Skip to content
English

Switch theme

// MiniMax · Vision-language

MiniMax-VL-01: VRAM & local deployment

456.4B (MoE) parameters. About 305 GB at 4-bit and 1.00 TB at BF16 with one 8K-token request.

released 2025-01-12licence: Licence not statedmemory figures are estimates

01Specs

PublisherMiniMax · MiniMaxAI
Family / seriesMiniMax · MiniMax-VL-01
TypeVision-language (image-text-to-text)
Parameters456.4B (MoE) Source: counted from official safetensors weights
Context length8K tokens (8,192)max_position_embeddings in config.json
Architecture80 layers (10 full-attention) · GQA+HYBRID · MoE
Released weightsBF16 · 915 GBdownload
Licence— · Licence not statedThe model card metadata does not state a licence; check the repository files.
Released2025-01-12 (Hugging Face repo created)

02How much VRAM it needs (estimate)

Total = weights + 10% + 0.5GB overhead + KV cache for one request (FP16 KV). More concurrent users multiply the KV column.

PrecisionWeightsTotal @ 8KSmallest GPUMac
BF16 / FP16 (original)913 GB 1.00 TB 8 × H200 (1,128GB)—
FP8 / INT8 / Q8_0485 GB 534 GB 8 × H100 (640GB)—
4-bit (Q4_K_M / AWQ / GPTQ-Int4)277 GB 305 GB 4 × RTX PRO 6000 (384GB)Mac 512GB

KV cache per request: 8K 0.3 GB

Bits per weight: BF16 16, FP8 8.5, 4-bit 4.85. “Fits” = total ≤ 90% of memory; Mac GPU can use ~2/3 of unified memory (≤36GB) or ~3/4 (larger). For MoE, all experts must be in memory, so size by total parameters. Method →

03How to run it

Only runtimes the publisher or the runtime’s official library names for this model.

Which inference engine to choose →

04Official repositories & quantised builds

1 official repo(s) from MiniMaxAI.

05Other MiniMax models

All 11 →
ModelParams4-bit totalReleased
MiniMax-Music32.43B2.1 GB2026-08-07
MiniMax-H333.1B22.6 GB2026-07-28
MiniMax-M3428B / 23B286 GB2026-06-02
MiniMax-M2.7228.7B155 GB2026-04-09
MiniMax-M2.5228.7B155 GB2026-02-12
MiniMax-M2.1228.7B155 GB2025-12-20
MiniMax-M2230B / 10B156 GB2025-10-22
MiniMax-M1-80k456B / 45.9B305 GB2025-06-13

06Sources

Data checked 2025-07-03. Parameters, context and licence are copied from the official repository; memory figures are our estimates.

Data updated: Self-host open models(UTC+8)