Overview
DeepSeek-V4-Flash-0731 is the official release of DeepSeek V4 Flash, published on July 31, 2026 and superseding the preview. It keeps the structure of the preview and adds a speculative decoding module to the same checkpoint, with the agentic training that shows up across the benchmark table below.
| Specification | DeepSeek V4 Flash 0731 |
|---|---|
| Total parameters | 284B |
| Active parameters | 13B per token |
| Architecture | Mixture of experts, 256 routed experts with 6 active, 1 shared |
| Layers | 43 |
| Context window | 1,048,576 tokens |
| Modalities | Text input, text output |
| Reasoning | Effort levels low, high and max |
| Speculative decoding | DSpark module attached to the checkpoint |
| Release date | July 31, 2026 |
| License | MIT |
The routing is what makes a 284B model tractable: six of 256 experts run per token, so generation costs about what a 13B model costs while all 284B parameters stay resident in memory. The million-token context is the other headline, and it is the reason this model gets used on whole repositories rather than single files.
DeepSeek V4 Flash 0731 benchmarks
DeepSeek's launch numbers, from the model card, compare the 0731 release with the two V4 previews, with GLM-5.2 and with Claude Opus 4.8:
| V4 Flash 0731 | V4 Flash (Preview) | V4 Pro (Preview) | GLM-5.2 | Claude Opus 4.8 | |
|---|---|---|---|---|---|
Terminal Bench 2.1 Agentic terminal-use tasks | 82.7 | 61.8 | 72.1 | 81.0 | 85.0 |
NL2Repo Repo-level code generation | 54.2 | 39.4 | 38.5 | 48.9 | 69.7 |
Cybergym Security and exploit reasoning | 76.7 | 38.7 | 52.7 | - | 83.1 |
DeepSWE Agentic coding | 54.4 | 7.3 | 12.8 | 46.2 | 58.0 |
Toolathlon-Verified Long-horizon tool use | 70.3 | 49.7 | 55.9 | 59.9 | 76.2 |
Agents' Last Exam Frontier agentic tasks | 25.2 | 15.8 | 16.5 | 23.8 | 25.7 |
AutomationBench Office automation tasks | 25.1 | 10.8 | 12.8 | 12.9 | 27.2 |
The gains over the preview are large enough to change what the model is for: DeepSWE goes from 7.3 to 54.4. It also clears V4 Pro (Preview) on every row here while activating fewer parameters. Opus 4.8 stays ahead across the board, closest on Agents' Last Exam.
DeepSeek V4 Flash 0731 hardware requirements
The system requirement to check is memory. We quantized the model from the original weights and published the builds as AtomicChat/DeepSeek-V4-Flash-0731-GGUF.
| Memory | Build to pick | File size |
|---|---|---|
| 96 GB | AD-IQ1_M | 70.2 GB |
| 128 GB | AD-IQ2_S | 93.4 GB |
| 192 GB | AD-IQ3_S | 130.8 GB |
| 256 GB and up | AD-MXFP4 | 154.5 GB |
These sizes put the model on a large Mac Studio or a multi-GPU box rather than a laptop. On a PC you can keep the attention layers on a 24 GB card and offload the experts to system RAM, which works here because only 13B parameters are active per token. Holding the full 1M context needs room for the KV cache on top of the weights.
How to run DeepSeek V4 Flash 0731 in Atomic Chat
Atomic Chat is a free local app for macOS, Windows and Linux. It includes a Hugging Face model browser and a built-in chat, with no manual llama.cpp build required.
- Download Atomic Chat for your platform and open it.
- Search for DeepSeek-V4-Flash-0731 in the model browser and open Download Options.
- Pick the build that fits the memory you have, then start a chat.
The full walkthrough, with the quant table and the llama.cpp commands, is in our guide to running DeepSeek V4 Flash locally.
For the rest of the lineup see every DeepSeek model you can run locally, including DeepSeek-V4-Pro and the earlier V4 Flash preview.
License
DeepSeek-V4-Flash-0731 is released under the MIT license. It permits commercial use, modification, redistribution and private deployment as long as the copyright and license notice stay with the code.