FITMYLLM · JULY 28, 2026 · OPEN METHODOLOGY · COMMUNITY BENCHMARKS
LIVE · UPDATED DAILY
236B
Parameters (21B active)
Quantization Options Select your GPU for speed estimates Context length: 4K 8K 16K 32K 64K 128K
Quant Bits VRAM @ 16K Quality IQ2_XXS 2.38 84.8 GB
70.7 + 14.1 KV
low IQ2_M 2.93 101.0 GB
86.9 + 14.1 KV
low Q2_K 3.16 107.8 GB
93.7 + 14.1 KV
low IQ3_XXS 3.25 110.4 GB
96.4 + 14.1 KV
low IQ3_XS 3.5 117.8 GB
103.7 + 14.1 KV
low Q3_K_S 3.64 121.9 GB
107.9 + 14.1 KV
low IQ3_M 3.76 125.5 GB
111.4 + 14.1 KV
low Q3_K_M 4 132.6 GB
118.5 + 14.1 KV
low Q3_K_L 4.3 141.4 GB
127.3 + 14.1 KV
moderate IQ4_XS 4.46 146.1 GB
132.1 + 14.1 KV
moderate Q4_K_S 4.67 152.3 GB
138.3 + 14.1 KV
moderate Q4_K_M 4.89 158.8 GB
144.7 + 14.1 KV
good Q5_K_S 5.57 178.9 GB
164.8 + 14.1 KV
good Q5_K_M 5.7 182.7 GB
168.6 + 14.1 KV
good Q6_K 6.56 208.1 GB
194.0 + 14.1 KV
excellent Q8_0 8.5 265.3 GB
251.2 + 14.1 KV
lossless FP16 16 486.6 GB
472.5 + 14.1 KV
lossless
Select your GPU above to see speed estimates and compatibility for each quantization.
Too big for a single GPU — plan a multi-GPU deployment
Even the lightest quant needs ~85 GB. Size GPUs, replicas, TCO and scaling for a production setup. Open in Enterprise →
▸ READY TO RUN THIS? RENT BY THE HOUR
RENT A GPU AND RUN DEEPSEEK-V2.5 236B NOW
Spin up an A100 / H100 / 4090 in ~60s. Pay by the second. Cancel anytime.
Community Ratings Chat Coding Reasoning Creative Vision Roleplay Agentic
Loading ratings...
Run this model IQ2_XXS — 70.7 GB VRAM IQ2_M — 86.9 GB VRAM Q2_K — 93.7 GB VRAM IQ3_XXS — 96.4 GB VRAM IQ3_XS — 103.7 GB VRAM Q3_K_S — 107.9 GB VRAM IQ3_M — 111.4 GB VRAM Q3_K_M — 118.5 GB VRAM Q3_K_L — 127.3 GB VRAM IQ4_XS — 132.1 GB VRAM Q4_K_S — 138.3 GB VRAM Q4_K_M — 144.7 GB VRAM Q5_K_S — 164.8 GB VRAM Q5_K_M — 168.6 GB VRAM Q6_K — 194.0 GB VRAM Q8_0 — 251.2 GB VRAM FP16 — 472.5 GB VRAM
Ollama llama.cpp vLLM LM Studio KoboldCpp Jan Docker
▸ Easiest way to get started · Beginners
DOCS ↗ curl -fsSL https://ollama.com/install.sh | shCOPY
$ ollama run deepseek-v2.5:236bCOPY
Tag may need adjustment — check ollama.com/library/deepseek-v2.5 for available tags.
▸ SETUP GUIDE >_
Auto-setup with fitmyllm CLI Detects your GPU, recommends the best model, downloads it, and starts chatting — zero config. Benchmarks your speed and contributes anonymous data to improve predictions.
Auto-detect GPU Live tok/s in chat Speed benchmarks 9 inference engines
GPUs that can run this model At Q4_K_M quantization. Sorted by minimum VRAM.
Find the best GPU for DeepSeek-V2.5 236B
Build Hardware for DeepSeek-V2.5 236B DeepSeek V2.5 — merged chat and coder. Better writing and coding.
Read full model card ▸ COLOPHON FITMYLLM · INDEPENDENT · DATA-DRIVEN
FITMYLLM · EST. 2025 · © 2026
RECOMMENDATIONS FROM PUBLISHED MATH, CORRECTED BY THE COMMUNITY — 30.
▸ SPEC SHEET
DeepSeek-V2.5 236B — 236B MoE. ▸ SPECIFICATIONS
PARAMETERS 236B (21B active)
ARCHITECTURE Mixture of Experts
CONTEXT LENGTH 128K tokens
CAPABILITIES chat, coding
RELEASE DATE 2024-09-05
PROVIDER DeepSeek
FAMILY deepseek ▸ VRAM REQUIREMENTS
QUANT BPW VRAM QUALITY IQ2_XXS 2.38 70.7 GB 65% IQ2_M 2.93 86.9 GB 75% Q2_K 3.16 93.7 GB 78% IQ3_XXS 3.25 96.4 GB 82% IQ3_XS 3.5 103.7 GB 84% Q3_K_S 3.64 107.9 GB 85% IQ3_M 3.76 111.4 GB 86% Q3_K_M 4 118.5 GB 88% Q3_K_L 4.3 127.3 GB 90% IQ4_XS 4.46 132.1 GB 92% Q4_K_S 4.67 138.3 GB 93% Q4_K_M 4.89 144.7 GB 94% Q5_K_S 5.57 164.8 GB 96% Q5_K_M 5.7 168.6 GB 96% Q6_K 6.56 194.0 GB 97% Q8_0 8.5 251.2 GB 100% FP16 16 472.5 GB 100%
§ 01 BENCHMARK SCORES
HumanEval 87.6
MMLU-PRO 71.0
MATH 74.7
IFEval 80.0
BBH 82.0
BigCodeBench 40.4
Arena Elo 1273.0
AA Intelligence 12.3
§ 02 RUN COMMAND
Run DeepSeek-V2.5 236B locally with Ollama — needs 144.7 GB VRAM at Q4_K_M:
$ ollama run deepseek-v2.5:236b
§ 03 COMPATIBLE GPUs
13 @ Q4_K_M Feedback 