What GPU do you need to run MiniMax-M3.1-Flash Unsloth 2-bit?
A new Framework Desktop ($6,799). No priced GPU card holds MiniMax-M3.1-Flash Unsloth 2-bit (about 145.4 GB at UD-Q2_K_XL); this system holds it with room for about 302K tokens of context (569K with 8-bit KV cache), at about 25 tok/s estimate. 1 of the 157 GPUs and systems we track fit it.
427.05B parameters, 23B active (mixture-of-experts) (Hugging Face, accessed 2026-10-26). UD-Q2_K_XL GGUF: 143.91 GB (source). Architecture config (Hugging Face config.json; sparse_attention_freq is 1 on all 60 layers, including the 3 leading dense-MLP layers (moe_layer_freq). llama.cpp minimax-m3 sets leading_dense_block_count from those leading zeros and allocates the indexer cache only on the other 57 layers: 1 KV head, K and V both sparse_index_dim 128, same element type as the main cache (src/llama-kv-cache-msa.cpp). The graph writes indexer K only; V is still allocated): 60 attention layers × 4 KV heads × 128 head dim, plus an indexer cache on 57 sparse layers (1×128 K and V), so about 149 KB of FP16 KV cache per token of context (how we estimate context).
Same checkpoint as MiniMax-M3.1-Flash. This row is Unsloth's UD-Q2_K_XL (Unsloth Dynamic 2-bit): 143,910,068,580 bytes = 143.91 GB, the sum of the four shards in the UD-Q2_K_XL folder. Unsloth says this 2-bit quant reaches 91.5% top-1 accuracy and is about 45% smaller; that is Unsloth's own claim, not verified by RigPerDollar. Unsloth recommends budgeting 160 GB+ of RAM + VRAM (or unified memory) at 16K context; its table lists about 150 GB for 2-bit. Needs a llama.cpp build with MiniMax-M3.1 support (llama.cpp PR #31192) or Unsloth Desktop. Fit ignores the vision projector. Guide: https://unsloth.ai/docs/models/minimax-m3.1-flash . Announced at https://x.com/UnslothAI/status/2114730104228286936. Use the Q4_K_M page (minimax-m3.1-flash, 258.01 GB) when the machine has room for it; AtomicChat's plain Q2_K is minimax-m3.1-flash-q2-k (155.91 GB).
Other places to get it
Checked Nov 14, 2026. We list only copies we opened and checked. Check file hashes for copies from other uploaders.
- ModelScope: Lab's own account
- Ollama: Cloud only (runs on Ollama's servers)
ollama run minimax-m3.1-flash:cloud
Single GPUs that fit, cheapest first
| Single GPU that fits | VRAM | Speed estimate | Max context estimate | Price |
|---|---|---|---|---|
| Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB | 192 GB | ~25 tok/s | 302K · 569K (8-bit) | low $6,799 |
Every device that runs it
Graphics cards, Macs and mini PCs from the wider hardware list behind My rig, smallest usable memory first. Not sure about yours? Check your exact rig.
Show all 4 devices that run it
| Device | Type | Memory | Speed estimate | Max context estimate | Price |
|---|---|---|---|---|---|
| Apple M3 Ultra 256GB | Mac | 256 GB (~192 usable) | ~64 tok/s | 302K · 569K (8-bit) | No price yet |
| Apple M5 Ultra 256GB | Mac | 256 GB (~192 usable) | ~81 tok/s | 302K · 569K (8-bit) | No price yet |
| Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB | Mini PC | 192 GB | ~25 tok/s | 302K · 569K (8-bit) | low $6,799 |
| Apple M5 Ultra 512GB | Mac | 512 GB (~384 usable) | ~81 tok/s | 1M (full) | No price yet |
Multi-GPU: cards that need two or more
| Cards | Total VRAM | Speed estimate | Max context estimate |
|---|---|---|---|
| 4 × NVIDIA RTX A6000 | 192 GB | ~76 tok/s | 302K · 569K (8-bit) |
Speed calculator
estimateEstimate = bandwidth ÷ GB read per token (the model's listed GGUF file: Q4_K_M for most models, otherwise the quant named in the list; scaled by active/total parameters for MoE), scaled by a per-platform efficiency and fixed per-token costs calibrated to 45 measured llama.cpp benchmarks (2025–26). Fit = file size + 1.5 GB headroom ≤ total VRAM (on Macs, the part of unified memory macOS lets the GPU use by default), and room for at least 4K tokens of context with an 8-bit KV cache. For a model with no KV cache data, only the file and headroom are checked, so a fit says nothing about context. Multi-GPU: VRAM adds, speed stays one card's (a layer split runs the cards in turn); max context uses the total. Speed is for short context. Max context = what is left after the file and a 1 GB runtime reserve ÷ KV cache bytes per token, default FP16 and 8-bit (q8_0) cache. How speed works · how context works