Autor: Jakub Rusinowski · Ostatnia aktualizacja: 21 lipca 2026
PREVIEW (June 2026, specs unverified). Experimental sparse-attention snapshot of DeepSeek V3.2 (already present in Hugging Face Transformers). ~685B/37B MoE, MIT, 128K context — a research preview of the V3.2 line (which is itself slated for retirement in favour of V4). Confirm the exact served variant against the Hugging Face model card.
| DeepSeek V3.2-Exp 685B | Min 414 GB VRAM · Q4_K_M · 128,000 ctx · |
The cheapest GPU that runs DeepSeek V3.2-Exp locally (min 414 GB VRAM) is the Apple M3 Ultra (512 GB).
Install Ollama then run: ollama run
Minimum VRAM: 414 GB. For best results use Q4_K_M quantization.
DeepSeek V3.2-Exp needs about 414 GB VRAM at Q4_K_M quantization for its smallest variant. Variants: DeepSeek V3.2-Exp 685B (414 GB, Q4_K_M). On Apple Silicon, unified memory counts toward this requirement.
DeepSeek V3.2-Exp's smallest variant needs about 414 GB, which exceeds a single RTX 4090 (24 GB). Use multiple GPUs, a higher-VRAM card, or Apple Silicon with large unified memory.
Q4_K_M is the best balance of quality and VRAM for DeepSeek V3.2-Exp in most cases. Choose Q8_0 for near-lossless quality if you have spare VRAM, or smaller quants (Q3/Q2) only when memory is tight.
Install Ollama, then run: ollama run . This downloads DeepSeek V3.2-Exp and starts a local, OpenAI-compatible endpoint — no internet connection is needed after the initial download.