What can a Mac Studio M3 Ultra (256 GB) run locally?

Only on an experimental path, in RigCheck's quick setup: IQ3_S; No measured speed for this hardware and setup yet; Measured on one Mac (M5 Max 128 GB). A chip with less memory bandwidth is slower. Apple Silicon (M1 or newer, macOS 15 or newer) runs Strata's Metal engine, ready-made in the Mac app (.dmg). Strata calls its Mac status new: measured on three Macs, other chips untested; RigCheck counts it as experimental. The full checker takes your exact setup (RAM, disk, operating system, processor) and gives the install steps.

Yes, by RigCheck's capacity estimate, in RigCheck's quick setup: Q8_0 · unified memory; No measured speed for this hardware and setup yet; On a Mac, raise the GPU wired-memory limit first; the full checker gives the command; RigCheck capacity estimate, not a measured setup. The full checker takes your exact setup (context length, RAM, image input) and gives the command to start it.

Yes, by RigCheck's capacity estimate, in RigCheck's quick setup: Q8_0 · unified memory; No measured speed for this hardware and setup yet; On a Mac, raise the GPU wired-memory limit first; the full checker gives the command; RigCheck capacity estimate, not a measured setup. The full checker takes your exact setup (context length, RAM, image input) and gives the command to start it.

Yes, by RigCheck's capacity estimate, in RigCheck's quick setup: BF16 · unified memory; No measured speed for this hardware and setup yet; On a Mac, raise the GPU wired-memory limit first; the full checker gives the command; RigCheck capacity estimate, not a measured setup. The full checker takes your exact setup (context length, RAM, image input) and gives the command to start it.

Yes, by RigCheck's capacity estimate, in RigCheck's quick setup: Gemma 4 31B: fits (capacity estimate) (BF16, unified memory); Gemma 4 26B-A4B: fits (capacity estimate) (BF16, unified memory); Gemma 4 12B: fits (capacity estimate) (BF16, unified memory); Gemma 4 E4B: fits (capacity estimate) (BF16, unified memory); Gemma 4 E2B: fits (capacity estimate) (BF16, unified memory); No measured speed for this hardware and setup yet; On a Mac, raise the GPU wired-memory limit first; the full checker gives the command; RigCheck capacity estimate, not a measured setup. The full checker takes your exact setup (context length, RAM, image input) and gives the command to start it.

Only as a planning estimate: the checker's uncalibrated memory budget, offload peak unverified, in RigCheck's quick setup: Mac GGUF is within the planning budget; Metal execution unverified here; Uncalibrated allowances, not a minimum or a guarantee. Full checkpoint files are counted in host RAM even when mmap or sequential loading can use less. Actual peak depends on attention, prefix caches, VAE tiling, reference sizes and backend. No speed prediction. The full checker lets you change the task, resolution, references and offload (batch is fixed at one), and lists the files.

A published route or recipe matches, with caveats, in RigCheck's quick setup: Nemotron 3.5 Bolt 74B-A9B: a published route or recipe matches, with caveats (Matches Unsloth's short-context budget: about 32 GB of RAM or unified memory); Nemotron 3.5 Super VL 120B-A12B: a published route or recipe matches, with caveats (Matches Unsloth's 192 GB budget: Q8_0 (8-bit, 132 GB)); On a Mac, raise the GPU wired-memory limit first; the full checker gives the command; Nemotron 3.5 Bolt 74B-A9B: NVIDIA names no Mac setup; Unsloth's llama.cpp guide covers Metal builds. Unsloth budgets around 32 GB of RAM or unified memory for short-context use (48–64 GB for more headroom) with its 22 GB IQ2_XXS GGUF in llama.cpp; longer contexts, images and video need more. This unified memory is 256 GB. That is Unsloth's planning estimate, not a measured run on this hardware. On a Mac, raise the GPU wired-memory limit first. Nemotron 3.5 Super VL 120B-A12B: Unsloth's planning estimate for a modest context, including the OS, runtime, KV cache and vision file; not a measured run. This setup counts 256 GB of unified memory (counted once). Longer contexts and video need more. On a Mac, raise the GPU wired-memory limit first. The full checker lets you change the route or recipe settings and lists the files and the published recipe.

A published route or recipe matches, with caveats, in RigCheck's quick setup: Matches Unsloth's 4-bit figure (210 GB): UD-Q4_K_XL (199.7 GB); On a Mac, raise the GPU wired-memory limit first; the full checker gives the command; Unsloth's total-memory table (RAM + VRAM, or unified memory) for its GGUF files with llama.cpp; it does not say at what context length, and which file of a bit width fits at 210 GB is our reading. This setup counts 256 GB of unified memory. Image input adds the 1.2 GB vision file. Use Unsloth's llama.cpp branch (glm5next/upstream) or Unsloth Desktop, as Unsloth's guide says. No published speed for this hardware. On a Mac, raise the GPU wired-memory limit first. The full checker lets you change the route or recipe settings and lists the files and the published recipe.

A published route or recipe matches, with caveats, in RigCheck's quick setup: DeepSeek V4.1 Flash September release: a published route or recipe matches, with caveats (Matches DwarfStar: Q2 main weights fully resident on a 256 GB Mac (September weights)); DeepSeek V4.1 Flash 1017 update: experimental path (Experimental: DwarfStar's September setup applied to its 1017 file: Q2 main weights fully resident on a 256 GB Mac (1017 weights)); DeepSeek V4.1 Flash September release: DwarfStar says a 256 GB or larger Mac can hold all main weights of its Q2 file (151.77 GiB of the 365.7 GB file). Its Q4 file needs SSD streaming below 512 GB. DwarfStar publishes no V4.1 speed for this setup (its one Mac figure is Q4 resident on a 512 GB M3 Ultra). Image input works on Metal with its separate 0.97 GB vision encoder. DwarfStar is a community engine, not DeepSeek's code; these are its documented setups, not a RigCheck estimate. Keep the GGUF on a fast local SSD (the 188.83 GiB of Engram tables are always read from the file). DeepSeek V4.1 Flash 1017 update: DwarfStar says a 256 GB or larger Mac can hold all main weights of its Q2 file (151.77 GiB of the 365.7 GB file). Its Q4 file needs SSD streaming below 512 GB and is September only. DwarfStar publishes no V4.1 speed for this setup (its one Mac figure is Q4 resident on a 512 GB M3 Ultra, September weights). Experimental for 1017: Unsloth says its 3-bit UD-IQ3_XXS (394.7 GB) works on 256 GB RAM setups such as a Mac Studio, the bottom of its 256–290 GB estimate for 3-bit, in its unmerged llama.cpp pull request #30817 branch or Unsloth Desktop. DwarfStar documents and tests these setups with its September file only. On October 19, 2026 it published DeepSeek-V4.1-Flash-1017-IQ2_XXS-Q2_K-imatrix.gguf (365.7 GB, the size of its September Q2; its name gives the same IQ2_XXS / Q2_K mix) and pointed its Q2 download target at it, whose help text names SSD streaming on one 128 GB Mac, tensor parallelism on two, or a larger resident Mac. RigCheck applies the September setup to it. No 1017 speed or 1017 Q4 file is published; its 1017 MXFP4-experts GGUF (501.6 GB) has no documented setup. DwarfStar's vision encoder is from the September checkpoint; it does not say whether it works with the 1017 file. DwarfStar is a community engine, not DeepSeek's code; these setups are documented for its September file, not a RigCheck estimate. Keep the GGUF on a fast local SSD (the 188.83 GiB of Engram tables are always read from the file). The full checker lets you change the route or recipe settings and lists the files and the published recipe.