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 expertsagenticreasoning2 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 sit in memory, but only 23B are read per generated token — so it needs the memory of a 427.04B model and generates at roughly the speed of a 23B one. That is why large unified-memory machines suit it and fast 32 GB GPUs do not.
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 2026-09-07 Official spec source
Released
1 Apr 2026 Official spec
Official source
Quantizations and memory
| Quantization | Format | Bits/weight | Weights | Quality kept |
|---|---|---|---|---|
MLX 4-bit | mlx | 4.5 | 232.7 GB | 98.0% |
Q4_K_Mdefault | gguf | 4.85 | 248.3 GB | 98.5% |
Q8_0 | gguf | 8.5 | 431.0 GB | 99.9% |
Weight sizes are computed from the parameter count and bits per weight plus a format-specific overhead for the layers that stay at higher precision — not read from a specific published file. Quality retention is an assumption, not a measured evaluation.
Quality benchmarks
| Benchmark | Category | Score | Reported by | Date | Source |
|---|---|---|---|---|---|
| SWE-bench Pro | coding | 59 | canonical model_repository | 23 Jun 2026 | link |
| SWE-bench Verified | coding | 80.5 | 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
~28–40.3 t/s · €9.199
Lowest purchase price among configurations where the model fits at some quantization in our catalogue. Speed is not considered.
Cheapest above 20 t/s
~28–40.3 t/s · €9.199
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
~49.4–71.1 t/s · €12.599
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 euro
~49.4–71.1 t/s · €12.599
Highest decode tokens/second per EUR 1,000 of purchase price. Ignores running costs and resale — see the economics section for the full picture.
Hardware that runs MiniMax M30 measured, 2 estimated
2 of 2 rows
| Hardware↕ | Quant | Memory↕ | Decode▼ | Prefill↕ | Context | Price↕ | Fit | Confidence↕ |
|---|---|---|---|---|---|---|---|---|
| Mac Studio M5 Ultra 512 GB Apple · 512 GB · 1,200 GB/s | Q4_K_M | 257.7 GB | ~49.4–71.1 t/s | ~293–608 t/s | 1024K | €12.599 | Comfortable | Estimated |
| Mac Studio M3 Ultra 512 GB Apple · 512 GB · 819 GB/s | Q4_K_M | 257.7 GB | ~28–40.3 t/s | ~70.6–147 t/s | 1024K | €9.199 | Comfortable | Estimated |
Similar models
Qwen3.8 Flash Next
Alibaba Qwen · 180B (6B active) · Qwen Community 1.0
Kimi K2.6
Moonshot AI · 1.0T (32B active) · Modified MIT
GLM-5.3 Flash
Z.ai · 321.323B (18B active) · MIT
DeepSeek-V4 Flash
DeepSeek · 304.18B (13B active) · MIT
Hunyuan Hy3
Tencent · 298.786B (21B active) · Tencent Hunyuan Community
DeepSeek-V3.2
DeepSeek · 685.397B (37B active) · MIT