DeepSeek236B params21B active (MoE)128k context
DeepSeek Coder V2 locally
DeepSeek Coder V2 is an open-weight LLM from the DeepSeek family with 236B parameters. Main use: code and developer agents. Detected minimum hardware: NVIDIA B100 (192 GB).
Technical facts
Parameters236B
Max context128k
Q4_K_M148.4 GB
Q5_K_M181.3 GB
Q8263.8 GB
FP16527.5 GB
FamilyDeepSeek
Last sync2026-05-12
Available quantizations
GGUF weightsQ4_K_M
148.4GB
Acceptable. Good compromise when VRAM is limited.
Q5_K_M
181.3GB
Good quality. Sweet spot for size and precision.
Q8
263.8GB
Near-FP16 quality. Comfortable for production.
FP16
527.5GB
Reference precision. Maximum quality, doubled VRAM.
Compatible GPUs
10 single-GPUGPUs that can run DeepSeek Coder V2 on a single card, ranked by VRAM headroom.
NVIDIA B100
NVIDIA192 GB · Blackwell DC
148.4 / 192 GBcomfortable · Q4
NVIDIA B200
NVIDIA192 GB · Blackwell DC
148.4 / 192 GBcomfortable · Q4
NVIDIA GB200 (per GPU)
NVIDIA192 GB · Blackwell DC
148.4 / 192 GBcomfortable · Q4
Instinct MI300X
AMD192 GB · Instinct CDNA 3+
148.4 / 192 GBcomfortable · Q4
Mac Studio M2 Ultra (192GB)
Apple192 GB · Mac Studio
148.4 / 192 GBcomfortable · Q4
Mac Studio M3 Ultra (192GB)
Apple192 GB · Mac Studio
148.4 / 192 GBcomfortable · Q4
Mac Pro M2 Ultra (192GB)
Apple192 GB · Mac Pro
148.4 / 192 GBcomfortable · Q4
Instinct MI325X
AMD256 GB · Instinct CDNA 3+
181.3 / 256 GBcomfortable · Q5
Mac Studio M3 Ultra (256GB)
Apple256 GB · Mac Studio
181.3 / 256 GBcomfortable · Q5
Mac Studio M3 Ultra (512GB)
Apple512 GB · Mac Studio
263.8 / 512 GBcomfortable · Q8
Similar models
VRAM estimates: parameters x bits/8 plus margin. Real performance varies by engine, context length and batch size.
sync: 2026-05-12