Mistral 24B VRAM: Small 3.2 vs Devstral vs Magistral

Short answer: Mistral Small 3.2, Devstral Small 1.1 and Magistral Small 1.2 are all 24B-class releases, but they target different work. Choose Mistral Small 3.2 for general multimodal instruction, Devstral Small 1.1 for agentic coding, and Magistral Small 1.2 for reasoning. Their familiar 14.3GB Q4, 16.8GB Q5, 25.1GB Q8 and roughly 47–48GB BF16 sizes create similar capacity tiers—constrained 16GB Q4, practical 24GB Q4/Q5, 32GB Q8 and 64GB BF16—but modality, templates, agent scaffolds and context behavior differ.

Mistral 24B models compared

Model	Primary use	Vision on the discussed path	Important deployment constraint
Mistral Small 3.2 24B Instruct	General instruction and multimodal work	Yes in the official checkpoint; community packaging varies	Count consolidated or sharded weights once; verify projector/runtime support
Devstral Small 1.1 24B	Agentic software engineering	No on the official GGUF path; vision encoder removed	Budget the agent scaffold, tools, containers and repository separately
Magistral Small 1.2 24B	Reasoning	No on the official recommended GGUF path	Use mistral-common formatting; integrated stringified template is warned against

Artifact-size comparison

Artifact tier	Mistral Small 3.2	Devstral Small 1.1	Magistral Small 1.2
Q4_K_M	14.3GB community example	14.3GB official GGUF	14.3GB official GGUF
Q5_K_M	16.8GB community example	16.8GB official GGUF	16.8GB official GGUF
Q8_0	25.1GB community example	25.1GB official GGUF	25.1GB official GGUF
BF16	48.023GB official checkpoint	47.2GB official GGUF	47.2GB official GGUF

What 16GB, 24GB, 32GB and 64GB mean

• 16GB: a 14.3GB Q4 file leaves roughly 1.7GB before cache, runtime buffers, display use and any modality-specific state. Treat it as constrained for every model.
• 24GB: the practical Q4/Q5 tier. Q5 leaves roughly 7.2GB on decimal labels before other allocations.
• 32GB: the first listed tier above a 25.1GB Q8 file, with roughly 6.9GB before cache and runtime state.
• 48GB: a 47.2GB or 48.023GB BF16 file is a file-size trap; it leaves essentially no complete-workload margin.
• 64GB: the practical BF16 starting tier, still subject to context, image work, agent processes and runtime support.

Use the 24GB/32GB/48GB capacity guide and VRAM calculator for the surrounding hardware decision.

Choose Mistral Small 3.2 when

• You need the general-purpose instruct model rather than a coding- or reasoning-specialized release.
• Your deployment requires supported image input and you have verified the exact multimodal runtime path.
• You can budget its Pixtral vision encoder, projector, image processing and text cache.
• You have recorded whether a community conversion integrates, separates or omits vision components.

Open the Mistral Small 3.2 VRAM guide for the official 48.023GB checkpoint, 320KiB-per-token text-cache arithmetic and repository double-counting warning.

Choose Devstral Small 1.1 when

• Your target is repository-level coding and tool use rather than general multimodal chat.
• You have selected the exact 2507 Devstral Small 1.1 release, not 2505 or Devstral Small 2.
• Your agent scaffold and function-calling integration are compatible.
• You can separately budget host RAM, CPU, storage, containers, builds, tests and browser/terminal tools.

Open the Devstral Small 1.1 VRAM guide for official GGUF sizes and the complete coding-agent resource checklist.

Choose Magistral Small 1.2 when

• You specifically need the reasoning-focused release and have evaluated its output quality for your task.
• You can pair llama.cpp with the recommended mistral-common version.
• Your integration uses the reference tokenization and message formatting rather than assuming the embedded stringified template is correct.
• You accept that the official GGUF/recommended path discussed here does not support vision.

Open the Magistral Small 1.2 VRAM guide for Mistral’s template, reasoning-chunk, context and vision-path warnings.

Context is not interchangeable

Each release advertises a large context window, but maximum accepted positions do not prove fit, speed or quality at that length. Small 3.2’s official language geometry produces a 320KiB-per-token FP16/BF16 cache: 2.5GiB at 8K, 10GiB at 32K and 40GiB at 128K for one sequence before vision and runtime state. Magistral’s official page additionally warns that performance may degrade beyond 40K. Devstral coding agents can fill context with repository content, tool results and edit history.

Run one fair comparison

1. Record exact model release, repository, revision, files and checksums.
2. Use the same quantization tier only when the artifacts are actually comparable.
3. Record runtime, templates/tokenization, device placement and all supporting software.
4. Use the same input budget, output target, concurrency and cache precision.
5. For Small 3.2, disclose image dimensions/count and projector packaging.
6. For Devstral, disclose agent scaffold, tools, container and repository state.
7. For Magistral, verify reasoning chunks and role formatting.
8. Measure peak accelerator memory, peak system RAM, time to first token, throughput and task success.

Use the quantization guide, run the same disclosed workload, and share a reproducible result.

Official sources and limits

• Mistral Small 3.2 official repository
• Devstral Small 1.1 official GGUF repository
• Magistral Small 1.2 official GGUF repository
• HighVRAM methodology

Reviewed October 25, 2026. Artifact sizes and deployment constraints come from the linked Mistral repositories; Small 3.2 GGUF examples are labeled community conversions. HighVRAM has not independently benchmarked every model or runtime. This is not a speed or quality ranking.

Hardware shortlist: use the best-GPU-for-24B guide after choosing the exact Mistral release.