llamaperf

Strata benchmarks

Strata is a local inference runtime for Qwen3.8-Flash-Next.

101 community reports

Strata results by GPU

Hardware with reported Strata results, with each report's model, quant and speed, newest first. Runs on several cards, with speculative decoding, with batched requests or with part of the model in system RAM say so, since each describes a different setup. Prompt processing is how fast the input is read, not how long you wait for the first word. How to measure time to first token.

Strata on NVIDIA RTX 309025 reports

All Strata results on the NVIDIA RTX 3090Filter Strata reports on the NVIDIA RTX 3090

Strata on NVIDIA RTX 509014 reports

All Strata results on the NVIDIA RTX 5090Filter Strata reports on the NVIDIA RTX 5090

Strata on NVIDIA RTX 40908 reports

All Strata results on the NVIDIA RTX 4090Filter Strata reports on the NVIDIA RTX 4090

Strata on NVIDIA RTX 5070 Ti8 reports

All Strata results on the NVIDIA RTX 5070 TiFilter Strata reports on the NVIDIA RTX 5070 Ti

Strata on NVIDIA RTX 5060 Ti 16GB5 reports

Filter Strata reports on the NVIDIA RTX 5060 Ti 16GB

Strata on AMD RX 7900 XTX4 reports

Filter Strata reports on the AMD RX 7900 XTX

Strata on NVIDIA RTX 4070 Ti Super3 reports

Filter Strata reports on the NVIDIA RTX 4070 Ti Super

Strata on NVIDIA RTX 50703 reports

Filter Strata reports on the NVIDIA RTX 5070

Strata on AMD RX 7900 XT2 reports

Filter Strata reports on the AMD RX 7900 XT

Strata on NVIDIA L40S2 reports

Filter Strata reports on the NVIDIA L40S

Strata on NVIDIA RTX 30702 reports

Filter Strata reports on the NVIDIA RTX 3070

Strata on NVIDIA RTX 3090 Ti2 reports

Filter Strata reports on the NVIDIA RTX 3090 Ti

Strata on NVIDIA RTX 4070 Super2 reports

Filter Strata reports on the NVIDIA RTX 4070 Super

Strata on NVIDIA RTX 40802 reports

Filter Strata reports on the NVIDIA RTX 4080

Strata on NVIDIA V100 16GB2 reports

Filter Strata reports on the NVIDIA V100 16GB

Strata on NVIDIA V100 32GB2 reports

Filter Strata reports on the NVIDIA V100 32GB

Strata, GPU not identified30 reports

What is Strata?

Strata is a local inference runtime for Qwen3.8-Flash-Next. It spreads work across the GPU, system RAM and storage, so available VRAM alone does not describe the setup. Its upstream project documents Windows and Linux installation, a chat interface and APIs for local applications. Check the repository and build in each report before comparing results.

Start with the upstream project for its supported configurations. Community ports can have different installation requirements and hardware-specific optimizations.

Source: Upstream Strata

Strata-V100 is a community fork

Strata-V100 maintains a Volta-focused version of Strata and periodically incorporates upstream changes. Its documentation requires CUDA 12.x and a build containing sm_70 support. CUDA 13 cannot compile that target. Features added for newer GPU architectures do not automatically improve V100 performance.

Record whether a result used this fork, upstream Strata or another build. Keep the V100 memory variant, card count, PCIe layout, system RAM and expert placement beside the result. The V100 32GB hardware page provides additional reported configurations.

Source: Strata-V100 documentation

V100 32GB reports

Strix Halo needs its own evidence

Upstream Strata lists Strix Halo among its experimental community paths, built from source on Linux. Use the Strix Halo reports to inspect the model, operating system, runtime build and speculation settings actually tested. A result from an NVIDIA fork cannot establish performance on the AMD system.

Source: Upstream platform documentation

Strix Halo reports

Read the workload before the speed

Compare actual input and output counts, prefix reuse, quantization, MTP settings and request concurrency. Configured context capacity does not tell you how much text a benchmark processed.

The V100 fork preserves a historical long-context sweep that recorded thermal slowdown during its 128K and 256K prompts. Those measurements belong to the documented build and cooling conditions. They are not measurements of every later revision.

Use the benchmark-reading guide before comparing engines. Keep prefill, generation and application latency separate, and treat missing settings as unknown.

Source: V100 measurement details

How to compare benchmarks

Source-check footer: 13 October 2026. Community-reported evidence, not tests performed by LlamaPerf. Recheck changing platform requirements before publishing.

Compare GPUs running Strata

Every card is compared on one model, Qwen3.6 35B · 3B active at 8-bit, the size and bit class that most cards here have plain runs of. A card with no run of that model shows none rather than a figure from a different model.

GPUVRAMReportsMedian speed, Qwen3.6 35B · 3B active 8-bit
NVIDIA RTX 3090nvidia24GB33no Qwen3.6 35B · 3B active 8-bit run reported
NVIDIA RTX 5090nvidia32GB16no Qwen3.6 35B · 3B active 8-bit run reported
NVIDIA RTX 4090nvidia24GB11no Qwen3.6 35B · 3B active 8-bit run reported
NVIDIA RTX 3060 12GBnvidia12GB10no Qwen3.6 35B · 3B active 8-bit run reported
AMD RX 7900 XTXamd24GB9no Qwen3.6 35B · 3B active 8-bit run reported
NVIDIA RTX 5070 Tinvidia16GB9no Qwen3.6 35B · 3B active 8-bit run reported
NVIDIA RTX 5080nvidia16GB7no Qwen3.6 35B · 3B active 8-bit run reported
AMD Radeon AI PRO R9700 32GBamd32GB6no Qwen3.6 35B · 3B active 8-bit run reported

Strata against other engines

Pairs of plain runs on the same card, of the same model size at the same bit class: one device, one request, no speculative decoding, the whole model in memory. Context length and build still differ between the two sides, and each side shows its own.

No matched pair yet. No card has plain runs of one model size at one bit class on Strata and on another engine, so llamaperf can't say how it compares on speed. Add a run.

Top models on Strata

Frequently asked

Is Strata faster than other engines?

llamaperf has no matched comparison for Strata yet: no card has plain runs of the same model size at the same bit class on Strata and on another engine. Speed claims about engines need that pairing, so this page doesn't make one.