Can You Run Mistral Large 4 Locally? Only on 512GB, and Not Until the Weights Ship

Last updated: October 2026

No, not yet, and when you can, it will take about 450GB of memory. Mistral Large 4 is a 1-trillion-parameter mixture-of-experts model with 49 billion parameters active per token, available today only as a paid API preview. When the weights land at the end of October, a single 512GB Mac Studio M5 Ultra is the smallest one-box machine that can run it, at 2-bit.

The open weights, not the API, are the part of this launch you can own. Here is the memory math, the machines that clear it, and how it compares with open models you can already run.

What Mistral Actually Announced, and What It Left Out

Mistral confirmed the size, the active parameters, and a month-end weights date. It did not publish a license, quantized formats, or hardware guidance. The October 6 announcement and the model documentation page disagree on two numbers, so both values appear below.

Sources: Mistral announcement and model docs, October 6, 2026; Artificial Analysis model page; TNW. The weights date comes from TNW's reporting, not from Mistral.

Mistral's own wording on timing is short: "We will release the weights by the end of the month." The company nicknamed the model Le Chonk, and pitches it for security operations that need to run on private cloud or on-premise. It has published no hardware guidance for doing that.

One open question: Mistral is red-teaming Large 4 with security firms, vetted partners, and state authorities, who get the same model with lighter moderation and expanded cyber capabilities. Mistral has not said whether the public weights will match that partner build.

The Memory Math: How Big Mistral Large 4 Will Be at Each Quantization

Expect about 400GB at 2-bit, about 470GB at 3-bit, about 605-650GB at 4-bit, about 1.1TB at 8-bit, and about 2.1TB at full BF16 precision. No quantized files exist yet, so these are estimates, built from two independent anchors.

• Anchor one: Unsloth's GGUF builds of Mistral Large 3, the 675B predecessor, scaled up 1.56x to Large 4's 1.05T.
• Anchor two: Unsloth's GGUF builds of Kimi K2, a 1T model, scaled up 5 percent.

At every quantization level, the two estimates land within 3 percent of each other. That agreement makes the table below trustworthy in shape, though not a guarantee of final file sizes.

Estimates scaled from Unsloth's published GGUF sizes for Mistral Large 3 (675B) and Kimi K2 (1T). Real files will differ. Long contexts add KV-cache memory on top of the figures shown.

Memory to load includes the operating system, compute buffers, and a working context. Mistral's docs list a 1M-token context, but no home machine in this table has room for one that long.

Which Machines Can Actually Run It

One 512GB Mac Studio M5 Ultra at 2-bit, two of them or a 768GB-plus EPYC server at 4-bit, and nothing smaller without accepting 1-bit quality.

One 512GB Mac Studio M5 Ultra: 2-bit, maybe 3-bit

Apple's Mac Studio spec page lists the M5 Ultra with up to 512GB of unified memory and 1.2TB/s of memory bandwidth. At launch, Apple said the 512GB configuration would arrive in late October, per Notebookcheck. That is the same month the weights are due.

There is a catch. By default, macOS lets the GPU use roughly three quarters of unified memory, about 384GB on a 512GB machine. That is less than the 2-bit build needs. You have to raise the limit, for example to 448GB with sudo sysctl iogpu.wired_limit_mb=458752, which leaves 64GB for macOS. The setting resets at reboot.

At 2-bit, each generated token reads about 19GB of active weights. At 1.2TB/s, that caps decode speed at roughly 64 tokens per second. Real throughput lands well below a bandwidth ceiling, and we will publish measured numbers once the weights and builds exist. Our M5 Ultra Mac Studio guide covers what else 512GB runs.

Two Macs, or a 768GB-Plus EPYC Server: 4-bit

The 4-bit build needs about 650-700GB, more than any single consumer machine holds. Two 512GB Mac Studios pool 1TB of memory through clustering tools such as exo or llama.cpp's RPC backend. Splitting a model across machines adds network overhead.

The other route is a single-socket AMD EPYC server with 12 memory channels. Twelve 64GB DDR5 modules give 768GB, and at DDR5-6400 the theoretical bandwidth is about 614GB/s. That works out to a decode ceiling of about 22 tokens per second at 4-bit. Adding one GPU helps: llama.cpp can keep the expert layers in system RAM and run the rest on the card. Memory is the expensive part of that build; our RAM price explainer covers why, and the local AI hardware guide covers build tiers.

What Does Not Work

• 128GB boxes such as a DGX Spark, a Ryzen AI Max+ 395 mini PC, or a 128GB Mac cannot hold even the 1-bit build.
• A 256GB Mac Studio falls short of the 1-bit build once macOS takes its share.
• A desktop with 256GB of DDR5 and an RTX 5090 totals 288GB, enough for the smallest 1-bit file with almost no context. Unsloth's own guidance for 1T models recommends 2-bit XL builds or higher for best results.
• An RTX 5090 alone has 32GB of VRAM, a small fraction of the smallest build.

The 49B-Active Tax

Large 4 reads 49 billion parameters for every token: about 1.5x Kimi K2's 32 billion, and about three times what DeepSeek V4.1-Flash (16B active) and GLM-5.3-Flash (18B active) read. Because home hardware is limited by memory bandwidth, Large 4 will generate tokens at a fraction of a Flash model's speed on the same machine.

Is It Worth It? Large 4 Against the Open Models You Can Already Run

For general use on home hardware, no. The preview scores 38 on the Artificial Analysis Intelligence Index, and five open-weight models score higher, two of which run on 128GB-256GB machines.

Scores from the Artificial Analysis Intelligence Index v4.3.2 as of October 6, 2026. Large 4 was scored as a preview through the API and may score differently once the weights ship. Home setups come from our own reality checks, linked below.

Mistral scoped its benchmark claim carefully. It says Large 4 rivals the best open models worldwide and clearly beats every open-weight model built in the US or Europe. On the Artificial Analysis open-weights leaderboard, that claim holds on its own terms. The best-scoring open model from those regions, Thinking Machines' Inkling, scores 25. But the claim also tells you Large 4 is not the top open model, and the table shows by how much.

Where Large 4 leads, it leads on Mistral's own numbers, which have not yet been independently reproduced:

• Security work: 93 percent of Cybench challenges solved, and 82 percent on vulnerability reproduction and patching, which Mistral calls the highest of any model.
• Visual grounding: 42 percent on Dense 200, against 41 percent for GPT-6-Astra.
• Languages: training data spanning 160-plus languages, including every official EU language.

If your work is security operations, multilingual document work, or visual grounding, Large 4 may earn its 450GB. For coding agents and general chat on hardware you already own, a Flash-class model gives you more per gigabyte.

GLM-5.3-Flash is the strongest open model that runs on a single 128GB box today. Its 2-bit build is about 109GB, which fits tightly on Linux; our GLM-5.3-Flash reality check has the file sizes and tested machines. Two 128GB DGX Sparks also run DeepSeek V4.1-Flash at measured speeds.

Check Price on Amazon: GMKtec EVO-X2 (128GB)

Check Price on Amazon: NVIDIA DGX Spark (128GB)

For a broader map of what fits each memory tier, see the best local AI models by VRAM and memory.

Why the Open Weights Matter More Than the API

Because the API and Mistral Cloud are still rented infrastructure. Only the weights let you set the policy, keep the model when the vendor changes course, and keep your data on hardware you control.

Mistral also sells hosting in its own European datacenters, run under European law. That changes your landlord's jurisdiction, not the fact that you have one: prompts still leave your network, and access still depends on an account.

Mistral makes the ownership argument itself. Its announcement argues that open weights let organizations run sensitive security work under their own rules instead of a provider's. That is the same argument we made after the June suspension of frontier model access: customers of the affected APIs lost access, while models already running on local hardware kept working.

Open weights are not a license, though. Mistral Large 3 shipped under Apache 2.0, which allows commercial use, according to its model card. Mistral has not said what Large 4 ships under. Read the LICENSE file before you build anything commercial on it.

What to Check the Day the Weights Drop

Seven checks, in order, before you download 400GB:

1. Source. Download only from Mistral's official Hugging Face organization. Treat any Large 4 repository that appears elsewhere around release day as suspect until it matches the official files.
2. License. Open the LICENSE file and check the commercial-use terms.
3. Architecture. Check config.json. If Large 4 reuses Large 3's architecture, which llama.cpp already supports, community GGUF builds should follow quickly. A new architecture means waiting for runtime support.
4. Official formats. Look for FP8 or NVFP4 checkpoints from Mistral. Large 3 shipped both.
5. Real file sizes. Compare Unsloth's actual GGUF sizes with the estimates above.
6. Ollama. Check whether the listing offers downloadable tags or only a cloud tag. The Mistral Large 3 listing offers only a cloud tag today, which means your prompts run on someone else's servers.
7. Independent score. Wait for Artificial Analysis to score the released weights, not the API preview.

We will update this page in place when the weights ship, with real file sizes, runtime support, and measured speeds. For how the same math played out on a larger model, see our Kimi K3 reality check.

Mistral Large 4 has no downloadable weights yet, so nothing can be graded today. Our LLM VRAM calculator lists it under announced models and grades every open model you can run now against your graphics card, Mac or mini PC.

Check what your hardware runs today

Frequently Asked Questions

Can I run Mistral Large 4 on an RTX 5090?

Not on the GPU alone. An RTX 5090 has 32GB of VRAM, and the smallest Mistral Large 4 build is estimated at 255-300GB. Paired with 256GB of system RAM, a 5090 desktop could load only a 1-bit build with almost no context. For usable quality, plan on a 512GB Mac Studio at 2-bit or a server with 768GB or more for 4-bit.

When will the Mistral Large 4 open weights be released?

Mistral says it will release the weights by the end of October 2026, and TNW reports a date of October 27. Until then, Mistral Large 4 is available only as a public preview through Mistral's API, under the model ID mistral-large-4. No official files exist on Mistral's Hugging Face organization yet.

Is Mistral Large 4 open source?

Mistral Large 4 is announced as open-weight, but Mistral has not published its license. Its predecessor, Mistral Large 3, shipped under Apache 2.0, which permits commercial use. Open weights let you download and run the model; the license decides what you may do with it. Check the LICENSE file on release day before commercial use.

How much RAM do I need to run Mistral Large 4?

Estimates put the 2-bit build at about 400GB, so plan on roughly 430-450GB of memory to load it. A 4-bit build needs about 650-700GB, and 8-bit about 1.2TB. These figures are scaled from the published GGUF sizes of Mistral Large 3 and Kimi K2, and long contexts add KV-cache memory on top.

Is Mistral Large 4 in Ollama?

Not as of October 6, 2026, and the weights have not shipped. When a listing appears, check whether it offers downloadable tags or only a cloud tag. Ollama lists Mistral Large 3 only as a cloud model today, which runs on remote servers rather than on your hardware. A cloud tag is not a local model.

Is Mistral Large 4 better than Kimi K3 or DeepSeek V4.1?

Not on aggregate benchmarks. On the Artificial Analysis Intelligence Index, the Large 4 preview scores 38, Kimi K3 (Max) scores 44, and DeepSeek V4.1-Flash (Max) scores 39. Large 4 leads on Mistral's own security and visual-grounding results, which have not been independently reproduced. For home hardware, the Flash models need far less memory per point.