Strata 0.2.9 adds 1-bit DeepSeek, 1.8x faster prompts on Strix Halo and 1.45x parallel speedups
danielhanchen · x · 2026-10-31
Unsloth's Strata 0.2.9 ships 1-bit DeepSeek V4.1 Flash quantized sets (UD-IQ1S 241GB, UD-IQ1M 258GB) runnable on PC, DGX Spark and Mac.
Key gains:
- Strix Halo: 8K prompts jump from 791 to 1,424 tok/s (1.8x) on Qwen3.8 IQ3S; 32K from 778 to 1,460.
- DeepSeek 2.25: 1.24x faster on 4K prompts, 1.58x on 16K; 2x on DGX Spark short prompts.
- Linux NVIDIA parallel requests: 1.45x at 8 concurrent requests (RTX PRO 5000), identical outputs.
- 2x Intel Arc Pro B65: up to +15.8% prompt throughput (42.1 → 53.8 tok/s).
1-bit DeepSeek decode: 48-88 tok/s on 4x R700, 15-26 on DGX Spark, 15-18 on M5 Max 128GB.
More from Infra
- At SF Compute's "no more compute" townhall, someone pitched me cheap B300 capacity — daniel_mac8 · 2026-10-31
- FastSDCPU Halloween release adds Fibo 2.0 Edit Turbo image editing via OpenVINO on CPU — simpleuserhere · 2026-10-31
- Hot-swapping 18GB models RAM↔VRAM on one 24GB GPU: 1.5s physics vs 29s llama.cpp reality — bjivanovich · 2026-10-31
- Memgraph 3.14's breaking changes will break column-position parsers — release notes double as a regression checklist — JeremyCMorgan · 2026-10-31
- Altman Meets ~50 Samsung Execs in Korea to Deepen Memory, Data-Center Ties — rohanpaul_ai · 2026-10-31
- Alphabet's free cash flow negative again: $9.1B in the red, up to $205B capex planned — YvesMulkers · 2026-10-31