DeepSeek-V4 Guide: VRAM Requirements, GGUF Builds

September 26, 2026September 27, 2026

About This Model

DeepSeek-V4 is DeepSeek’s MoE model series, built to be “highly efficient" enough to handle million-token contexts routinely. The centerpiece of this page, DeepSeek-V4-Pro-0813, is the official release of the flagship DeepSeek-V4-Pro and replaces the earlier preview. It keeps the preview’s architecture and adds a DSpark module for speculative decoding.

According to the technical report’s abstract, DeepSeek-V4-Pro is an MoE model with 1.6T total parameters and 49B active per token, pre-trained on more than 32T tokens. It combines two compressed attention mechanisms (CSA and HCA) for long-context efficiency. At a one-million-token context, it needs only 27% of the per-token inference FLOPs and 10% of the KV cache of the previous-generation DeepSeek-V3.2. Reasoning depth can be set to low, high or max with reasoning_effort.

What Makes It Stand Out

1. Agentic performance improved sharply over the preview. An excerpt from the publisher’s comparison table (bold marks the best value in each row):

• It beats the preview on every row. DeepSWE rose from 12.8 to 62.7 and AutomationBench from 12.8 to 31.8.
• On many agentic tasks it stands level with the top models. On Terminal Bench 2.1 (terminal work) it is within 0.4 points of Kimi K3 and Fable-5, and it has the highest scores in the table on Cybergym and AutomationBench.

Caveats and Weaknesses

• It still trails the top models on hard knowledge questions and repository generation. On HLE without tools it scores 42.7, 10.6 points behind Fable-5 (53.3). It is 8.2 points behind Opus-4.8 on NL2Repo and 7.3 points behind Fable-5 on DeepSWE.
• Across the full table in the model card, it is lower than Kimi K3 on 6 of the 10 figures where both have a value. It cannot be called the best open-weight model outright.
• Watch the evaluation conditions. The code-agent tasks were measured with DeepSeek’s own agent framework (DeepSeek Harness). DSBench-FullStack and DSBench-Hard (omitted above) are DeepSeek’s internal test sets and cannot be reproduced by third parties under the same conditions.

Running It Locally

• This is not a model for consumer GPUs. The model card’s vLLM example runs on a single node with four GB300s, and the memory requirements this site computes from the distributed files, shown in the table below, are data-center scale.
• This release does not include a Jinja-format chat template, which many tools rely on. Instead, the encoding folder in the repository provides Python scripts that build the input and parse the output.
• In vLLM and SGLang, DSpark speculative decoding is enabled with a single flag; no separate draft model is needed.
• For high and max reasoning, a maximum output length of 384K tokens is recommended.
• The license is MIT, which allows commercial use.
• This family also covers NVIDIA’s NVFP4 build, DeepSeek-V4-Pro-0813-nvfp4-DSpark. The successor, DeepSeek-V4.1, has its own page.

Sources: deepseek-ai/DeepSeek-V4-Pro-0813 model card and the technical report’s abstract (arXiv:2606.19348), as of 2026-09-26. Benchmark figures are as published by DeepSeek.

Our Coverage and Data

Everything Local Model Watch has published about the DeepSeek-V4 family: 2 article(s) covering the base model and its fine-tunes, plus converted builds we tracked after publication. Memory requirements below are computed by this site from file sizes, not quoted from model cards. Part of our model family index.

At a Glance

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 1650.5B parameters

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

Runs in Ollama, LM Studio and llama.cpp via a converted build.

The publisher ships safetensors, but unsloth/DeepSeek-V4-Pro-0813-GGUF provides a GGUF build you can use.

License — mit (Commercial use allowed): Permits commercial use, modification and redistribution, provided the copyright notice and license text are retained.

Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Quantized and Converted Variants

File sizes of each build:

• Available builds in unsloth/DeepSeek-V4-Pro-0813-GGUF: Q4_K_XL 791.3GB / Q8_K_XL 813.5GB
• Available builds in unsloth/DeepSeek-V4-Pro-0813: FP8 831.4GB
• Available builds in DevQuasar/deepseek-ai.DeepSeek-V4-Pro-0813-GGUF: Q2_K 530.3GB / Q3_K_M 697.0GB / Q4_K_M 885.6GB

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files.

Articles (the family’s own models first, then newest)

Repositories

• deepseek-ai/DeepSeek-V4-Pro-0813
• nvidia/DeepSeek-V4-Pro-0813-nvfp4-DSpark

Last updated 2026-09-27 (JST). The explanation at the top of this page was written with the help of AI from the primary sources it cites. The tables and lists under “Our Coverage and Data" are assembled by code from our article log.