What fits: the builds

Mistral Large 4 is a mixture-of-experts model with about 1 trillion parameters, 49 billion of them active per token (52 billion counting embeddings and output layers): 128 routed experts, four active per token, plus one shared expert. It takes text and image input, has a 1M-token context and offers three reasoning efforts, low, medium and high (the default), with no setting that turns reasoning off. Mistral’s model card reports 26 on the Artificial Analysis Intelligence Index v5, at launch the highest of any western open-weight model (Chinese open models score higher; Thinking Machines’ Inkling 2 Small has since scored 31), and 48% on Artificial Analysis’s Coding Agent Index, which it calls effectively tied with GLM-5.3.

Mistral Large 4 builds, smallest first. File sizes in decimal GB from the Hugging Face repos; bits per weight = size ÷ 1,050B parameters.
BuildRepoSizeBits/weightOne 512 GB Mac256 GB Mac
UD-IQ1_SUnsloth GGUF233.0 GB1.8YesYes — the only build that fits
IQ2_XXSggml-org GGUF297.7 GB2.3YesNo
2-bit MLXmlx-community (text only)338.1 GB2.6YesNo
Q2_Kggml-org GGUF380.7 GB2.9Yes — the build Hugging Face ranNo
UD-IQ3_XXSUnsloth GGUF423.5 GB3.2Yes, with the GPU memory limit raisedNo
3-bit MLXmlx-community (text only)459.3 GB3.5Yes, tightNo
UD-Q3_K_XLUnsloth GGUF492.7 GB3.8Very tightNo
4-bit MLXmlx-community (text only)590.4 GB4.5No — two MacsNo
UD-Q4_K_XLUnsloth GGUF604.8 GB4.6NoNo
Q4_K_Mggml-org GGUF634.8 GB4.8NoNo

Apple’s 512 GB is 512 GiB, about 550 decimal GB, which is why a 493 GB file can still load. The index’s methodology assumes macOS lets the GPU use about three-quarters of memory by default, so the 3-bit builds need the GPU memory limit raised (the iogpu.wired_limit_mb setting; how much RAM you need explains it), and every build needs room left for macOS and the context.

The index lists the model’s minimum as 383 GB: the Q2_K file plus about 2 GB. Its usual rule takes the smallest build of at least 3 bits, which here is UD-IQ3_XXS at 424 GB, but a 2-bit build sets the minimum where a published Mac run used one, and Hugging Face’s did.

How fast it runs

MacBuild and runtimeDecodeSource
Mac Studio M3 Ultra, 512 GBQ2_K, llama.cpp (Llama App)sourced15 tok/sHugging Face launch post, 29 October; prompt length not stated
Mac Studio M5 Ultra, 512 GBThe 383 GB build, costed at 3.0 bits per weightestimated~21 tok/sThe index’s estimate, scaled from the M3 Ultra run (model page)
Provenance key
sourced
Published figure — vendor or press review
estimated
community
Linked individual run — blog, forum or submission

No one has published a 512 GB M5 Ultra run yet, so the M5 Ultra figure is an estimate. It rests on bandwidth: the M5 Ultra reads memory 1.5× as fast as the M3 Ultra (1,228 against 819 GB/s), and decode on these chips tracks that ratio (M5 Ultra vs M3 Ultra). At Q2_K each token reads about 18 GB of active weights (49B parameters at 2.9 bits), which caps the M3 Ultra near 46 tok/s; Hugging Face’s 15 is about a third of that ceiling.

Prompt processing is the slow half. Hugging Face measured about 90 tok/s of prefill on the M3 Ultra, so a 10,000-token prompt waits close to two minutes for its first token. One community run on a 256 GB M5 Ultra, with the 1.8-bit build, processed a 32K-token prompt at about 184 tok/s, which is about three minutes.

Three more runs have been reported. The index does not list them, for the reasons given:

Which Mac runtimes load it

The licence

Mistral Large 4 ships under a new Mistral Ecosystem Open-Weight License, a modified MIT licence. It grants the usual MIT rights with an attribution notice and adds one condition: a licensee, or any of its affiliates, that operates a Model-as-a-Service business must sign a separate commercial licence with Mistral before using the model commercially. Model-as-a-Service means giving third parties inference or post-training access, by API or hosted endpoint for instance, with meaningful control over the inputs, parameters or training data. Internal use is exempt, and so are end-user products that embed the model in specific features and services that only relay requests to models hosted elsewhere.

Running it on your own Mac is internal use. The index scores the licence 5 of 10 points, the same as other licences that need a separate agreement for API hosting, such as the Muse Model License and the StepFun Model License (how licences are scored). Mistral Large 3 was Apache 2.0, which scores 8.

Where it ranks

In the LLMCheck index, Mistral Large 4 scores 58 and ranks #92 of the 130 models: 45 of 50 for capability, 5 of 25 for speed (from the estimate above), 3 of 15 for accessibility (the tier for models that need more than 64 GB) and 5 of 10 for the licence. Speed is the term a measured M5 Ultra run would move.

512 GB or 256 GB

Apple sells the M5 Ultra Mac Studio with 256 GB for $9,499 (30-core CPU, 64-core GPU) or $10,799 (36-core CPU, 80-core GPU), and with 512 GB for $22,299, only with the 80-core GPU and a 1 TB SSD (Apple US list prices, checked 14 November 2026). The step from 256 to 512 GB costs $11,500 on the 80-core model, or $12,800 from the 64-core one.

A 256 GB Mac Studio already runs the index’s pick for every Mac from 192 GB up, MiMo-V2.6-Flash, at community77 tok/s (community). It is tied at Score 81 with Qwen 3.6-35B-A3B for the index’s #1, listed first on the alphabetical tie-break; Qwen needs only 24 GB and is the best pick at 32 GB or less. What 512 GB adds are the models that need more than 256 GB, or need it for a 4-bit build:

Models a 512 GB Mac Studio adds over 256 GB, in LLMCheck Score order: every model with a Mac runtime that needs more than 256 GB and scores at least Mistral Large 4’s 58, plus three whose 4-bit builds need 512 GB. “Needs” is the index’s minimum RAM.
ModelScore (rank)NeedsLicenceIndex, M5 Ultra
DeepSeek V4.1 Flash70 (#18)434 GBMITestimated~29 tok/s
GLM 5.2 at 4-bit69 (#27)256 GB at 2-bitMITestimated~22 tok/s
Inkling 2 Small at 4-bit68 (#32)228 GB at ~3-bitApache 2.0estimated~36 tok/s
Kimi K2.667 (#39)429 GBModified MITestimated~27 tok/s
Kimi K2.7 Code67 (#40)421 GBModified MITestimated~27 tok/s
Ling-3.1-flash67 (#41)360 GBMITestimated~27 tok/s
GLM-5.3 at 4-bit66 (#44)256 GB at 2-bitGLM-5.3estimated~17 tok/s
Kimi K2.566 (#46)417 GBModified MITestimated~28 tok/s
GLM-5.165 (#50)390 GBMITestimated~17 tok/s
Muse Spark Open65 (#53)342 GBMuse Model Licenseestimated~31 tok/s
Hy4 preview63 (#65)292 GBApache 2.0estimated~14 tok/s
DeepSeek V3.2-Speciale62 (#71)360 GBMITestimated~18 tok/s
DeepSeek V3.259 (#84)360 GBMITestimated~18 tok/s
Mistral Large 458 (#92)383 GBMistral Ecosystem Open-Weightestimated~21 tok/s

Every model in the table decodes more slowly than MiMo-V2.6-Flash and scores lower, so the index’s pick for a 512 GB Mac stays MiMo-V2.6-Flash. What the extra memory buys is choice: 1T-class models on one machine, 4-bit builds of the GLM models, and room to keep two large models loaded at once. A 256 GB Mac can load Mistral Large 4 only as the 1.8-bit UD-IQ1_S, below the 2–3-bit floor the index uses to set a model’s minimum memory. The 512 GB M5 Ultra page ranks everything that fits.

64-core or 80-core GPU

M5 Ultra30-core CPU, 64-core GPU36-core CPU, 80-core GPU
Memory bandwidth1.2 TB/s1.2 TB/s
96 GB$5,499$6,799
256 GB$9,499$10,799
512 GBNot offered$22,299

Sources: Apple’s Mac Studio specs and store configuration pages, US list prices with a 1 TB SSD, checked 14 November 2026.

Decode reads weights from memory, and both GPUs read at the same 1.2 TB/s, so the index gives them the same figures. It does not treat the 64-core part as binned, because binned in the index means a lower-bandwidth configuration. The community runs agree: MiMo-V2.6-Flash measured 75.5–88.6 tok/s on 64-core machines and 61.0–89.4 on 80-core ones (oMLX, 4-bit or MXFP4, multi-token prediction off, 1K–8K prompts), so the spread between sessions is wider than any gap between the two GPUs. The extra 16 GPU cores should help most in prompt processing, which is compute-bound, but the index has no matched pair that isolates the gain. The $1,300 matters most if you want 512 GB, which Apple sells only with the 80-core GPU.

4-bit needs two Macs

The smallest 4-bit builds are 590 GB (MLX) and 605 GB (Unsloth UD-Q4_K_XL), more than one 512 GB Mac holds, so 4-bit Mistral Large 4 means two Mac Studios linked over Thunderbolt 5: $44,598 for two 512 GB M5 Ultras at list price. mlx-vlm 0.7.12 runs it with distributed inference, and its maintainer showed it on two M3 Ultras at about 20 tok/s with a batch of three requests (2 November; the post does not say whether that is per request or in total, or which quantization it used). It is a two-machine, batched figure, so the index does not list it.

MLX’s RDMA backend over Thunderbolt 5 needs macOS 26.2 and MLX 0.30.1 or later. The two bugs that hung or crashed a run when a peer dropped out (#3910, #4192) were closed by a fix merged on 30 October (#4767), which is not yet in an MLX release; the latest, 0.32.4, dates from 14 October.

Who should buy 512 GB

Buy 512 GB if you need a specific 1T-class open model on one machine: Mistral Large 4, which Mistral pitches on European languages and visual grounding, the Kimi K2 line for agentic coding, DeepSeek V4.1 Flash, or the GLM models at 4-bit. The index’s figures for those models run from about 17 to 29 tok/s (estimated), against 77 for MiMo-V2.6-Flash.

For everyone else, a 256 GB Mac Studio with the 64-core GPU ($9,499) runs the index’s pick at the same decode speed as the 80-core model and costs $12,800 less than the 512 GB one. A used 512 GB M3 Ultra is the machine Hugging Face measured. Apple stopped offering that memory option in March 2026, and its 819 GB/s gives about two-thirds of the M5 Ultra’s decode speed.

Frequently Asked Questions

Can a Mac run Mistral Large 4?

Yes, but only a 512 GB Mac Studio holds it on one machine, at 2 to 3 bits: the Q2_K build is 381 GB and Unsloth's UD-IQ3_XXS is 424 GB. Hugging Face published 15 tok/s with Q2_K in llama.cpp on a 512 GB M3 Ultra. Every 4-bit build is 590 GB or more and needs two Macs.

How fast is Mistral Large 4 on the 512 GB M5 Ultra?

No run has been published yet. According to the LLMCheck index, it runs at about 21 tok/s (estimated), scaled from Hugging Face's published 15 tok/s on an M3 Ultra with the index's bandwidth model (1.5x the M3 Ultra's bandwidth, with the build costed at 3.0 bits per weight). Prompt processing is slower: about 90 tok/s on the M3 Ultra, so a 10,000-token prompt takes close to two minutes to start.

Does Mistral Large 4 fit a 256 GB Mac Studio?

Only as Unsloth's 1.8-bit UD-IQ1_S build (233 GB). A community run on a 256 GB M5 Ultra measured 18.4 tok/s at a 32K-token context with it, but that is below the 2 to 3 bits the index uses to set a model's minimum memory, so the index lists Mistral Large 4 as needing 383 GB.

How much does the 512 GB Mac Studio M5 Ultra cost?

$22,299 in the US, with a 1 TB SSD and the 36-core CPU and 80-core GPU, the only chip Apple offers with 512 GB. Apple has taken orders since 28 October 2026. The 256 GB model costs $9,499 with the 64-core GPU or $10,799 with the 80-core GPU (Apple list prices, checked 14 November 2026).

Is the 64-core M5 Ultra slower than the 80-core for local LLMs?

Not for decode. Both have 1.2 TB/s of memory bandwidth, which sets decode speed, and community runs of MiMo-V2.6-Flash measured 75.5 to 88.6 tok/s on 64-core machines against 61.0 to 89.4 on 80-core ones. The extra GPU cores should help prompt processing. Only the 80-core model can be ordered with 512 GB.

Can I use Mistral Large 4 commercially?

Yes, under the Mistral Ecosystem Open-Weight License, a modified MIT licence, with one exception: a company that operates a Model-as-a-Service business, giving third parties API or hosted access with meaningful control over inputs, parameters or training data, needs a separate commercial licence from Mistral. Internal use and products that embed the model in specific features are exempt.

What is the best local LLM for a 512 GB Mac Studio?

According to the LLMCheck index, MiMo-V2.6-Flash, which already fits 192 GB and runs at 77 tok/s community-measured on an M5 Ultra. It is tied at Score 81 with Qwen 3.6-35B-A3B for the index's #1, listed first on the alphabetical tie-break; Qwen needs only 24 GB and is the best pick at 32 GB or less. 512 GB adds 1T-class models such as DeepSeek V4.1 Flash (Score 70), the Kimi K2 line and Mistral Large 4 (58).