Skip to content
English

Switch theme

// MiniCPM / VoxCPM · Text LLM

MiniCPM4-8B: VRAM & local deployment

8.19B parameters. About 6.2 GB at 4-bit and 18.8 GB at BF16 with one 8K-token request.

released 2025-06-05licence: Commercial OKmemory figures are estimates

01Specs

PublisherOpenBMB · openbmb
Family / seriesMiniCPM / VoxCPM · MiniCPM4
TypeText LLM (text-generation)
Parameters8.19B Source: counted from official safetensors weights
Context length32K tokens (32,768)max_position_embeddings in config.json
Architecture32 layers · GQA
Released weightsBF16 · 16.4 GBdownload
LicenceApache-2.0 · Commercial OKPermissive licence (Apache-2.0 / MIT / CC-BY / OpenMDW): commercial use allowed.
Released2025-06-05 (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 @ 8KTotal @ 32KSmallest GPUMac
BF16 / FP16 (original)16.4 GB 18.8 GB19.6 GB RTX 4090 24GBMac 36GB
FP8 / INT8 / Q8_08.7 GB 10.3 GB11.1 GB 12GB GPUMac 24GB
4-bit (Q4_K_M / AWQ / GPTQ-Int4)5.0 GB 6.2 GB7.0 GB 8GB GPUMac 16GB

KV cache per request: 8K 0.3 GB · 32K 1.1 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

3 official repo(s) from openbmb.

05Other MiniCPM / VoxCPM models

All 23 →
ModelParams4-bit totalReleased
MiniCPM4-0.5B · same series434M0.9 GB2025-06-05
MiniCPM-V-4.7-35B-A3B35.2B / 3B24.2 GB2026-10-07
MiniCPM-V-4.7-1B1.55B1.6 GB2026-10-06
MiniCPM5-2B2.52B2.5 GB2026-09-06
MiniCPM5-1B1.08B1.4 GB2026-05-21
MiniCPM-V-4.6-Thinking1.3B1.5 GB2026-05-08
MiniCPM-V-4.61.3B1.5 GB2026-04-13
VoxCPM22.29B2.0 GB2026-04-03
MiniCPM-SALA9.48B7.1 GB2026-02-11

06Sources

Data checked 2025-10-24. Parameters, context and licence are copied from the official repository; memory figures are our estimates.

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