Strata v0.2.7: DeepSeek V4.1 Flash on Strix Halo, 512K/1M context
Strata 0.2.7 ships a dedicated gfx1151 engine package so DeepSeek V4.1 Flash runs on Ryzen AI Max Windows PCs, speeds up prompt reads (8K prompt 342 to 399 tok/s on four R9700 cards), and exposes 512K/1M context options.
NEW STRATA RELEASE: v0.2.7 (published 2026-10-27T14:54:42Z, stable) previous known: v0.2.4 by @coldniko (2026-10-24T02:18:00Z) | fleet now: P2005 Control v0.2.4, Qwen engine 0.1.48 (running, brain), DeepSeek engine 0.1.49 + IQ2_XS pack on disk (not running); Mac Studio Control v0.2.4, Metal engine 0.1.48.1 (Qwen serve); P1005 offline, P2002 offline 1. NVIDIA cards with 12 GB or less answer faster. The router now prefers experts that are already on the card, which saves page-ins from RAM. On an RTX 3060 12 GB (Qwen3.8-Flash-Next IQ3_XXS, five paired runs a... 2. Radeon RX 9000 and AI PRO R9700: faster Qwen. Two changes: the large per-token lookup table stays in RAM when the PC has room for it, and the decode split between card and processor starts at a better value.... 3. DeepSeek-V4.1-Flash runs on Ryzen AI Max (Strix Halo) PCs with Windows. It has its own engine package for the gfx1151 chip, and the app or strata launch downloads it with the model. The new guide docs/STRIX_... 4. DeepSeek reads prompts faster on most PCs. The engine no longer re-reads a batch of experts from the SSD after every prompt, picks the attention buffer size by itself and changes how the RAM store is filled.... 5. Four Radeon AI PRO R9700 cards run DeepSeek faster, with five experts per word by default. With the launcher's own settings, DeepSeek 2.25: an 8K prompt 342 to 399 tok/s on the first read and 967 to 1,350 on... 6. DeepSeek can use 512K and 1M contexts. Setup, the Terminal and the app offer them up to the limit of the model files, with a plain "not tested yet" note past 256K. The default stays 8K, and the app now loads... 7. Two NVIDIA cards read long prompts much sooner on a layer split. On a layer split over two CUDA cards the prompt is now read layer by layer in large pieces (layer-major prefill, on by default), so each layer... 8.... notes: https://github.com/Niko1221/Strata/releases/tag/v0.2.7 Nothing installed or changed yet - waiting for your go-ahead.