Unsloth/DeepSeek-V4-Pro-0813-GGUF

asoberbeck2 pts0 comments

unsloth/DeepSeek-V4-Pro-0813-GGUF · Hugging Face

Log In<br>Sign Up

In progress will take a while

See below for instructions on the Flash model for now:

Read our How to Run DeepSeek-V4 Guide!

Unsloth Dynamic 2.0 achieves superior accuracy & outperforms other leading quants.

See our DeepSeek-V4 guide for quantization analysis and run instructions.

DeepSeek-V4-Pro-0813 is a 1.57T parameter model with 48B active parameters per token, so it needs substantially more memory than DeepSeek-V4-Flash-0731.

For DeepSeek-V4-Flash-0731 GGUFs, see unsloth/DeepSeek-V4-Flash-0731-GGUF.

Quants are uploaded to this repository as they finish converting.

Model architecture

DeepSeek-V4-Pro-0813<br>DeepSeek-V4-Flash-0731

Total parameters<br>1.57T<br>745B

Active parameters per token<br>48B<br>26B

Layers<br>61<br>43

Hidden size<br>7168<br>4096

Attention heads<br>128<br>64

Routed experts<br>384<br>256

Active experts per token

Shared experts

Expert FFN size<br>3072<br>2048

Context length<br>1M<br>1M

DeepSeek-V4-Pro-0813

Technical Report

Introduction

DeepSeek-V4-Pro-0813 is the official release of DeepSeek-V4-Pro , superseding the preview version, with greatly enhanced agentic capabilities and performance improvements that are especially pronounced in production environments. It is built on the DeepSeek-V4-Pro (Preview) model structure, with a DSpark speculative decoding module attached.

DeepSeek-V4-Pro-0813 outperforms DeepSeek-V4-Pro (Preview) on the benchmarks listed below, and is broadly competitive with the strongest proprietary models available.

Benchmark<br>DeepSeek-V4-Pro-0813<br>DeepSeek-V4-Flash-0731<br>DeepSeek-V4-Pro (Preview)<br>DeepSeek-V4-Flash (Preview)<br>GLM-5.2<br>Kimi K3<br>Opus-4.8<br>Fable-5 (w/ fallback)

HLE (wo / w tools)<br>42.7 / 60.0<br>37.8 / 51.5<br>37.7 / 48.2<br>34.8 / 45.1<br>40.5 / 54.7<br>43.5 / 56.0<br>49.8 / 57.9<br>53.3 / 63.0

Terminal Bench 2.1<br>87.9<br>82.7<br>72.1<br>61.8<br>81.0<br>88.3<br>85.0<br>88.0

NL2Repo<br>61.5<br>54.2<br>38.5<br>39.4<br>48.9<br>69.7

Cybergym<br>83.3<br>76.7<br>52.7<br>38.7<br>80.0<br>78.3<br>83.1

DeepSWE<br>62.7<br>54.4<br>12.8<br>7.3<br>46.2<br>67.5<br>58.0<br>70.0

Toolathlon-Verified<br>74.1<br>70.3<br>55.9<br>49.7<br>59.9<br>76.5<br>76.2<br>77.9

Agents' Last Exam<br>25.7<br>25.2<br>16.5<br>15.8<br>23.8<br>27.6<br>25.7

AutomationBench (Public)<br>31.8<br>25.1<br>12.8<br>10.8<br>12.9<br>30.8<br>27.2<br>29.1

DSBench-FullStack †<br>71.1<br>68.7<br>41.8<br>37.0<br>61.8<br>73.7<br>71.6<br>77.2

DSBench-Hard †<br>67.2<br>59.6<br>31.1<br>25.8<br>54.5<br>63.0<br>71.7<br>68.3

Notes:

For the code-agent tasks among the public benchmarks above, DeepSeek-V4-Pro-0813 is evaluated with the minimal mode of DeepSeek Harness as the agent framework, using the max reasoning effort level with temperature = 1.0, top_p = 0.95.

† DSBench-FullStack is an internal full-stack development test set; DSBench-Hard is an internal test set of difficult coding-agent problems.

Chat Template

This release does not include a Jinja-format chat template. Instead, DeepSeek provide a dedicated encoding folder with Python scripts and test cases demonstrating how to encode messages in OpenAI-compatible format into input strings for the model, and how to parse the model's text output.

The reasoning_effort parameter supports three levels, low, high and max, which control how much deliberation the model spends before answering.

Sampling parameters

For local deployment, DeepSeek recommend setting the sampling parameters to temperature = 1.0, with top_p = 0.95 for agentic scenarios and top_p = 1.0 otherwise. For the high and max reasoning effort levels, they recommend a maximum output length of 384K tokens.

License

This repository and the model weights are licensed under the MIT License.

Citation

@misc{deepseekai2026deepseekv4,<br>title={DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence},<br>author={DeepSeek-AI},<br>year={2026},

Downloads last month -

Downloads are not tracked for this model. How to track

Inference Providers NEW<br>This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for unsloth/DeepSeek-V4-Pro-0813-GGUF<br>Base model

deepseek-ai/DeepSeek-V4-Pro-0813

Quantized<br>(1)<br>this model

Paper for unsloth/DeepSeek-V4-Pro-0813-GGUF<br>Paper • 2606.19348 • Published Apr 26 • 33

deepseek model flash unsloth gguf parameters

Related Articles