MiniMax M3
A long-context agentic MoE aimed squarely at tool-calling workloads. Its official weights contain about 427B parameters, with about 23B activated per token according to the model card.
MiniMaxMixture of expertsagenticreasoning5 machines can run it
Total parameters
427.04B
Official spec sets memory need
Active parameters
23B
Official spec sets decode speed
Context
1M
Official spec tokens
4-bit weights
233 GB
Calculated before KV cache
This is a sparse mixture of experts. All 427.04B of weights must be available to the runtime, but only 23B are read per generated token. Ordinary runtimes keep the full quantized model in memory; a runtime-specific SSD-streaming artifact can retain a smaller working set and fetch expert data on demand, trading speed for capacity.
Model specification
Publisher
MiniMax Official spec
Architecture
Sparse mixture of experts Official spec
Total parameters
427.04B Official spec
Active parameters
23B per token Official spec
Context length
1,048,576 tokens Official spec
Attention ⓘ
GQA · 4 KV heads × 128 dim × 60 layers Official spec
Licence
Apache 2.0 Official spec source
Specification confidence ⓘ
High · verified 2026-10-19 Official spec source
Released
1 Apr 2026 Official spec
Official source
Quantizations and memory
| Quantization | Format | Bits/weight | Resident | Download | Quality kept |
|---|---|---|---|---|---|
MLX 4-bit | mlx | 4.5 | 232.7 GB | 232.7 GB | 98.0% |
Q4_K_Mdefault | gguf | 4.85 | 248.3 GB | 248.3 GB | 98.5% |
Q5_K_M | gguf | 5.69 | 291.4 GB | 291.4 GB | 99.3% |
Q8_0 | gguf | 8.5 | 431.0 GB | 431.0 GB | 99.9% |
Generic weight sizes are computed from the parameter count and bits per weight plus a format-specific overhead. Runtime-specific artifacts use their published resident and download footprints; streamed models can therefore require much more disk than memory. Quality retention is an assumption, not a measured evaluation.
Coding & quality benchmarksCompare coding results →
| Benchmark | Category | Score | Run details | Reported by | Date | Source |
|---|---|---|---|---|---|---|
| SWE-bench Pro | coding | 59 | Not supplied by source | canonical model_repository | 23 Jun 2026 | link |
| SWE-bench Verified | coding | 80.5 | Not supplied by source | canonical model_repository | 23 Jun 2026 | link |
We show source metrics rather than deriving one opaque quality number. Different benchmarks measure genuinely different things, and collapsing them into a single score would hide exactly the disagreements worth seeing.
Recommendations
Cheapest that can run it
~24–34.5 t/s · $8,287
Lowest purchase price among configurations where the model fits at some quantization in our catalogue. Speed is not considered.
Cheapest above 20 t/s
~24–34.5 t/s · $8,287
20 tokens/second is roughly the point at which generation keeps pace with reading. Below it, interactive use feels like waiting.
Cheapest above 40 t/s
~42.5–61.1 t/s · $11,350
40 tokens/second is the threshold most people describe as comfortable for coding agents, where output arrives faster than you can review it.
Fastest with real measurements
Nothing in the database qualifies.
Highest throughput among configurations with an actual published measurement rather than our estimate.
Best throughput per purchase-price unit
~42.5–61.1 t/s · $11,350
Highest decode tokens/second per 1,000 units of the displayed purchase currency. Ignores running costs and resale — see the economics section for the full picture.
Hardware that runs MiniMax M30 measured, 5 estimated
5 of 5 rows
| Hardware↕ | Quant | Memory↕ | Decode▼ | Prefill↕ | Context | Price↕ | Fit | Confidence↕ |
|---|---|---|---|---|---|---|---|---|
| NVIDIA DGX H200 (8x H200, 1,128 GB) NVIDIA · 1,128 GB · 38,400 GB/s | Q8_0recommended | 448.7 GB | ~334–481 t/s | ~27000–56070 t/s | 1024K | $359,430 | Comfortable | Estimated |
| Lenovo ThinkStation PX (4x RTX PRO 6000, 384 GB) Lenovo · 384 GB · 7,168 GB/s | Q4_K_Mrecommended | 258.9 GB | ~109–157 t/s | ~1300–2690 t/s | 512K | $58,553 | Comfortable | Estimated |
| Mac Studio M5 Ultra 512 GB Apple · 512 GB · 1,200 GB/s | Q5_K_Mrecommended | 302.0 GB | ~42.5–61.1 t/s | ~293–608 t/s | 1024K | $11,350 | Comfortable | Estimated |
| NVIDIA DGX Station GB300 (748 GB) NVIDIA · 748 GB · 7,100 GB/s | Q8_0recommended | 445.9 GB | ~29.3–42.1 t/s | ~3610–7490 t/s | 1024K | $90,082 | Comfortable | Estimated |
| Mac Studio M3 Ultra 512 GB Apple · 512 GB · 819 GB/s | Q5_K_Mrecommended | 302.0 GB | ~24–34.5 t/s | ~70.6–147 t/s | 1024K | $8,287 | Comfortable | Estimated |
Submit a benchmarkContributions are reviewed before publication.
Similar models
MiMo-V2.6 Flash 309B-A15B
XiaomiMiMo · 309B (15B active) · MIT
DeepSeek-V4.1 Flash
DeepSeek · 763.205B (16B active) · MIT
MiniMax Text 01 456B-A46B
MiniMax · 456.09B (46B active) · MiniMax
Step 3.7 Flash 198B-A11B
StepFun · 201.365B (11B active) · Apache 2.0
Qwen3.8 Flash Next
Alibaba Qwen · 180B (6B active) · Qwen Community 1.0
MiMo-V2.6 Pro 1.02T-A42B
XiaomiMiMo · 1.0T (42B active) · MIT