Community performance index
Real hardware.
Real numbers.
A living map of what local models can do on AMD-powered Macs. Standardized runs, honest context, and community data you can reproduce.
Submit your benchmark ↗Runs indexed18and growing
Models tested13Dense + MoE
GPU families07RDNA + GCN
Fastest generation250tokens / sec
| Model / quant | Hardware | Architecture | Prompt | Generation | Source |
|---|---|---|---|---|---|
Llama 3.2 1BQ4_K_M · Dense | RX 6700 XT12 GB VRAM | RDNA 2 | 2,994tok/s | 250tok/s | ToshLLM LabJul 15, 2026 |
Qwen3 4BQ4_K_M · Dense | RX 6900 XT16 GB VRAM | RDNA 2 | 291.4tok/s | 97.5tok/s | Community testerJul 16, 2026 |
Qwen3 4BQ4_K_M · Dense | RX 6700 XT12 GB VRAM | RDNA 2 | 1,005tok/s | 94tok/s | ToshLLM LabJul 15, 2026 |
gpt-oss 20BQ4_K_M · MoE · Full VRAM | RX 6700 XT12 GB VRAM | RDNA 2 | 972tok/s | 93tok/s | ToshLLM LabJul 15, 2026 |
Gemma 3 4BQ4_K_M · Dense | RX 6700 XT12 GB VRAM | RDNA 2 | 1,168tok/s | 90tok/s | ToshLLM LabJul 15, 2026 |
Qwen3 8BQ4_K_M · Dense | RX 6700 XT12 GB VRAM | RDNA 2 | 476tok/s | 60tok/s | ToshLLM LabJul 15, 2026 |
Qwen3.6 14B A3BQ5_K_M · MoE · Full VRAM | RX 6700 XT12 GB VRAM | RDNA 2 | 736tok/s | 56tok/s | ToshLLM LabJul 15, 2026 |
Gemma 4 12BQ4_K_M · Dense | RX 6950 XT16 GB VRAM | RDNA 2 | 595.7tok/s | 53.4tok/s | Community testerJul 16, 2026 |
Qwen3 4BQ4_K_M · Dense | RX 5600 XT6 GB VRAM | RDNA 1 | 100tok/s | 52.1tok/s | Community testerJul 16, 2026 |
Qwen K-quantK-quant · Dense · Wave64 path; model details pending verification | RX 5808 GB VRAM | GCN / Vega | 285tok/s | 51tok/s | Community testerJul 11, 2026 |
Gemma 4 26B A4BQ4_K_M · MoE | Radeon Pro Vega II32 GB VRAM | GCN / Vega | 190.3tok/s | 47.9tok/s | Mac Pro communityJul 16, 2026 |
Qwen3-Coder 30B A3BQ4_K_M · MoE · ncmoe 23 | RX 6950 XT16 GB VRAM | RDNA 2 | 587.3tok/s | 47.8tok/s | Community testerJul 16, 2026 |
Qwen3.5 9BQ4_K_M · Dense | RX 6700 XT12 GB VRAM | RDNA 2 | 411tok/s | 45tok/s | ToshLLM LabJul 15, 2026 |
Qwen3.6 35B A3BQ4_K_S · MoE · All experts in VRAM | Radeon Pro Vega II32 GB VRAM | GCN / Vega | 284.7tok/s | 42.1tok/s | Mac Pro communityJul 16, 2026 |
Gemma 4 12BQ4_K_XL · Dense | RX 6700 XT12 GB VRAM | RDNA 2 | 342tok/s | 34tok/s | ToshLLM LabJul 15, 2026 |
Qwen3.5 122B A10BQ5_K_M · MoE · Four-GPU split | 4× W6800X128 GB VRAM | RDNA 2 | 226.5tok/s | 26.1tok/s | Basten7Jul 15, 2026 |
Qwen3.6 35B A3BQ4_K_S · MoE · CPU expert offload | RX 6700 XT12 GB VRAM | RDNA 2 | 434tok/s | 25tok/s | ToshLLM LabJul 15, 2026 |
Gemma 4 26B A4BMXFP4 · MoE · CPU expert offload | RX 6700 XT12 GB VRAM | RDNA 2 | 552tok/s | 22tok/s | ToshLLM LabJul 15, 2026 |
∅
No matching runs
Try another model, GPU, or filter.
Reproduction protocol
Same test.
Useful comparison.
01
Control the context
Prompt processing uses 512 tokens; generation uses 128. Run three repetitions after a warm-up.
02
Record the machine
GPU, VRAM, quantization, model hash, ToshLLM version and offload settings travel with every result.
03
Preserve provenance
App-signed runs remain labeled as community data. Internally inconsistent or extreme results wait for manual review.
Standard llama-bench command
TOSH_FA_AMD=1 GGML_METAL_CONCURRENCY_DISABLE=1 GGML_SCHED_PREFETCH_EXPERTS=1 GGML_CPU_NO_REPACK=1 ./llama-bench -m MODEL.gguf -ngl 99 --mmap 0 -fa auto -p 512 -n 128 -r 3 --progressMeasurement note: the environment flags match ToshLLM's AMD and MoE runtime. MTP speculative decoding is not supported by llama-bench; measure it with real generation through llama-server.