Strata 0.2.10 doubles dual-GPU decode speed, runs Qwen3.8 at 20+ tok/s on 8GB cards
evilsocket · x · 2026-11-03
Strata 0.2.10 makes DFlash2 the default drafter and delivers big local-inference gains: Qwen3.8 Flash Next now runs above 20 tokens/s on an 8GB GPU with 128K context.
- 2× RTX A6000: decode 201 → 399 tok/s (+98%); 2× RTX PRO 5000 +70%; 4 concurrent users +85%
- 4× Radeon R9700: single user 101.6 → 180.9 tok/s, 8 users +42%
- Strix Halo +12%, Apple M5 Max +10.3% for agents, long-document wait −44%
- DFlash2 beats MTP by 4–12% decode speed across single cards
Also adds one model across all cards on Linux and DeepSeek image models on Mac. Full benchmarks in the GitHub release.
More from Infra
- NetBird reverse proxy adds path-based auth: per-path SSO, bypass, or block on public apps — philipvollet · 2026-11-04
- Intel and Micron say supply, not demand, caps server growth ahead of AMD's Q4 guide — tengyanAI · 2026-11-04
- DeepSeek open-sources DeepInfer inference engine delivering 26,000+ tokens/sec per GPU — bycloud · 2026-11-04
- Musk: AI boom hits a wall by end of 2026 — power, not chips, is the bottleneck — elonmusk · 2026-11-04
- PROCS paper: parallel snapshot compression cuts inference cold-start latency up to 44.7% — charles_irl · 2026-11-04
- Shanghai field trip: Greater China investors so bearish some doubt even 2027 AI capex — terryyuezhuo · 2026-11-04