Strix Halo inference engine. Qwen Flash Next Q4_K_XL: 1,628.52pp, 59.41tg single user, 157.22 tok/s 8 users; Qwen27B Q4_K_XL: 656.33pp, 70.56tg tok/s single user with DFlash2
For dual DGX spark users; GLM 5.3 flash got a 50%+ performance boost
For the last few months, I ran DeepSeek v4.0 flash (NVFP4). First 0731, then visionexp because it was a free improvement. I got around 65 tps decode…
Welcome to the Strix Halo Wiki!
Welcome to the Strix Halo Wiki! The purpose of this website it to gather important information and practical guides for systems powered by AMD Ryzen AI MAX and MAX+ processors. Our guides cover various topics and of course there is a separate section…
Niko1221/Strata
Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
Mia on Twitter / X
Fast and Furious ⚡️Update your DeepSeek v4.1 Flash on 3x DGX Sparks- Prose decode tok/s improved by 17%- Code decode tok/s improved by 39%- Mixed decode tok/s improved by 29%Prose single stream is now 50+ tok/s and 85 tok/s at 4 concurrent streams.Changes to 4x DGX… pic.twitter.com/0cMC1kgvDR— Mia (@MiaAI_lab) September 24, 2026
Overview - GroqDocs
Fast LLM inference, OpenAI-compatible. Simple to integrate, easy to scale. Start building in minutes.
Zhijian Liu on Twitter / X
🔥 DFlash x MLX is happening!Shoutout to @aryagm01 for the early work on this. We're building on the momentum. Native MLX support, more models (Qwen3.5), up to 4x faster. Lossless!👉 https://t.co/9CtLKDptNI pic.twitter.com/pMdaCH4oKi— Zhijian Liu (@zhijianliu_) April 15, 2026
Artur Chakhvadze on Twitter / X
We are releasing our first quantized checkpoints for the Qwen3.5 series of models, co-designed jointly with our inference engine to achieve maximum possible performance on Apple hardwareStarting from 0.8B, 2B and 4B modelshttps://t.co/2R8BdhAfzv— Artur Chakhvadze (@norpadon) June 8, 2026
Open Models Inference for Coding · Umans AI
Hosted Kimi K3, GLM 5.2, and DeepSeek V4 Flash. Pay per token, on infrastructure we own.
Benchmarking Subquadratic’s latest model & SSA Kernel | Appen
56× faster than FlashAttention-2 at 1M tokens. Independent efficiency, retrieval, and SWE-Bench benchmark of sparse self-attention. Download the full report.
Daniel Han on Twitter / X
OpenAI's OSS model possible breakdown:1. 120B MoE 5B active + 20B text only2. Trained with Float4 maybe Blackwell chips3. SwiGLU clip (-7,7) like ReLU64. 128K context via YaRN from 4K5. Sliding window 128 + attention sinks6. Llama/Mixtral arch + biasesDetails:1. 120B MoE… https://t.co/bMFp3Z6Gs5 pic.twitter.com/1NFO4utPqr— Daniel Han (@danielhanchen) August 1, 2025
Reverse-engineering GGUF | Post-Training Quantization
Understanding The Math Behind ReSTIR GI
Recently, I had the pleasure of contributing to Nvidia’s Zorah project, the flagship demo for the RTX 50 Series GPUs. My primary role was to provide technical support for light transport in Zorah, which included collaborating with my colleague Daqi Lin to implement a brand new ReSTIR-based global illumination solution, specifically, ReSTIR PT[1], within the NvRTX branch of Unreal Engine. A GDC presentation video on Zorah was released a few weeks ago.
Hardware Arbitrage with Luminal and Positron AI
Luminal compiles off-the-shelf models to Positron Atlas and Nvidia Blackwell to unlock fast and cheap inference across heterogeneous chips.
AMD Ryzen™ AI MAX+ 395 Processor: Breakthrough AI Performance in Thin and Light
The AMD Ryzen™ AI MAX+ 395 (codename: “Strix Halo”) is the most powerful x86 APU in the market today and delivers a significant performance boost over the competition. Powered by 16 “Zen 5” CPU cores...
Strix Halo on Ubuntu looks great - Netstatz
The 2TB NVMe drive that came with this box is shown below. It was modified using gparted on the Ubuntu USB live boot prior to installation. New Linux users
