COLLECTION / RECIPES

Recipes

One artifact × hardware × engine unit. Validated rows are launch-safe; measured speed sweeps attach to each recipe as evidence.

11 records

Apple M3 Ultra 96GB 60-core GPU1 × 96 GBgemma-4-12b-coder-fable5-composer2.5-v18bitmlxreference64.1Kcontext40.5 tok/smeasuredcandidate mlx evidence measured

Apple M3 Ultra 96GB 80-core GPU1 × 96 GBgemma-4-12b-coder-fable5-composer2.5-v18bitmlxreference64.1Kcontext39.4 tok/smeasuredcandidate mlx evidence measured

Apple M4 Max 36GB1 × 36 GBgemma-4-12b-coder-fable5-composer2.5-v18bitmlxreference64.1Kcontext25.1 tok/smeasuredcandidate mlx evidence measured

Apple M4 Max 128GB1 × 128 GBgemma-4-12b-coder-fable5-composer2.5-v18bitmlxreference64.1Kcontext31 tok/smeasuredcandidate mlx evidence measured

Apple M4 Pro 48GB1 × 48 GBgemma-4-12b-coder-fable5-composer2.5-v18bitmlxreference64.1Kcontext17.3 tok/smeasuredcandidate mlx evidence measured

Apple M5 Max 128GB1 × 128 GBgemma-4-12b-coder-fable5-composer2.5-v18bitmlxreference64.1Kcontext35.8 tok/smeasuredcandidate mlx evidence measured

Apple M5 Max 128GB1 × 128 GBGemma-4-12B-itQ8_0llama.cppreference2Kcontext38.4 tok/smeasuredcandidate llama.cpp evidence measured

Apple M5 Pro 64GB1 × 64 GBgemma-4-12b-coder-fable5-composer2.5-v18bitmlxreference64.1Kcontext19.9 tok/smeasuredcandidate mlx evidence measured

GeForce RTX 50901 × 32 GBgemma-4-12B-coder-fable5-composer2.5-v1-GGUFNVFP4vllmreferenceUnknown contextcontext183 tok/smeasuredcandidate vllm evidence measured

GeForce RTX 50901 × 32 GBGemma-4-12B-itQ8_0lmstudioreference2Kcontext85.3 tok/smeasuredcandidate lmstudio evidence measured

Ryzen AI Max+ 3951 × 128 GBGemma-4-12B-itQ4_K_Mllama.cppreference4Kcontext24.8 tok/smeasuredcandidate llama.cpp evidence measured