DeepSeek

DeepSeek is one of the most aggressively permissive frontier-weight families we track: every model on this page is MIT licensed, and the range spans from dense 8B distills that fit in 12GB to a 862B MoE flagship whose Q4 weights alone exceed the VRAM of 4×96GB cards. The family splits into two lines — the R1 reasoning series (including three Qwen-based distills) and the V4 chat series — united by a signature MoE architecture that keeps active parameters small relative to total size.

All VRAM figures below use our standard estimate: Q4 weights + KV cache at 8K context + 1 GB overhead. MoE KV cache scales with active parameters, not total — which is why even the 862B Pro adds far less per-context-K than a dense model would.

Size ladder

• DeepSeek-R1-0528-Qwen3-8B (8.2B dense, 128K ctx) — Q4 ≈ 4.5 GB → any 12GB card with room to spare; the cheapest entry ticket to R1-style reasoning.
• DeepSeek-R1-Distill-Qwen-14B (14.8B dense, 128K ctx) — Q4 ≈ 8.1 GB → 12GB at Q4 with modest headroom; 16GB for longer context.
• DeepSeek-R1-Distill-Qwen-32B (32.8B dense, 128K ctx) — Q4 ≈ 18.0 GB → 24GB card (RTX 3090/4090 class); the biggest R1 distill that fits a single consumer GPU.
• DeepSeek-R1 (684.5B MoE, 37B active, 128K ctx) — Q4 ≈ 376.5 GB → multi-GPU server (4×96GB minimum); the original R1 reasoning flagship.
• DeepSeek-V4-Flash (158.1B MoE, 13B active, 1M ctx) — Q4 ≈ 87.0 GB → 96GB-class card; the V4 family's efficiency play with only 13B active per token.
• DeepSeek-V4-Pro (861.6B MoE, 49B active, 1M ctx) — Q4 ≈ 473.9 GB → 5×96GB server minimum; the family's maximum-quality V4 variant.

Family highlights

• All MIT, no asterisks. Every single DeepSeek model in our catalog uses the MIT license — commercial use, fine-tuning, redistribution, all unrestricted. This is rare among frontier-weight families and one reason GGUF quants of these models are everywhere.
• MoE across the board. The big models (R1, V4 Flash, V4 Pro) are MoEs with active parameters far smaller than total — 37B active inside 684.5B total for R1, 13B active inside 158B for V4 Flash. That means decode speed tracks the active count, not the total, once the model is loaded.
• 1M context on V4. Both V4 Flash and V4 Pro ship with a 1M-token context window — one of the longest we track. But remember that KV cache grows with context: stretching from 8K to 1M adds roughly 25 GB of KV on V4 Flash (13B active) and 98 GB on V4 Pro (49B active). Long-context deployments should budget accordingly.
• R1 distills bridge the gap. If you cannot afford 4×96GB for the full R1, the Qwen-based distills give you R1-style reasoning at consumer-friendly sizes. The 32B distill at Q4 on a 24GB card is the best quality-per-dollar reasoning model available today.

Which one should you pick

• 12GB VRAM — R1-Distill-Qwen-8B or 14B at Q4. The 8B is the zero-friction starting point; the 14B fits with careful context budgeting.
• 16GB VRAM — R1-Distill-14B at Q4 with comfortable headroom. The 8B distills are trivial at this tier.
• 24GB VRAM — R1-Distill-32B at Q4. This is the sweet spot for reasoning models on a single consumer GPU — near-frontier reasoning quality at a 24GB footprint.
• 96GB VRAM — V4 Flash at Q4 (≈87 GB) fits one card; this gets you a V4-generation chat model with 13B-active MoE speed. Full R1 does not fit.
• Multi-GPU (4×96GB+) — R1 at Q4 (≈377 GB) across 4 cards. V4 Pro needs 5×96GB or more at Q4.

For most single-GPU buyers, the R1 distills — especially the 32B on a 24GB card — are the practical DeepSeek experience. The full R1 and V4 models are datacenter-scale, but MIT licensing makes even those accessible to builders who own the hardware.

Links

• DeepSeek-R1-Distill-Qwen-32B — fit & speed estimates
• DeepSeek-V4-Flash — fit & speed estimates
• Best GPUs for reasoning models
• Best MoE models to run locally
• Guide: How much VRAM do local LLMs need?
• Guide: LLM quantization explained