MiniMax M2.7
workstationThe MiniMax workhorse that quietly passed a million downloads. M2.7 (April 2026) is the generation before M3: 62 layers, 256 routed experts with 8 active per token, GQA 8x128 attention, 200K context, ~115B total parameters in 230GB of bf16 weights, text-only. 1.14 million downloads on Hugging Face against M3’s 168k says which one the local crowd actually runs.
What it is for. Coding and agent work at the 96GB tier: the unsloth Dynamic quants put UD-Q4_K_XL at 140.8GB (512GB-class) and UD-Q2_K_XL at 75.3GB, which fits an M5 Ultra 96GB or a single RTX PRO 6000 box with KV headroom (KV runs 254MB per 1,000 tokens at fp16). Nothing in the M2.7 line clears a 32GB Mac or a 24GB 3090 for quality runs - the 1-bit UD-IQ1_M at 60.7GB is a degraded last resort, and a 64GB M4 Max cannot hold it with cache.
The ecosystem is real: unsloth, bartowski, and mudler all ship GGUF sets, including an abliterated community variant. If you want the current-generation upgrade path, M3 adds multimodal input, 1M context via sparse attention, and 4x the experts - at 4x the memory.
- 200k
- minimax community
- MiniMax
- Apr 2026
Related models
Save your hardware and every model page answers the real question: will it run on your machine, and how fast?
Join free - save your rig →Run it locally
Per-quant memory needs and a static "can you run it?" reference - no rig entry required
The reference hardware
22 reference configs, drawn in-house. Scroll for more.
Can you run it? - reference rigs
| Rig | UD-IQ1_M | UD-Q2_K_XL | UD-Q4_K_XL |
|---|---|---|---|
| NVIDIA Jetson Orin NX 16GB | no -> cloud | no -> cloud | no -> cloud |
| Jetson AGX Orin 64GB | tight | no -> cloud | no -> cloud |
| Single GTX 1080 Ti (11GB) | no -> cloud | no -> cloud | no -> cloud |
| Single RTX 4090 (24GB) | no -> cloud | no -> cloud | no -> cloud |
| Single RTX 5090 (32GB) | no -> cloud | no -> cloud | no -> cloud |
| 2x RTX 3090 (48GB) | offload | offload | no -> cloud |
| 2x RTX 5090 (64GB) | tight | offload | no -> cloud |
| 4x H100 80GB (320GB) | fast 119.4t/s | fast 96.6t/s | fast 52.0t/s |
| NVIDIA DGX Station 748GB | fast 71.3t/s | fast 57.7t/s | fast 31.0t/s |
| 8x RTX 3090 rack (192GB) | fast 66.8t/s | fast 54.0t/s | fast 29.1t/s |
| 4x RTX 5090 (128GB) | fast 63.9t/s | fast 51.7t/s | offload |
| AMD Instinct MI300X (192GB) | fast 47.5t/s | fast 38.4t/s | fast 20.7t/s |
| 4x RTX 4090 (96GB) | fast 35.9t/s | fast 29.1t/s | offload |
| RTX PRO 6000 Blackwell (96GB) | ok 16.0t/s | ok 12.9t/s | offload |
| Mac Studio M4 Ultra 192GB | ok 10.6t/s | ok 8.6t/s | slow 4.6t/s |
| Mac Studio M4 Ultra 512GB | ok 10.6t/s | ok 8.6t/s | slow 4.6t/s |
| MacBook Pro M5 Max 128GB | slow 6.0t/s | slow 4.8t/s | no -> cloud |
| Dual EPYC 9004 + 768GB DDR5-4800 | slow 4.1t/s | slow 3.3t/s | slow 1.8t/s |
| DGX Spark 128GB unified | slow 2.4t/s | slow 2.0t/s | no -> cloud |
| Ryzen AI Max+ 395 128GB | slow 2.3t/s | slow 1.8t/s | no -> cloud |
| Epyc + 512GB DDR4-3200 + 2x RTX 3090 | slow 1.8t/s | slow 1.5t/s | slow 0.8t/s |
| Epyc + 512GB DDR4-2400 + 2x RTX 3090 | slow 1.4t/s | slow 1.1t/s | slow 0.6t/s |
Fit tiers use the same will-it-run logic as the rig finder. For comfortable fits, the badge reflects decode speed: fast >=20 t/s, ok 8-20 t/s, slow <8 t/s. t/s is a bandwidth estimate, not a measured benchmark.
Download options
Or run it in the cloud
No per-token API provider pricing tracked for MiniMax M2.7 yet. For flagship list prices, see the calculator.
Inference cost over time
Data accumulates from the first daily sync - longer ranges populate over time. Prices come from OpenRouter snapshots, not a historical API.