Strata Patches Deliver 3-6x Speedups and Fix Tool Calls on Dual-GPU Rig
Fz1zz · reddit · 2026-11-10
On an RTX 5090 + RTX 4070 Ti SUPER (31 GB RAM), the author benchmarked a tuned Strata v0.2.1 build with 8 patches (leveraging Opus 5.5 and Fable 5.2 optimizations) against upstream v0.2.13's default install — same PC, same model files (Qwen3.8-Flash-Next UD-IQ3XXS at full 262,144-token context):
- Decode speed: 251 vs 85 tok/s (3.0×)
- 20K-token prompt: 10.3s vs 65.0s (6.3×)
- 80K-token prompt: 15.4s vs 89.2s (5.8×)
- 250K-token prompt: 46.7s vs 257.4s (5.5×)
- Per-turn at 256K context: 2.1s vs 12.7s (6.0×)
- Tool-call probe: 30/30 vs 9/30 passed
The author is surprised these optimizations are "left on the table"; patches and benchmarks are open-sourced at github.com/ExTV/strata-5090-4070.
More from Infra
- Daytona CEO: Model inference is just a small slice of the AI agent infrastructure opportunity — mattturck · 2026-11-10
- Coherent signs $20B in laser orders shipping from 2028 for AI CPO optics — demian_ai · 2026-11-10
- Invest in engineering, not agent factories: Geoffrey Huntley says Nix and fast builds get uplifted by model drops — sloppenheimer · 2026-11-10
- fast_float benchmarks: parser hints double throughput to 355M numbers/sec — lemire · 2026-11-10
- Anthropic retracts IPO filing after sending bankers wrong price range; Google, Nvidia, Microsoft take $26B private slice — GaryMarcus · 2026-11-10
- Taiwan supply chain: Nvidia CEO controversy won't derail AI demand long-term — nordicinst · 2026-11-10