MiniMax

MiniMax is a Chinese AI lab whose open-weight releases target one specific niche: very large Mixture-of-Experts (MoE) models that pair huge total parameter counts with small active-parameter counts, plus extremely long context windows. The family on AI Hashrate has two members: MiniMax-M2.7 (229B total, 10B active, 200K context) and MiniMax-M3 (428B total, 23B active, 1M context). Both are community favorites for agent and coding workloads, with million-plus Hugging Face downloads — but their weight size alone puts them beyond any ordinary gaming GPU. Here is what running them locally actually takes.

All VRAM figures below use our standard estimator: Q4 weights + KV cache at 8K context + 1 GB runtime overhead. Each model page shows the full 4K–32K matrix.

Size ladder

• MiniMax-M2.7 (229B, MoE with 10B active per token) — Q4 needs ≈129 GB VRAM at 8K context → a 192GB+ setup (multi-GPU or high-capacity unified memory). At Q2 it drops to ≈67 GB, which a single 96GB card can hold. See MiniMax-M2.7 details
• MiniMax-M3 (428B, MoE with 23B active per token) — Q4 needs ≈241 GB VRAM at 8K context → beyond the 192GB+ tier; plan for 256GB+ across multiple GPUs. At Q2 it needs ≈125 GB, which fits a 192GB-class rig. See MiniMax-M3 details

What makes MiniMax different

MoE economics. M2.7 stores 229B parameters but activates only about 10B per token (256 experts, 8 per token), so decode speed is closer to a 10B-class dense model than to anything its file size suggests. M3 activates 23B per token (128 experts, 4 per token) — heavier per token, but remarkably light for a 428B model. This is the family's whole pitch: flagship-tier knowledge at mid-tier compute cost per token.

Long context — and what it costs. The KV cache grows linearly with context length. On our estimator M2.7 adds ≈0.25 GB per 1K tokens, so its full 200K window means roughly +50 GB on top of the 8K baseline. M3 adds ≈0.58 GB per 1K tokens; running its full 1M context locally would demand ≈589 GB of KV cache alone. In practice, run M3 at 32K–128K and you still get a context window most workloads never exhaust.

Licensing. Both models ship under MiniMax's own licenses (MiniMax License for M2.7, MiniMax Community License for M3) rather than OSI-approved terms. Read the terms before any commercial deployment.

Which one should you pick

• Under 96GB VRAM: neither model fits, even at Q2. These are workstation and server-class models — look at smaller families like Qwen3 or Llama for a single gaming GPU.
• 96GB: M2.7 runs at Q2 (≈67 GB) with comfortable headroom. Q4 does not fit.
• 192GB: the sweet spot for M2.7 — Q4 at ≈129 GB leaves room for long-context KV cache. M3 fits only at Q2 (≈125 GB), and it is tight.
• 256GB+ (multi-GPU or maxed-out unified memory): M3 at Q4 becomes possible. This is firmly multi-card territory.
• Speed expectations: tok/s tracks active parameters, not total ones. On the same rig, M2.7 (10B active) will feel noticeably quicker than M3 (23B active) — assuming the weights actually fit in VRAM.

If you must pick one: M2.7 for most local rigs, because 192GB-class hardware can already run it at full Q4 quality; M3 only if you own 256GB+ of VRAM, or you genuinely need the 1M context and can tolerate Q2.

Links

• MiniMax-M2.7 on AI Hashrate — full VRAM matrix and GPU fit checks
• MiniMax-M3 on AI Hashrate — full VRAM matrix and GPU fit checks
• Best GPUs for 70B+ models — cards that fit this size class
• How much VRAM do local LLMs need? — our estimation methodology
• Running large models on consumer GPUs — quantization and offload strategies
• MiniMax-M2.7 on Hugging Face / MiniMax-M3 on Hugging Face