Mistral AIMistralMinistral3ForCausalLM

Devstral 2 123B Instruct 2512 — Hardware Requirements & GPU Compatibility

ChatFunctions

Devstral 2 123B Instruct 2512 is Mistral AI's large agentic coding model for software engineering tasks such as exploring codebases, editing multiple files, and running autonomous coding agents. Released as an FP8-quantized checkpoint of a roughly 125-billion-parameter dense model, it is a direct step up from the smaller Devstral Small line and pairs with Mistral's own Vibe CLI as well as third-party scaffolds like OpenHands and Claude Code. It scores 72.2% on SWE-Bench Verified, 61.3% on SWE-Bench Multilingual, and 32.6% on Terminal-Bench 2, competitive with or ahead of several much larger open models. At this size it needs a multi-GPU workstation even when quantized. Context length is 262,144 tokens. It is released under a Modified MIT License that blocks companies with over $20 million in monthly consolidated revenue from using it without a separate commercial license from Mistral AI. It was published in November 2025.

125.0B params · dense262K contextOtherNov 20256.5K downloads360 likes25.5K quant downloads

Your hardware

Verdicts below are for the reference RTX 4090. Set yours to personalise the page.

Detection reads your browser's GPU name with no data sent anywhere.

Devstral 2 123B Instruct 2512 runs on a 96 GB+ Mac or a workstation GPU

Best local pick → Mac Studio (M3 Ultra, 96GB) · Q4_K_M · 81.2 GB · ~8 tok/s

Macs from 48 GB (IQ2_XS · ~11 tok/s)

Other

Not set up for it?

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run Devstral 2 123B Instruct 2512

One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.

Ways to runOther ways

each opens step-by-step

Run in cloud

Runs on B200 (192 GB) (110 GB headroom) · Q4_K_M

Generation speed: ~68 tok/s; generation speed
Cost per 1M output tokens: $24.29; per 1M output tokens

Deployon RunPodBest valuefrom $5.98/hr

$5+ bonus for new users

Compare GPUs →

Run Devstral 2 123B Instruct 2512

81.2 GB VRAM at 16K context

Works on macOS, Windows and Linux.

Install Ollama (Linux)
$curl -fsSL https://ollama.com/install.sh | sh
Install Ollama (macOS / Windows)
Download and run the app; it installs the ollama command.
Download Ollama ↗
Run Q4_K_M
$ollama run devstral-2:123b
Official Ollama library build. For the exact Q4_K_M file from unsloth/Devstral-2-123B-Instruct-2512-GGUF: ollama run hf.co/unsloth/Devstral-2-123B-Instruct-2512-GGUF:Q4_K_M
Raise the context window
Type it inside the chat. For the API server instead, start it with OLLAMA_CONTEXT_LENGTH=16384 ollama serve.
$/set parameter num_ctx 16384
A larger context needs more memory than the weights alone.
Use the API
Ollama serves its API on http://localhost:11434.

Works on macOS, Windows and Linux.

Install LM Studio
Desktop app for macOS (Apple silicon), Windows and Linux.
lmstudio.ai ↗
Download the model
In the app press Ctrl/Cmd + Shift + M and search mistralai/devstral-2-2512, then pick Q4_K_M. Or use the CLI (run LM Studio once first):
$lms get mistralai/devstral-2-2512@q4_k_m
Open on Hugging Face ↗
Load and chat
Open the model loader, select the model and set Context Length to 16384. Leave GPU offload at maximum if it fits.
Start the local server (optional)
$lms server start

Works on macOS, Windows and Linux.

Install llama.cpp (macOS)
MacPorts, Nix and conda-forge also work.
$brew install llama.cpp
Install options ↗
Install llama.cpp (Windows)
$winget install llama.cpp
Install llama.cpp (Linux)
Download the prebuilt llama-*-bin-ubuntu-* archive for your GPU backend (CPU, Vulkan) from the releases page, unpack it and run the binaries from that folder. Homebrew (brew install llama.cpp) and conda-forge also work on Linux.
llama.cpp releases ↗
Start the server
Downloads the model on first run. The chat UI and an OpenAI-compatible API are at http://localhost:8080.
$llama-server -hf unsloth/Devstral-2-123B-Instruct-2512-GGUF:Q4_K_M -c 16384 -ngl 99 --port 8080

Works on macOS, Windows and Linux.

Option A: Docker Model Runner
Docker Desktop: Settings > AI > Enable Docker Model Runner. Docker Engine on Linux: install the docker-model-plugin package.
$docker model run hf.co/unsloth/Devstral-2-123B-Instruct-2512-GGUF:Q4_K_M
Model Runner docs ↗
The quant tag after the repo is how Hugging Face cards write it; if your Docker version rejects it, drop the tag and pick the quant by repo.
Option B: llama.cpp server image (NVIDIA)
Needs the NVIDIA Container Toolkit. Models are cached in the llama-cache volume. Open http://localhost:8080.
$docker run --gpus all -p 8080:8080 -v llama-cache:/root/.cache ghcr.io/ggml-org/llama.cpp:server-cuda -hf unsloth/Devstral-2-123B-Instruct-2512-GGUF:Q4_K_M -c 16384 -ngl 99 --host 0.0.0.0 --port 8080
Image tags (cuda13, vulkan, rocm) ↗
Cache path inside the image is not in the docs; remove the -v flag if the volume complains.

Linux only.

Install vLLM
Linux only; on Windows run it inside WSL2. NVIDIA GPU shown; AMD ROCm, Intel and CPU builds have their own install pages.
$uv venv --python 3.12 --seed
$source .venv/bin/activate
$uv pip install vllm --torch-backend=auto
vLLM installation ↗
Serve the model
OpenAI-compatible API at http://localhost:8000/v1.
$vllm serve mistralai/Devstral-2-123B-Instruct-2512 --max-model-len 16384
Runs the original weights, not the Q4_K_M GGUF; needs ~275 GB of GPU memory.

macOS only.

MLX supports only macOS, so follow these steps for that setup.

Install mlx-lm
$pip install mlx-lm
Chat in the terminal
$mlx_lm.chat --model mlx-community/Devstral-2-123B-Instruct-2512-4bit
Downloads the model on first run. The MLX repo has its own quantisation; it may not match the quant selected above.
Or start a local server
Listens on http://localhost:8080.
$mlx_lm.server --model mlx-community/Devstral-2-123B-Instruct-2512-4bit

Works on macOS, Windows and Linux.

Install Jan
jan.ai ↗
Find the model
Open the Hub tab, paste this model ID in Search, and choose Q4_K_M if offered.
$unsloth/Devstral-2-123B-Instruct-2512-GGUF
Use the local API (optional)
Settings > Local API Server > Start Server. It listens on http://127.0.0.1:1337.

Rent GPUs hourly to run Q4_K_M right there. Pricing comes from partner marketplaces and varies.

Available GPU rentals for Q4_K_M, prioritized by value
GPU	Speed	$/1M tokens	Rent
B200 (192 GB)CheapestCheapest110 GB free	~68 tok/s	$24.29	Deploy on RunPodfrom $5.98/hr (opens in new tab)	192 GB · 110 GB free · ~68 tok/s · $24.29 / 1M tokens
H200 SXM (141 GB)59 GB free	~41 tok/s	$24.32	Deploy on RunPodfrom $3.59/hr (opens in new tab)	141 GB · 59 GB free · ~41 tok/s · $24.32 / 1M tokens
H200 NVL (141 GB)59 GB free	~41 tok/s	$25.55	Deploy on Vast.aifrom $3.77/hr (opens in new tab)	141 GB · 59 GB free · ~41 tok/s · $25.55 / 1M tokens
H200 SXM (141 GB)59 GB free	~41 tok/s	$33.94	Deploy on Vast.aifrom $5.01/hr (opens in new tab)	141 GB · 59 GB free · ~41 tok/s · $33.94 / 1M tokens
B300 (288 GB)206 GB free	~68 tok/s	$28.18	Deploy on RunPodfrom $6.94/hr (opens in new tab)	288 GB · 206 GB free · ~68 tok/s · $28.18 / 1M tokens
B200 (192 GB)110 GB free	~68 tok/s	$30.47	Deploy on Vast.aifrom $7.50/hr (opens in new tab)	192 GB · 110 GB free · ~68 tok/s · $30.47 / 1M tokens

Compare GPUs →

How Much VRAM Does Devstral 2 123B Instruct 2512 Need?

Select a quantization level; the listed hardware adheres to it.

VRAM	+32K	+128K	File	Quality	On RTX 4090 (reference)
35.4 GB	46.5 GB	81.9 GB	34.4 GB	2-bit · i-quant extreme compression — Importance-weighted 2-bit, extreme compression — significant quality loss	Does not fit: needs 40.6 GB
38.5 GB	49.6 GB	85.1 GB	37.3 GB	2-bit · i-quant extra small — Importance-weighted 2-bit, extra small	Does not fit: needs 43.7 GB
40.1 GB	51.2 GB	86.6 GB	39.7 GB	2-bit · i-quant small — Importance-weighted 2-bit, small	Does not fit: needs 45.3 GB
43.2 GB	54.3 GB	89.7 GB	43.5 GB	2-bit · i-quant medium — Importance-weighted 2-bit, medium	Does not fit: needs 48.4 GB
49.5 GB	60.6 GB	96 GB	49 GB	3-bit · i-quant — Importance-weighted 3-bit	Does not fit: needs 54.7 GB
52.6 GB	63.7 GB	99.1 GB	51.7 GB	3-bit · i-quant extra small — Importance-weighted 3-bit, extra small	Does not fit: needs 57.8 GB
54.2 GB	65.2 GB	100.7 GB	46.6 GB	2-bit · K-quant — 2-bit quantization with K-quant improvements	Does not fit: needs 59.3 GB
55.7 GB	66.8 GB	102.3 GB	54.4 GB	3-bit small — 3-bit small quantization	Does not fit: needs 60.9 GB
57.3 GB	68.4 GB	103.8 GB	56.8 GB	3-bit · i-quant medium — Importance-weighted 3-bit, medium	Does not fit: needs 62.5 GB
62.0 GB	73.1 GB	108.5 GB	60.6 GB	3-bit medium — 3-bit medium quantization	Does not fit: needs 67.2 GB
63.5 GB	74.6 GB	110.1 GB	71 GB	4-bit · legacy — 4-bit legacy quantization	Does not fit: needs 68.7 GB
65.1 GB	76.2 GB	111.6 GB	66.1 GB	3-bit large — 3-bit large quantization	Does not fit: needs 70.3 GB
68.2 GB	79.3 GB	114.8 GB	67.1 GB	4-bit · i-quant compact — Importance-weighted 4-bit, compact	Does not fit: needs 73.4 GB
71.4 GB	82.4 GB	117.9 GB	70.9 GB	4-bit · i-quant non-linear — Importance-weighted 4-bit, non-linear	Does not fit: needs 76.5 GB
71.4 GB	82.4 GB	117.9 GB	78.5 GB	4-bit · legacy+ — 4-bit legacy quantization with offset	Does not fit: needs 76.5 GB
71.4 GB	82.4 GB	117.9 GB	71.2 GB	4-bit small — 4-bit small quantization	Does not fit: needs 76.5 GB
76.0 GB	87.1 GB	122.6 GB	74.9 GB	4-bit medium — 4-bit medium quantization — most popular sweet spot	Does not fit: needs 81.2 GB
77.6 GB	88.7 GB	124.1 GB	76.1 GB	4-bit large — 4-bit large quantization	Does not fit: needs 82.8 GB
87.0 GB	98.1 GB	133.5 GB	86.2 GB	5-bit small — 5-bit small quantization	Does not fit: needs 92.2 GB
90.1 GB	101.2 GB	136.6 GB	88.3 GB	5-bit medium — 5-bit medium quantization — good quality/size tradeoff	Does not fit: needs 95.3 GB
104.2 GB	115.3 GB	150.7 GB	102.6 GB	6-bit, very good — 6-bit quantization, very good quality	Does not fit: needs 109.4 GB
126.1 GB	137.1 GB	172.6 GB	132.9 GB	8-bit, near-lossless — 8-bit quantization, near-lossless	Does not fit: needs 131.2 GB
251.1 GB	262.2 GB	297.6 GB	250.1 GB	16-bit · bfloat — Brain floating point 16 — preferred for training	Does not fit: needs 256.3 GB

Selected Q4_K_M

4-bit medium quantization

81.2GBat 16K context

Weights: 75.0 GB
KV cache: 5.9 GB
Overhead: 0.3 GB

Doesn't fit RTX 4090 (reference) · needs 57.2 GB more

Or .

Which Hardware Can Run Devstral 2 123B Instruct 2512?

Q4_K_M · 76.0 GB

Devstral 2 123B Instruct 2512 (Q4_K_M) requires 76.0 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 99+ GB is recommended. Using the full 262K context window can add up to 93.8 GB, bringing total usage to 169.8 GB. No consumer GPU has enough memory. 30 devices with unified memory can run Devstral 2 123B Instruct 2512, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Studio (M3 Ultra, 96GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M

Typical hardwareQuantVRAMtok/sContext that fits

8–16 GB · RTX 4060, RTX 3060 12GB, RTX 5060 Ti 16GBDoesn't fit — smallest quant needs 40.6 GB8 GB tier →12 GB tier →16 GB tier →

24 GB · RTX 4090referenceDoesn't fit — smallest quant needs 40.6 GB24 GB tier →

Mac 32 GB · Mac Mini M4Doesn't fit — smallest quant needs 40.6 GBMacs →

Mac 64 GB · Mac Studio M4 MaxQ2_K · 59.3 GB · ~7 tok/s · 16K59.3~716KMacs →

How good is Devstral 2 123B Instruct 2512?

llmrun Score blends public benchmark boards into one number; 100 = the average open model.

Not enough data yet for an llmrun Score.

A Score needs results on enough independent boards. How the Composite Score is built

Benchmarks

SWE-bench Verified
SWE-bench · Dec 2025Independent
53.8%
#15/20 open
Independent
LMArena WebDev
LMArena · Oct 2026Arena
1197.1
#54/55 open
Arena

All 2 Devstral 2 123B Instruct 2512 benchmarks

Alternatives to Devstral 2 123B Instruct 2512

Mistral family · 35 models →

Similar models

6tok/s on RTX 4090

Model	Score	VRAM	tok/s	Why it is similar
Qwen3.5 122B A10BAlibaba	109.4	75.5 GB	—needs > 24 GB	same size
GPT OSS 120BOpenAI	97.5	70.8 GB	—needs > 24 GB	same size
Ornith 1.5 122Bornith-ai	—	75.5 GB	—needs > 24 GB	same size
NVIDIA Nemotron 3.5 Super VL 120B A12B BF16NVIDIA	—	87.7 GB	—needs > 24 GB	same size
Ling 3.0 Flash VLInclusion AI	—	76 GB	—needs > 24 GB	same size
Laguna S 2.1poolside	—	72.5 GB	—needs > 24 GB	same size

Specifications

Publisher: Mistral AI
Family: Mistral
Parameters: 125.0B
Architecture: Ministral3ForCausalLM
Context Length: 262,144 tokens
Vocabulary Size: 131,072
Release Date: 2025-11-28
License: Other

Resources

Frequently Asked Questions

How much VRAM does Devstral 2 123B Instruct 2512 need?

Devstral 2 123B Instruct 2512 requires 76.0 GB of VRAM at Q4_K_M, or 251.1 GB at BF16. Full 262K context adds up to 93.8 GB (169.8 GB total).

VRAM = Weights + KV Cache + Overhead

Weights = 125.0B × 4.8 bits ÷ 8 = 75 GB

KV Cache + Overhead ≈ 1 GB (at 2K context + ~0.3 GB framework)

Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.

KV Cache + Overhead ≈ 94.8 GB (at full 262K context)

VRAM usage by quantization

Q4_K_M

76.0 GB

Q4_K_M + full context

169.8 GB

Learn more about VRAM estimation →

Can NVIDIA GeForce RTX 5090 run Devstral 2 123B Instruct 2512?

No — Devstral 2 123B Instruct 2512 requires at least 35.4 GB at IQ2_XXS, which exceeds the NVIDIA GeForce RTX 5090's 32 GB of VRAM.

What's the best quantization for Devstral 2 123B Instruct 2512?

For Devstral 2 123B Instruct 2512, Q4_K_M (76.0 GB) offers the best balance of quality and VRAM usage. Q4_K_L (77.6 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 35.4 GB.

VRAM requirement by quantization

IQ2_XXS

35.4 GB~53%

Q2_K

54.2 GB~75%

Q3_K_L

65.1 GB~86%

Q4_K_M ★

76.0 GB~89%

Q4_K_L

77.6 GB~90%

BF16

251.1 GB~100%

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run Devstral 2 123B Instruct 2512 on a Mac?

Yes — Mac Studio M5 Max (48 GB) and 55 other Macs can run Devstral 2 123B Instruct 2512, the smallest at IQ2_XS. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 76.0 GB of usable unified memory (RAM minus macOS overhead), from Mac Studio (M3 Ultra, 96GB).

Can I run Devstral 2 123B Instruct 2512 locally?

Yes — Devstral 2 123B Instruct 2512 can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is Devstral 2 123B Instruct 2512?

At Q4_K_M, Devstral 2 123B Instruct 2512 can reach ~63 tok/s on AMD Instinct MI350X. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.

tok/s = (bandwidth GB/s ÷ model GB) × efficiency

Example: NVIDIA B200 → 8000 ÷ 76.0 × 0.65 = ~68 tok/s

Estimated speed at Q4_K_M (76.0 GB)

NVIDIA B200

~68 tok/s

NVIDIA B300

~68 tok/s

AMD Instinct MI350X

~63 tok/s

Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.

Learn more about tok/s estimation →

What's the download size of Devstral 2 123B Instruct 2512?

At Q4_K_M, the download is about 75.02 GB. The full-precision BF16 version is 250.05 GB. The smallest option (IQ2_XXS) is 34.38 GB.

Which GPUs can run Devstral 2 123B Instruct 2512?

No single consumer GPU has enough VRAM to run Devstral 2 123B Instruct 2512 at Q4_K_M (76.0 GB). Multi-GPU or professional hardware is required.

Which devices can run Devstral 2 123B Instruct 2512?

30 devices with unified memory can run Devstral 2 123B Instruct 2512 at Q4_K_M (76.0 GB), including ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB), GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.