🍋 LemonMatrix Leaderboard

Community-submitted benchmark results for AMD's Lemonade runtime, collected with LemonMatrix -- every number here was measured by a contributor's own machine talking to its own local Lemonade instance; this site only renders what Lemonade itself reported, never anything read from the machine directly.

2.5 tok/s refused
Qwen3.8-27B-GGUF-UD-Q4_K_XL · UD-Q4_K_XL on AMD Radeon RX 7900 XTX, llamacpp-rocm: another model was busy during the run, so it is not ranked. Refused run · clean run: 14.2 tok/s
63.6% saved
Best router policy vs. all-cloud: LM-D-three-tier on Radeon-AI-PRO-R9700, 65% of 120 requests kept local, measured baseline
320 tok/s
Fastest on Radeon, Qwen3-1.7B-GGUF · Q4_0: AMD Radeon AI PRO R9700 (gfx1201) (Radeon-AI-PRO-R9700), llamacpp-vulkan
386 tok/s
Fastest on Instinct, Qwen3-1.7B-GGUF · Q4_0: gfx950 GPU (AMD-Instinct-MI350X), llamacpp-rocm
197
LLM runs
194
Valid runs
11
Profiles (machines)
143
Other modality runs
3
Router runs

LLM leaderboard

Runs are grouped by model and quantization, and ranked by decode speed within each group. Repeated runs of the same machine and backend are folded into their best one; untick the box to see them all. Click a column header to rank by another column. ↑ higher is better, ↓ lower is better. Filters are applied together (AND).

* Prefill and TTFT were measured with the server's prompt cache on (runs before 2026-11-14, and router runs), so they time a cache hit and are likely too fast. Decode is not affected. Prompt is the prompt length per request: prefill rises and decode falls with it, so compare runs of a similar prompt size.

Other modalities

Non-LLM Lemonade recipes -- text-to-speech, speech-to-text, image/audio/3D generation, embeddings, reranking, classification. Each has its own throughput metric, so these aren't ranked against the table above. Every metric here is higher-is-better; for audio, × real time is seconds of audio per second of processing.

DateProfileModelBackendEngine Metric ↑Valid

Router cost runs

A collection.router run with a mixed prompt set: where the router sent each request, what that cost, and what sending every request to the baseline model would have cost. A measured baseline sends each prompt once to the baseline model and prices its own token counts; an estimated one reuses the routed model's counts, which understates a reasoning model's cost. Prices are USD per million tokens, from Lemonade's own cost metadata or the submitter's --cost; non-cloud models count as free. Costs are shown per 1,000 requests.

DateProfileRouterSaved ↑Cost / 1k req Baseline / 1k reqBaselineSplitLocal RequestsValid
2026-11-14Radeon-AI-PRO-R9700user.LM-D-three-tier63.6%$0.12$0.34DeepSeek-V4-Pro (measured)Qwen3-1.7B-GGUF 28% · Qwen3.5-35B-A3B-GGUF 38% · DeepSeek-V4-Pro 35%65%120valid
2026-11-14Radeon-AI-PRO-R9700user.LM-C-llm-classifier1.1%$0.16$0.16Qwen3-235B-A22B-Instruct-2507 (measured)Qwen3-235B-A22B-Instruct-2507 95% · Qwen3.5-35B-A3B-GGUF 5%5%120valid
2026-11-14Radeon-AI-PRO-R9700user.LM-A-keyword-code-to-cloud53.6%$0.04$0.09deepseek-v3.2 (measured)Qwen3-4B-GGUF 62% · deepseek-v3.2 38%62%120valid