Magistral Small 1.2 24B VRAM: 16GB, 24GB, 32GB or 48GB?

Short answer: Mistral’s official Magistral Small 1.2 24B GGUF repository lists 14.3GB Q4_K_M, 16.8GB Q5_K_M, 25.1GB Q8_0 and 47.2GB BF16. That makes 16GB a constrained Q4-only target, 24GB the practical Q4/Q5 tier, 32GB the Q8_0 tier, and 48GB too tight for BF16 as a complete workload. Plan 64GB+ for BF16 working room.

Official Magistral Small 1.2 GGUF sizes

Official artifact	Displayed size	Practical memory tier
Q4_K_M	14.3GB	16GB is constrained; 24GB gives useful working room
Q5_K_M	16.8GB	Does not fit wholly on 16GB; target 24GB+
Q8_0	25.1GB	Does not fit wholly on 24GB; target 32GB+
BF16	47.2GB	Too tight on 48GB after cache/runtime allocations; target 64GB+

16GB, 24GB, 32GB or 48GB?

• 16GB: the 14.3GB Q4_K_M file leaves only about 1.7GB before cache, buffers, kernels and display use. Treat this as a tightly validated target, not a comfortable fit.
• 24GB: the practical Q4/Q5 tier. It leaves about 9.7GB beyond Q4 or 7.2GB beyond Q5 before other allocations.
• 32GB: the first common tier above the 25.1GB Q8_0 file, with about 6.9GB before cache and runtime state.
• 48GB: the 47.2GB BF16 file is a file-size trap; it leaves about 0.8GB before the rest of the workload.
• 64GB+: the practical BF16 starting tier, subject to runtime and platform support.

Compare those choices in the 24GB/32GB/48GB guide and use the VRAM calculator for context and concurrency.

Model identity and context

Mistral describes Magistral Small 1.2 as a 24B reasoning model built on Mistral Small 3.2, with supervised fine-tuning from Magistral Medium traces and reinforcement learning. The official page lists a 128K context window but warns that performance may degrade beyond 40K. That is a quality/performance caveat as well as a memory-planning signal: maximum context is not a recommended default or a fit guarantee.

Vision capability versus the recommended GGUF path

The base model has a vision encoder and multimodal capability, but Mistral states that this GGUF repository does not include the vision encoder and that its recommended llama.cpp plus mistral-common usage does not support vision. Plan this official GGUF path as text/reasoning unless you deliberately choose a different officially supported deployment.

The chat-template warning matters

A model can load and still behave incorrectly if tokenization, role formatting or reasoning chunks are mishandled. Mistral explicitly says it cannot guarantee correct behavior with the integrated stringified chat template, recommends mistral-common as the reference, and uses special [THINK]/[/THINK] chunks. Validate answer formatting and reasoning behavior before interpreting a load test as a successful deployment.

Do not turn file size into a quality ranking

The official repository supplies four capacity points, not a universal statement that BF16 is necessary or Q4 is sufficient. Compare task output, latency and memory together. The quantization guide explains format names without treating bits alone as a quality score.

A reproducible fit test

1. Record the exact Mistral filename, displayed size and checksum.
2. Use the recommended llama.cpp and mistral-common versions; record both commits/releases.
3. Record GPU or unified-memory capacity, OS, driver and device placement.
4. Measure available memory before loading.
5. Record context, output target, concurrency, cache precision and offload settings.
6. Verify reasoning chunks, role formatting and answer parsing—not just that tokens appear.
7. Capture peak memory, time to first token, throughput and any fallback or out-of-memory error.
8. Evaluate task quality at the chosen quantization.

Use the workload test and share a measured result.

Official sources and limits

• Mistral official Magistral Small 1.2 GGUF files
• Mistral official model card and usage guidance
• Mistral official BF16 model repository
• HighVRAM methodology

Reviewed October 25, 2026. File sizes and usage warnings are from Mistral’s official repository on that date. HighVRAM has not independently benchmarked every artifact or runtime. Fit does not establish speed, template correctness or output quality.

Mistral coding alternative: compare the same official artifact-size ladder with the Devstral Small 1.1 24B VRAM guide, which adds coding-agent, tool and host-resource constraints.

General-purpose multimodal alternative: compare the Mistral Small 3.2 24B VRAM guide, including its official 48.023GB checkpoint and vision-aware cache planning.

Compare adjacent Mistral releases: the Mistral 24B hub contrasts Magistral with Small 3.2 and Devstral by workload, vision path, templates and supporting software.

Hardware shortlist: map these exact files to 16GB through 64GB+ hardware paths for a 24B LLM.