Hy3-preview vs MiniMax-M2
Hy3-preview (298.8B parameters) and MiniMax-M2 (228.7B parameters) side by side: the memory each needs at every precision, what it costs to run on a live GPU, and the context window, KV cache and license where they are published. Numbers are computed from the models' published specs; this page does not rank quality.
Side by side
| Fact | Hy3-preview | MiniMax-M2 |
|---|---|---|
| Parameters | 298.8B | 228.7B (Mixture-of-experts: 8 of 256 experts active per token (10B active parameters, stated on the model card)) |
| Architecture | Grouped-query attention; mixture of 192 experts, 8 active per token | Grouped-query attention; mixture of 256 experts, 8 active per token |
| Context length | 262,144 tokens | 192K tokens (196,608) |
| License | – | Modified MIT |
| Published precision | BF16 | F8_E4M3 |
| VRAM needed, As published | 668 GB | 256 GB |
| VRAM needed, FP8 | 334 GB | Not a smaller option |
| VRAM needed, INT4 | 167 GB | 128 GB |
| Cheapest live fit, As published | RTX PRO 6000 × 7 · $10.29/hr | RTX 4080 Super × 8 · $2.71/hr |
| Cheapest live fit, FP8 | L40 × 7 · $5.31/hr | – |
| Cheapest live fit, INT4 | RTX A5000 × 7 · $1.23/hr | RTX A5000 × 6 · $1.06/hr |
| KV cache per token (16-bit) | 320 KB | 248 KB |
| KV cache at 32k tokens | 10.0 GB | 7.75 GB |
| KV cache at 128k tokens | 40.0 GB | 31.0 GB |
VRAM is the weight size at each precision times a flat 1.2 overhead; see the methodology. The FP8 and INT4 rows need a quantized checkpoint or an engine that quantizes on load. The fit is the lowest-priced single GPU type that holds the model at that precision, or the lowest-priced multi-GPU set (up to 8) when none does. KV cache is for one sequence at 16-bit, computed from each model's config where the attention layout is known.
Which to pick
- MiniMax-M2 needs less VRAM at its published precision (256 GB against 668 GB), so it fits on a smaller GPU.
- Hy3-preview lists the longer context window (262,144 tokens against 196,608).
- MiniMax-M2 caches less per sequence at 32k tokens (7.8 GB against 10.0 GB), leaving more memory for batching.
- MiniMax-M2 has the cheaper live GPU fit at its published precision ($2.71/hr against $10.29/hr).
These follow only from the facts in the table above. Whether either model does your task well is a separate question this page does not answer.
Keep reading
- Hy3-preview: full VRAM table and live GPU fit
- MiniMax-M2: full VRAM table and live GPU fit
- The Hunyuan model series
- The MiniMax model series
- All models that fit in 288 GB
Other comparisons
