The best local LLMs for the M5 Ultra 256GB

DeepSeek V4 Flash (284B (13B active)) fits comfortably in the M5 Ultra 256GB's memory at FP4, using about 168GB with 8K context and generating ~45 tokens/sec. It is the most capable model in our catalog that runs well on this card. It is also our top pick for coding and reasoning.

DeepSeek V4 Flash is the smaller model of DeepSeek's April 2026 V4 generation: 284B total parameters, 13B active, a 1M-token context, and a toggleable thinking mode, refreshed with improved 0731 weights in July. Its compressed attention makes long contexts remarkably cheap, and its expert weights ship natively in FP4, so the full-quality build is only about 162GB. Running it at full quality takes a 256GB Mac Studio or a multi-GPU server with about 170GB of memory; 3-bit dynamic builds of roughly 103GB to 110GB fit 128GB machines with some quality loss. DeepSeek's September V4.1 Flash is larger (552B, including 196B of Engram memory), so V4 Flash remains the easier DeepSeek to run locally.

Qwen3.8 27B (27B) fits comfortably in the M5 Ultra 256GB's memory at Q8_0, using about 32GB with 8K context and generating ~22 tokens/sec. It is the next-best model that runs well here, and the one to try if the top pick doesn't suit you. It is also our top pick for working with images.

Qwen3.8 27B, released in August 2026, is one of the strongest open models that fits a single 24GB GPU, with image input, a 262K native context, and switchable thinking. It uses the same hybrid layout as the Qwen3.5 and 3.6 27B models, where three of every four layers use linear attention, so long contexts cost far less memory than on older dense models. At Q4_K_M it needs about 18GB with 8K context, leaving room for 64K or more of context on a 24GB card, and 32GB cards can run Q6_K. It supersedes Qwen3.6 27B and Qwen3 32B, but its long reasoning traces make speed on slower hardware a real consideration.

gpt-oss-20b (20.9B (3.6B active)) fits comfortably in the M5 Ultra 256GB's memory at MXFP4, using about 15GB with 8K context and generating ~105 tokens/sec. Pick it when responsiveness matters more than squeezing out maximum quality.

gpt-oss-20b is OpenAI's small open-weight reasoning model, a 21B mixture-of-experts with 3.6B active parameters whose expert weights ship natively in 4-bit MXFP4. OpenAI has not released a successor, and it remains one of the best reasoning and tool-calling models that fits a 16GB GPU. The native weights take about 14GB, so 16GB cards and Macs with 24GB or more run it fully in memory, and 12GB cards run it with a little CPU offload. Reasoning effort is adjustable (low, medium, high) but cannot be turned off.

GLM-5.3-Flash (320B (18B active)) fits comfortably in the M5 Ultra 256GB's memory at Q3_K_M, using about 161GB with 8K context and generating ~39 tokens/sec. It is the largest model that runs fully on this GPU.

GLM-5.3-Flash is Z.ai's August 2026 mid-size model and its first natively multimodal one: 320B total parameters with 18B active, a 1M-token context, a thinking mode, and an MIT license. It ranks just behind the largest open models on the Artificial Analysis Intelligence Index while needing well under half their memory. Most of its layers use linear attention with a fixed-size state, so long contexts are cheap. Expect about 200GB at Q4_K_M with 8K context, which means a 256GB Mac Studio or a multi-GPU server; the smallest 1-bit GGUF is 93GB and fits 128GB machines, at a clear quality cost.