COLLECTION / RECIPES

Recipes

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

61 records

Apple M3 Ultra 96GB 80-core GPU1 × 96 GBDeepSeek-V4-FlashIQ2XXSllama.cppreference252.1Kcontext22.5 tok/smeasuredcandidate llama.cpp evidence measured

Apple M3 Ultra 96GB 80-core GPU1 × 96 GBDeepSeek-V4-FlashIQ2XXSllama.cppreference252.1Kcontext8.1 tok/smeasuredcandidate llama.cpp evidence measured

Apple M3 Ultra 96GB 80-core GPU1 × 96 GBDeepSeek-V4-FlashIQ2XXSllama.cppreference216.3Kcontext22.5 tok/smeasuredcandidate llama.cpp evidence measured

Apple M3 Ultra 96GB 80-core GPU1 × 96 GBDeepSeek-V4-FlashUD-Q2_K_XLllama.cppreference232.1Kcontext7.1 tok/smeasuredcandidate llama.cpp evidence measured

Apple M3 Ultra 512GB1 × 512 GBDeepSeek-V4-FlashIQ2XXSmlxreferenceUnknown contextcontext33 tok/smeasuredcandidate mlx evidence measured

Apple M4 Max 128GB1 × 128 GBDeepSeek-V4-FlashIQ2XXSllama.cppreference47.9Kcontext20.1 tok/smeasuredcandidate llama.cpp evidence measured

Apple M5 Max 128GB1 × 128 GBDeepSeek-V4-FlashIQ2_XXS (asym MoE)llama.cppreference32Kcontext26.3 tok/smeasuredcandidate llama.cpp evidence measured

Apple M5 Max 128GB1 × 128 GBDeepSeek-V4-FlashQ2llama.cppreference4Kcontext12.1 tok/smeasuredcandidate llama.cpp evidence measured

Apple M5 Max 128GB1 × 128 GBDeepSeek-V4-FlashIQ2XXSllama.cppreference255.9Kcontext25.5 tok/smeasuredcandidate llama.cpp evidence measured

Apple M5 Max 128GB1 × 128 GBDeepSeek-V4-FlashIQ2XXSllama.cppreference252.1Kcontext12.7 tok/smeasuredcandidate llama.cpp evidence measured

Apple M5 Max 128GB1 × 128 GBDeepSeek-V4-FlashIQ2XXSllama.cppreference255.9Kcontext25.5 tok/smeasuredcandidate llama.cpp evidence measured

Apple M5 Max 128GB1 × 128 GBDeepSeek-V4-FlashUD-Q2_K_XLllama.cppreference232.1Kcontext13 tok/smeasuredcandidate llama.cpp evidence measured

Apple M5 Max 128GB1 × 128 GBDeepSeek-V4-Flashunknownllama.cppreference255.9Kcontext32.2 tok/smeasuredcandidate llama.cpp evidence measured

Apple Max 128GB, generation unspecified1 × 128 GBDeepSeek-V4-Flash2bit-DQmlxreferenceUnknown contextcontext17 tok/smeasuredcandidate mlx evidence measured

GeForce RTX 30901 × 24 GBDeepSeek-V4-FlashIQ2_XXSllama.cppreference26.0Kcontext11 tok/smeasuredcandidate llama.cpp evidence measured

GeForce RTX 30901 × 24 GBDeepSeek-V4-Flash-0731IQ2_XSSllama.cppreference2Kcontext20.7 tok/smeasuredcandidate llama.cpp evidence measured

GeForce RTX 30904 × 24 GBDeepSeek-V4-Flash-0731UD-IQ2_Mllama.cppreference128Kcontext41.1 tok/smeasuredcandidate llama.cpp evidence measured

GeForce RTX 30907 × 24 GBDeepSeek-V4-Flash-0731UD-IQ4_XSllama.cppreference256Kcontext50.8 tok/smeasuredcandidate llama.cpp evidence measured

GeForce RTX 30907 × 24 GBDeepSeek-V4-Flash-0731UD-Q4_K_XLllama.cppreference256Kcontext33.9 tok/smeasuredcandidate llama.cpp evidence measured

GeForce RTX 30908 × 24 GBDeepSeek-V4-Flash-0731FP8_E4M3vllmreferenceUnknown contextcontext91.2 tok/smeasuredcandidate vllm evidence measured

GeForce RTX 30908 × 24 GBDeepSeek-V4-Flash-0731FP8vllmreferenceUnknown contextcontext404.5 tok/smeasuredcandidate vllm evidence measured

GeForce RTX 50904 × 32 GBDeepSeek-V4-Flash2-bit experts (vLLM-Moet W2) + FP8 KVvllmreferenceUnknown contextcontext212 tok/smeasuredcandidate vllm evidence measured