Modelpublic

backpack-run/Devstral-Small-2-24B-Instruct-2512-GGUF

sourceHugging Faceapache-2.0updated 19d agoView on Hugging Face

4likes4.1kdownloads

Model Card

Devstral-Small-2-24B-Instruct-2512 — Backpack GGUF

GGUF quantizations of mistralai/Devstral-Small-2-24B-Instruct-2512, tested for llama.cpp-compatible image-and-text inference and packaged for Backpack.

Model

Property	Value
Original model	`mistralai/Devstral-Small-2-24B-Instruct-2512`
Original publisher	mistralai
Upstream revision	`55c5b41e98c2dbd21b0c8afffc540dcfc9eb5128`
Architecture	Mistral3ForConditionalGeneration
Parameters	24,011,361,840
Context length	Not declared
Input modalities	text, image
Output modalities	text
License	apache-2.0

Available packages

Quantization	Size	Approx. RAM	Recommended for
Q4KM	13.3 GiB	21.23 GB	Most users
Q5KM	15.6 GiB	24.51 GB	Higher quality
Q8_0	23.3 GiB	35.7 GB	Plenty of memory

Memory values are estimates, not guarantees. Runtime configuration and context length change actual use.

Multimodal projector

File	Precision	Size
`Devstral-Small-2-24B-Instruct-2512-mmproj-F16.gguf`	F16	837.4 MiB

The projector is required for image input and must be used with one of the language-model GGUF files above.

Backpack recommendation

Recommended: Q4_K_M. It usually offers a practical quality, size, and speed balance for local inference.

Run with llama.cpp

Using the llama.cpp revision recorded below:

bash

llama-mtmd-cli --model Devstral-Small-2-24B-Instruct-2512-Q4_K_M.gguf --mmproj Devstral-Small-2-24B-Instruct-2512-mmproj-F16.gguf --image image.jpg --prompt "Describe this image."

Run with Backpack

These artifacts and backpack-model.yaml are prepared for the Backpack AI workspace.

Validation

Package	Integrity	Load	Inference	Tokenizer
Q4KM	passed	passed	passed	passed
Q5KM	passed	passed	passed	passed
Q8_0	passed	passed	passed	passed

Capability qualification: code

Tested with Devstral-Small-2-24B-Instruct-2512-Q4_K_M.gguf. These are deterministic smoke tests, not benchmark scores.

Capability	Status
chat	passed
code_generation	passed
structuredtoolarguments	passed
tool_calling	passed
multiturntool_loop	passed

—Packaged: 2026-08-30T16:36:59.071073+00:00
—llama.cpp revision: bdf3955159d7184f44b76091973eeff532890a35
—SHA-256 checksums: see checksums.sha256

—Devstral-Small-2-24B-Instruct-2512-Q4_K_M.gguf: e78a384e8f62862fa111a0ead8ff1d8f1373f0b17d00509b5cc77e91ee960a71

—Devstral-Small-2-24B-Instruct-2512-Q5_K_M.gguf: bc963f56b456eb681e8cc966314b7ba2919b0ac26dd514cd5fef5bff7db65706

—Devstral-Small-2-24B-Instruct-2512-Q8_0.gguf: d8e3a8fb757957d5beb11514e37b9b907707d08be38d5e494a63e514a4b7138f

—Devstral-Small-2-24B-Instruct-2512-mmproj-F16.gguf: 54fec1ceaf733d4f2ddc265c879997da7c675ade6d330f52d7e6e5f62a4738fd

Provenance

The source model was resolved to immutable revision 55c5b41e98c2dbd21b0c8afffc540dcfc9eb5128. It was converted with llama.cpp's convert_hf_to_gguf.py, including its multimodal projector, and quantized with llama-quantize; the exact tested revision is recorded above and in backpack-model.yaml.

License and attribution

Upstream declares apache-2.0. Review the upstream model card and comply with all applicable terms.

Backpack does not claim ownership of the original model. These artifacts are packaged and quantized distributions of the upstream model.

Disclaimer

Quantization can alter output quality. Memory estimates vary with runtime configuration, context length, and hardware.