Data sources & attribution
Measured speeds come from public benchmark data, credited here. Data generated 2026-10-16.
Sources
- own measurements (llm-advisor, llama.cpp b11658)
- LocalScore (Mozilla Builders / cjpais, Apache-2.0)
- llama.cpp community benchmark threads incl. the gpt-oss guide #15396 (ggml-org/llama.cpp, MIT)
- witcheer/rtx-5090-benchmarks (Hugging Face, MIT)
- iBlessi/rtx-5080-llm-throughput (Hugging Face, CC-BY-4.0)
- steven0226/speculative-decoding-bench-rtx4090 (Hugging Face, CC-BY-4.0)
- Hugging Face / GitHub open-licensed benchmark repos (see data/benchmarks/README.md)
LocalScore results are community-submitted averages published on localscore.ai (open-source project, Apache-2.0, by Mozilla Builders / Justine Tunney's llamafile team and cjpais). Only public pages were read, at a low request rate. This site is not affiliated with LocalScore or Mozilla.
Quality ratings
Quality sort uses the LMArena leaderboard dataset (text, overall; coding category when the Coding use case is selected), licensed CC-BY-4.0, data dated 2026-10-13. Only models matched by exact name and version are rated; others show “not rated”. The rank is as published by LMArena and includes closed models. Not affiliated with LMArena.
How “measured” and “estimated” work
efficiency = measured_tg / (memory_bandwidth_GBps / read_GB); dense models only; read_GB = params*bits_per_weight/8 >= 3.0 GB; single-device; median; per-hardware factor needs >= 3 rows, else per-kind median, else the global assumption. Global default efficiency before calibration: 0.55.
Coverage by hardware
| Hardware id | Data points | Sources |
|---|---|---|
| rtx-5090 | 38 | llamacpp-discussion-10879-scoreboard, llamacpp-discussion-15013-comments, llamacpp-discussion-15013-scoreboard, localscore-model-pages, hf-witcheer-rtx-5090-benchmarks, llamacpp-discussion-15396 |
| rtx-pro-6000 | 6 | llamacpp-discussion-15013-comments, llamacpp-discussion-15013-scoreboard, localscore-model-pages, llamacpp-discussion-15396 |
| rtx-4090 | 24 | llamacpp-discussion-10879-comments, llamacpp-discussion-10879-scoreboard, llamacpp-discussion-15013-comments, llamacpp-discussion-15013-scoreboard, localscore-model-pages, llamacpp-discussion-15396, own-measurement, hf-steven0226-speculative-decoding-bench-rtx4090 |
| rtx-5080 | 11 | llamacpp-discussion-10879-comments, llamacpp-discussion-10879-scoreboard, llamacpp-discussion-15013-comments, llamacpp-discussion-15013-scoreboard, localscore-model-pages, localscore-latest-runs, hf-iBlessi-rtx-5080-llm-throughput, llamacpp-discussion-15396 |
| rtx-5070ti | 5 | llamacpp-discussion-10879-comments, llamacpp-discussion-10879-scoreboard, llamacpp-discussion-15013-comments, llamacpp-discussion-15013-scoreboard, localscore-model-pages, llamacpp-discussion-15396 |
| rtx-3090 | 20 | llamacpp-discussion-10879-comments, llamacpp-discussion-10879-scoreboard, llamacpp-discussion-15013-comments, llamacpp-discussion-15013-scoreboard, localscore-model-pages, llamacpp-discussion-15396, own-measurement |
| rtx-5070 | 11 | llamacpp-discussion-15013-scoreboard, localscore-model-pages, own-measurement |
| rtx-3060-12 | 11 | llamacpp-discussion-15013-comments, llamacpp-discussion-15013-scoreboard, localscore-model-pages, own-measurement |
| dgx-spark | 19 | llamacpp-discussion-10879-comments, llamacpp-discussion-10879-scoreboard, llamacpp-discussion-15013-comments, llamacpp-discussion-15013-scoreboard, llamacpp-benches-dgx-spark, llamacpp-discussion-15396, llamacpp-discussion-16578-comments, llamacpp-benches-nemotron, localscore-model-pages |
| strix-halo | 26 | llamacpp-discussion-10879-comments, llamacpp-discussion-15013-comments, llamacpp-discussion-15021-comments, llamacpp-discussion-15021-scoreboard, llamacpp-discussion-23659-body, hf-axjns-strix-halo-inference-bench, github-strix-halo-guide-BENCHMARKS, localscore-model-pages |
| m4 | 9 | llamacpp-discussion-4167-comments, localscore-model-pages |
| m4-max-32 | 7 | llamacpp-discussion-4167-comments, localscore-model-pages, llamacpp-discussion-15396 |
| m5-max-40 | 8 | llamacpp-discussion-4167-comments, localscore-model-pages, hf-studioburnside-mlx-local-inference-benchmarks |
| rx-7900xtx | 8 | llamacpp-discussion-10879-comments, llamacpp-discussion-10879-scoreboard, llamacpp-discussion-15021-comments, llamacpp-discussion-15021-scoreboard, localscore-model-pages |
| m3-ultra | 5 | llamacpp-discussion-10879-comments, llamacpp-discussion-10879-scoreboard, localscore-model-pages, llamacpp-discussion-15396 |
| rtx-4070 | 11 | llamacpp-discussion-10879-scoreboard, localscore-model-pages, own-measurement |
| rtx-4060ti-16 | 14 | localscore-model-pages, own-measurement |
| rtx-4060 | 2 | localscore-model-pages |
| m5-max-32 | 3 | localscore-model-pages |
| m4-max-40 | 4 | localscore-model-pages, hf-pgmharikrishnan-mlx-serve-local-llm-benchmarks-2026-10 |
| m5-pro | 3 | localscore-model-pages |
| m4-pro | 3 | localscore-model-pages |
| m6 | 1 | localscore-model-pages |
| m5 | 3 | localscore-model-pages |
Privacy & analytics
This site uses Umami Cloud for anonymous, cookie-free analytics: page views and which options are chosen (hardware type, device, use case, context length, sort and filter names, whether a command or link was copied, theme, and clicks on submit/sponsor/affiliate links). We never record what you type in search or the benchmark form, and no personal data is stored in events. Browsers that send Do Not Track are not counted, and blockers work fine; the site behaves the same either way.
Community submissions
0 reviewed submissions, always shown as unverified. Submit yours.