Mistral & Mixtral Family: Complete VRAM Reference
Mistral AI: The French Challenger
Mistral AI's portfolio spans dense (3B-128B) and MoE (46B-675B) models. Key distinction: Mistral Large 2 (123B) was API-only, never downloadable, but that changed on December 2, 2025, when Mistral shipped its first MoE since the original Mixtral series: Mistral Large 3, a 675B-total/41B-active MoE model, fully open under Apache 2.0. That same release wave (dubbed "Mistral 3") also brought Small 4 (119B total/~6B active MoE) and the dense Ministral 3 line (3B, 8B, 14B). Open-weight dense/classic-MoE models: Mistral 7B, NeMo 12B, Small 22B, Small 3.1 24B, Mixtral 8x7B, Mixtral 8x22B, and now Mistral Medium 3.5 (128B dense, April 2026). Real Q4_K_M weight size across the classic open-weight lineup:
Dense Models
Mistral 7B Q4 (~4.4 GB): lightweight entry. NeMo 12B Q4 (~7.4 GB): designed for consumer GPUs, excellent performance-per-VRAM. Small 22B Q4 (~13.5 GB). Small 3.1 24B Q4 (~14.6 GB): fits 16 GB GPUs, adds vision. The 24B on 16 GB is the Mistral sweet spot for consumer hardware. Mistral Medium 3.5 (128B, April 2026) is the newest dense release: it folds Mistral's dedicated reasoning model (Magistral) and dedicated coding model (Devstral 2) into one configurable-effort model with standard GQA attention (8 KV heads) and 77.6% on SWE-bench Verified. Despite retiring the smaller Devstral name, Medium 3.5 is a genuinely large model: Q4_K_M needs roughly 78 GB for weights alone, more than any dense model covered above, so it belongs alongside the 48 GB+ workstation tier below, not the consumer sweet spot. Also worth knowing about: Ministral 3, a separate dense line at 3B/8B/14B released the same December 2025 wave, purpose-built for compute-constrained deployment via a technique Mistral calls Cascade Distillation, each size shipping base, instruct, and reasoning variants with image understanding, a plausible fit for the 8-16 GB tiers once quantized GGUFs are widely available, though this site does not yet have confirmed VRAM figures for it.
Mixtral MoE: Just Past Consumer Reach
Mixtral 8x7B (46.7B total, 12.9B active) at Q4 needs ~28 GB, just past single RTX 4090. Partial offloading works. Mixtral 8x22B (141B total, 39B active) at Q4 needs ~80 GB, workstation territory. UNIQUE: no competitor site has VRAM-by-quantization data for 8x22B.
Mistral 3: Large 3 and Small 4 (December 2025)
Mistral's December 2025 "Mistral 3" wave introduced its first MoE architecture since the original Mixtral series, and its first fully open frontier-scale model: Mistral Large 3 (675B total, 41B active, Apache 2.0, 256K context) ranks #2 among open-weight non-reasoning models on the LMArena leaderboard and is a legitimate frontier-class open model, not a research curiosity, but at these numbers it needs serious hardware. Small 4 (119B total, ~6B active per token, also MoE) is the more approachable sibling, unifying what used to be separate Small/Magistral/Pixtral/Devstral releases into one model, and Mistral itself frames it as running at roughly 6-8B-dense inference cost despite the 119B total footprint. This site doesn't yet have confirmed internal architecture details (layer count, attention head configuration) for either model, so the table below shows weights-only VRAM using the same total-parameter formula this site uses everywhere else: actual totals with KV cache and runtime overhead will run somewhat higher, as they do for every other model on this site.
| Model | Q4_K_M | Q3_K_M | Q2_K |
|---|---|---|---|
| Mistral Small 4 (119B total / ~6B active) | ~72.5 GB | ~57.2 GB | ~45.3 GB |
| Mistral Large 3 (675B total / 41B active) | ~411 GB | ~325 GB | ~257 GB |
What Can You Actually Download?
Apache 2.0: Mistral 7B, NeMo 12B, Small 22B, Mixtral 8x7B, Mixtral 8x22B, Small 3.1 24B, Mistral Large 3, Small 4, Medium 3.5, and the Ministral 3 line. Research License (non-commercial): none currently: Small 3.1 24B, previously research-only, now ships under the same open terms as the rest of the December 2025 wave. API-only (no weights): Mistral Large 2, superseded for local use by the now-open Mistral Large 3. Always verify the license on Mistral's own model cards before planning deployments, since licensing terms have shifted across releases.
Which Mistral for Your GPU?
8 GB: Mistral 7B Q4 (~4.4 GB) or Q5. 12-16 GB: NeMo 12B Q4 (~7.4 GB) or Small 3.1 24B Q4 (~14.6 GB, needs 16 GB). 24 GB: Mixtral 8x7B Q4 (~28.4 GB weights, partial offload). 48 GB+: Mixtral 8x22B Q4 (~85.9 GB weights, ~97 GB total) remains the practical local pick: even Small 4, the smaller of the new Mistral 3 MoE models, needs ~57 GB at Q3_K_M before KV cache, so it and Mistral Large 3 are effectively cloud/workstation-cluster territory rather than a single-GPU upgrade path. NeMo 12B is the practical pick for most users.
Frequently asked questions
- Can I run Mistral Large 2 locally?
- No, Large 2 itself was always API-only via Mistral's platform. But its open-weight successor, Mistral Large 3 (675B MoE, Apache 2.0, released December 2025), can be downloaded: you just need multi-GPU datacenter-class hardware to run it (roughly 411 GB at Q4_K_M). For a model you can realistically self-host, use Llama 3.3 70B or Qwen 2.5 72B instead.
- Why does Mixtral 8x7B need 28 GB at Q4?
- All 46.7B expert parameters must be loaded even though 12.9B are active per token. MoE saves compute, not memory. The same logic applies to the newer Mistral Large 3 and Small 4 MoE models: all their listed total parameters, not just the active ones, need to be loaded.
- Mistral 7B vs Llama 3.1 8B?
- Very close. Mistral 7B uses less VRAM (~4.4 vs ~5 GB) and has sliding window for long docs. Llama 3.1 8B has better benchmarks and ecosystem. Either is excellent.
- Best Mistral for 16 GB?
- Small 3.1 24B at Q4 (~14 GB): fits with room for context, adds vision. NeMo 12B at Q8 for near-lossless text-only.
- Is Mistral Large 3 actually usable on consumer hardware?
- No, not meaningfully. Even at Q2_K (~257 GB weights alone), it needs multiple 80 GB datacenter GPUs pooled together. Mistral's own guidance points to a single 8-GPU H200 node in FP8 for production use. Small 4 (119B total, ~6B active) is more approachable at roughly 57 GB at Q3_K_M, but still exceeds what a single consumer GPU offers; treat both as cloud-rental or workstation-cluster models, not local upgrades.