Short answer: Mistral’s official Mistral Small 3.2 24B Instruct repository contains a 48.023GB BF16 checkpoint, available either as one consolidated file or as ten shards representing the same weights. Community GGUF conversions list roughly 14.3GB Q4_K_M, 16.8GB Q5_K_M, 19.3GB Q6_K and 25.1GB Q8_0. That makes 16GB a constrained Q4 path, 24GB the practical Q4/Q5/Q6 tier, 32GB the Q8 tier, and 64GB the practical official-BF16 tier. Vision processing and long context can move those boundaries upward.
Repository counting warning: do not add the 48.023GB consolidated file to the ten sharded files. They are alternative representations of the same checkpoint, not two required copies. The ten shards total the same 48.023GB.
Mistral Small 3.2 file sizes
| Artifact | Displayed/calculated size | Practical memory tier |
|---|---|---|
| Community Q4_K_M GGUF | 14.3GB | 16GB is constrained; 24GB adds useful room |
| Community Q5_K_M GGUF | 16.8GB | Does not fit wholly on 16GB; target 24GB+ |
| Community Q6_K GGUF | 19.3GB | 24GB is plausible with bounded context |
| Community Q8_0 GGUF | 25.1GB | Does not fit wholly on 24GB; target 32GB+ |
| Official BF16 checkpoint | 48.023GB | 48GB is a file-size trap; target 64GB+ |
16GB, 24GB, 32GB or 64GB?
- 16GB: a 14.3GB Q4_K_M file leaves only about 1.7GB on decimal labels before cache, vision state, runtime buffers and display use. Treat it as tightly constrained.
- 24GB: the practical community Q4/Q5/Q6 lane. The 19.3GB Q6 example leaves about 4.7GB before the rest of the workload.
- 32GB: the first listed tier above a 25.1GB Q8_0 file, with roughly 6.9GB before cache and runtime state.
- 48GB: the official 48.023GB BF16 checkpoint already exceeds the nominal decimal label before any other allocation.
- 64GB: the practical BF16 starting tier, subject to context, image input and runtime support.
Compare the 24GB/32GB/48GB tiers and use the VRAM calculator as a starting range rather than a multimodal guarantee.
This is a multimodal Mistral 3 model
The official configuration identifies Mistral3ForConditionalGeneration with a Pixtral vision encoder. It records a 1,540-pixel image size, 14-pixel patches, 24 vision layers and a 1,024-dimensional vision hidden state. Image preprocessing, vision encoding and the multimodal projector are part of the complete workload.
A community GGUF may package the language model and projector differently, and some conversions are deliberately text-only. Record every required file and confirm that image input works before calling the deployment multimodal.
Exact language-cache arithmetic
Mistral’s official text configuration records 40 layers, 8 KV heads and a 128-dimensional attention head. For one sequence with a two-byte FP16/BF16 cache and every language layer on the accelerator:
bytes per token = 2 × 40 layers × 8 KV heads × 128 head dimension × 2 bytes
= 327,680 bytes
= 320 KiB per token| Allocated text-token slots | Illustrative FP16/BF16 KV payload |
|---|---|
| 4,096 | 1.25GiB |
| 8,192 | 2.5GiB |
| 32,768 | 10GiB |
| 131,072 | 40GiB |
The official 131,072-position limit is an architectural maximum, not a promise that a 64GB system can hold BF16 weights, a 40GiB full-precision cache and the multimodal runtime together.
Mistral Small 3.2 versus related 24B releases
- Mistral Small 3.2: general-purpose multimodal instruct model; this page.
- Devstral Small 1.1: agentic coding model with an official text-only GGUF path. See the Devstral VRAM guide.
- Magistral Small 1.2: reasoning-focused model; its official GGUF path has specific template and vision limitations. See the Magistral VRAM guide.
Similar parameter counts and GGUF sizes do not make the releases interchangeable. Check model identity, intended use, modality, license, template and runtime requirements.
Community GGUF evidence needs provenance
The Q4-through-Q8 examples above come from community conversions of the official model. Record the converter, revision, filename and checksum. Determine whether the vision projector is integrated, separate or omitted. Use the quantization guide without treating a smaller file as proof of unchanged text or image quality.
A reproducible multimodal fit test
- Record exact model repository, revision, every required file, size and checksum.
- Record whether you use the official consolidated or sharded checkpoint—never both as one weight total.
- Record language and vision precision, projector packaging and runtime/version.
- Record accelerator capacity, device placement, system RAM, OS and driver.
- Record image dimensions/count, allocated text slots, actual prompt, output target, concurrency and cache precision.
- Measure memory before load, after load, after image encoding and at peak generation.
- Capture preprocessing time, time to first token, throughput and any fallback or out-of-memory error.
- Evaluate both text and image-grounded output quality.
Run the same disclosed workload and share a reproducible result.
Continue planning
Official sources and limits
- Mistral official model card and checkpoint
- Mistral official repository files
- Mistral official configuration
- Community GGUF example
- HighVRAM methodology
Reviewed October 25, 2026. Official checkpoint size and configuration come from Mistral’s repository; community GGUF examples are labeled separately. HighVRAM has not independently benchmarked every conversion, image configuration or runtime. Fit does not establish speed or output quality.
Compare adjacent Mistral releases: the Mistral 24B hub contrasts Small 3.2 with Devstral and Magistral by workload, modality, artifact provenance and runtime constraints.
Cross-family vision comparison: the multimodal LLM VRAM guide compares Mistral Small 3.2 with Gemma 3 and Llama 3.2 Vision across 6GB-through-256GB paths.