Can You Run DeepSeek V4 Locally?
By Lefi Abdelmonem · LinkedIn ↗
Author · AI Local Check · Published July 17, 2026 · Updated August 4, 2026
Updated August 4, 2026 — the new V4-Flash-0731 revision is out, and we measured it. DeepSeek shipped a refreshed V4-Flash-0731 (MIT, 304B), now the #2 trending model on the hub. The takeaway has a twist: it's still far too big for a consumer GPU, but at its smallest quantization it has quietly crossed into 128 GB unified-memory territory — the same tier as Hy3 and Inkling-Small. Jump to the measured numbers.
DeepSeek V4 is one of the most talked-about open-weight releases of 2026 — and unlike Qwen or Llama at consumer sizes, the honest answer to "can I run it locally?" is: not on a normal gaming GPU, though the smallest Flash build now fits a high-memory workstation. Here's exactly why, with the real numbers, and which DeepSeek models you can run on an ordinary card.
Why DeepSeek V4 doesn't fit a consumer GPU
DeepSeek V4 ships in two sizes, both Mixture-of-Experts (MoE) models under the MIT license:
- DeepSeek V4-Flash — about 304 billion total parameters (latest 0731 revision), with a 1,048,576-token (1M) context.
- DeepSeek V4-Pro — the flagship, roughly 1.6 trillion parameters.
Even the smaller "Flash" model is enormous. Measured from its real GGUF files, DeepSeek V4-Flash-0731 ranges from about 78 GB at a 1-bit quantization to 152 GB at 8-bit:
| Quantization | Memory to run | Fits a 128 GB machine? |
|---|---|---|
| IQ1_S (~1-bit) | ≈ 78 GB | ✅ yes |
| Q2_K_XL (~2-bit) | ≈ 91 GB | ✅ yes |
| Q3_K_M (~3-bit) | ≈ 120 GB | tight |
| Q4_K_XL (~4-bit) | ≈ 146 GB | ❌ needs more |
| Q8 (8-bit) | ≈ 152 GB | ❌ server |
On a single 24 GB consumer GPU, none of these fit in VRAM — the model spills into system RAM and crawls (offloading a 300B+ model runs, but far too slowly to be pleasant). What changed with the 0731 build is the low end: at 1–2 bit (78–91 GB) it now fits a 128 GB unified-memory machine — an Apple Silicon box, an AMD Strix Halo, or an NVIDIA RTX Spark — just like Tencent's Hy3 (~87 GB) and Inkling-Small (~71 GB).
Which DeepSeek models CAN you run locally?
The good news: DeepSeek also releases much smaller distilled models that carry a lot of its reasoning ability into sizes that fit a normal GPU. These are open-weight and available as GGUF:
- DeepSeek-R1-Distill-Llama-8B — ~8B parameters; runs fully on an 8 GB GPU at Q6_K_L (about 7.7 GB).
- DeepSeek-R1-Distill-Qwen-7B — another 8 GB-friendly option.
- DeepSeek-R1-Distill-Qwen-1.5B — tiny, runs on almost anything (even CPU-only).
- Larger 14B, 32B and 70B distills exist for 16–48 GB cards.
These give you DeepSeek-style reasoning locally, privately, and for free — without the 78+ GB requirement of the full V4-Flash. See the best local reasoning models for how they pair with your GPU.
How to check what fits your GPU
Memory needs depend on the model size, quantization and your context length. Rather than guess, search any DeepSeek model to see its exact RAM/VRAM per quantization, or pick your graphics card to see which models fit. New to quantization? Read GGUF quantization explained, and the best LLM for your VRAM for size-by-size picks.
The bottom line
DeepSeek V4-Flash-0731 is a big-memory model — 304B, measured at 78–152 GB depending on quantization. It won't run on a gaming GPU, but at 1–2 bit it now fits a 128 GB unified-memory machine, and the flagship V4-Pro (~1.6T) stays data-center-only. For an ordinary card, DeepSeek's R1 distills (from 1.5B to 70B) run beautifully. Check exactly what your PC can run and pick the largest DeepSeek model that fits your hardware.