategy. The matched IU4 operator reached 104.66 TOPS—1.94× FP16 and 1.93× IU8. This opens a path toward faster, more efficient local AI without trading away model quality. I can't wait to see the things that @Italianclownz and other wizards do with it. Powered by Prompt Forge + Dual View, Kairic Edge acceleration, 262K context: HumanEval metrics: • 47.73 tok/s full-suite TG • +85% vs Unsloth Dynamic Q4 • +89% vs Unsloth Dynamic Q6 • Beat Unsloth Dynamic v3 q6 in human eval by 2 questions. This is also the first release powered by Kairic.ai, a new company focused on co-designing AI software around the hardware it actually runs on. Model, benchmarks, source, build guide, and runner: huggingface.co/jcbtc/Qwen3.8…

Want to run a 120b model on your 5090 with 256k context? I realized with a little bit of tweaking we can have the new @poolsideai Laguna S 2.1 in high precision NVFP4 at 256k context from home. So I made an NVFP4 model and configured it for you guys. Similar speeds to the AMD Halo, but now for those with GPU rigs. 256k context max, 20+ TPS 1k+ prefill and straight NVFP4. By far the smartest model you will ever run on your 5090. huggingface.co/jcbtc/Laguna-…