COLLECTION / RECIPES

Recipes

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

21 records

Apple M1 Max 64GB1 × 64 GBgemma-4-26B-A4BQ8_0llama.cppreference794context51.3 tok/smeasuredcandidate llama.cpp evidence measured

Apple M1 Max 64GB1 × 64 GBgemma-4-26B-A4BQ8_0ollamareference2Kcontext46.6 tok/smeasuredcandidate ollama evidence measured

Apple M3 Ultra 96GB 60-core GPU1 × 96 GBgemma-4-26B-A4B8bitmlxreference255.9Kcontext68.7 tok/smeasuredcandidate mlx evidence measured

Apple M3 Ultra 96GB 80-core GPU1 × 96 GBgemma-4-26B-A4B8bitmlxreference255.9Kcontext69.5 tok/smeasuredcandidate mlx evidence measured

Apple M3 Ultra 512GB1 × 512 GBgemma-4-26B-A4BINT4mlxreference2Kcontext98.9 tok/smeasuredcandidate mlx evidence measured

Apple M4 Max 36GB1 × 36 GBgemma-4-26B-A4B8bitmlxreference24.1Kcontext55.5 tok/smeasuredcandidate mlx evidence measured

Apple M4 Max 128GB1 × 128 GBgemma-4-26B-A4B8bitmlxreference255.9Kcontext64.8 tok/smeasuredcandidate mlx evidence measured

Apple M4 Pro 48GB1 × 48 GBgemma-4-26B-A4B8bitmlxreference255.9Kcontext41.7 tok/smeasuredcandidate mlx evidence measured

Apple M5 Max 128GB1 × 128 GBgemma-4-26B-A4B8bitmlxreference255.9Kcontext75 tok/smeasuredcandidate mlx evidence measured

Apple M5 Pro 64GB1 × 64 GBgemma-4-26B-A4B8bitmlxreference88.3Kcontext46.6 tok/smeasuredcandidate mlx evidence measured

GeForce RTX 30902 × 24 GBgemma-4-26B-A4BINT8_W8A16vllmreferenceUnknown contextcontext162.9 tok/smeasuredcandidate vllm evidence measured

GeForce RTX 40701 × 12 GBgemma-4-26B-A4BQ4_K_XLvllmreferenceUnknown contextcontext30.6 tok/smeasuredcandidate vllm evidence measured

GeForce RTX 50901 × 32 GBgemma-4-26B-A4BINT8vllmreferenceUnknown contextcontext296.4 tok/smeasuredcandidate vllm evidence measured

Intel Arc Pro B701 × 32 GBgemma-4-26B-A4BQ8_K_XLllama.cppreference2Kcontext43.3 tok/smeasuredcandidate llama.cpp evidence measured

Radeon AI PRO R97002 × 32 GBgemma-4-26B-A4BMXFP4_MOEllama.cppreference794context87.3 tok/smeasuredcandidate llama.cpp evidence measured

Radeon AI PRO R97003 × 32 GBgemma-4-26B-A4BQ8_0llama.cppreference794context73.7 tok/smeasuredcandidate llama.cpp evidence measured

Radeon AI PRO R97002 × 32 GBgemma-4-26B-A4BQ8_K_XLllama.cppreference256Kcontext131.4 tok/smeasuredcandidate llama.cpp evidence measured

RTX PRO 6000 Blackwell1 × 96 GBgemma-4-26B-A4Bint8_per_channel_weight_onlyvllmreferenceUnknown contextcontext276.7 tok/smeasuredcandidate vllm evidence measured

Ryzen AI Max+ 3951 × 128 GBgemma-4-26B-A4BQ4_K_Mllama.cppreference16Kcontext65.9 tok/smeasuredcandidate llama.cpp evidence measured

Ryzen AI Max+ 3951 × 128 GBgemma-4-26B-A4BQ6_Kllama.cppreference16Kcontext55.5 tok/smeasuredcandidate llama.cpp evidence measured

Ryzen AI Max+ 3951 × 128 GBgemma-4-26B-A4BQ8_0llama.cppreference16Kcontext47.2 tok/smeasuredcandidate llama.cpp evidence measured