Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarX: @victormustar→ original

DeepSeek updates V4-Flash: 82.7 on terminal tasks, up from 61.8

DeepSeek has updated its open V4-Flash model (304 billion parameters), and as a terminal agent it has improved markedly. On Terminal-Bench 2.1, where the model installs packages, edits configs and runs programs on its own, the score rose from 61.8 to 82.7 compared with the preview version. On Agents' Last Exam it scored 25.2 against 25.7 for the closed Opus-4.8. Teams building coding agents should consider running their own tasks on it.

AI-processed from X: @victormustar; edited by Hamidun News
DeepSeek updates V4-Flash: 82.7 on terminal tasks, up from 61.8
Source: X: @victormustar. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceX: @victormustar ↗
Editor-in-Chief
MIT licensethe weights can be downloaded and run in-house, including inside a product
DeepSWE (complex code-editing tasks)from 7.3 to 54.4 points
Key metric #3The gap with Opus-4.8 on Agents' Last Exam is just 0.5 points

Imagine an agent that receives the task "deploy the service and fix the failing tests" and then opens a terminal on its own, installs dependencies, edits configs and runs programs until everything works. Until now, agents like this have mostly been built on closed APIs, because open models lagged behind in this area. The update DeepSeek-V4-Flash-0731 narrows that gap: on Terminal-Bench 2.1 (a test of command-line work), the model rose from 61.8 to 82.7 points compared with the preview version. The figures were shared by Victor Mustar of Hugging Face in a post about AI's "open summer."

What exactly was released

DeepSeek-V4-Flash-0731 is an updated version of V4-Flash with a MoE (mixture of experts) architecture and a total size of 304 billion parameters. In a MoE model, only some of the "experts" inside the network process each token, so such a model is cheaper to run than a dense model of the same size. The source does not say how many parameters are active in this specific version.

For practitioners, the MIT license matters more. It is one of the most permissive licenses: the weights can be downloaded, deployed on your own servers, fine-tuned and built into a commercial product without separate approvals. For companies that do not want to send code and logs to an external cloud API, this is a serious argument.

What the numbers mean

The comparison is with the preview version of the same model, and the difference is significant.

  • Terminal-Bench 2.1 measures how reliably a model handles command-line tasks: installing packages, editing configs, running programs. The score rose from 61.8 to 82.7 points.
  • DeepSWE is a set of complex software development tasks that require understanding someone else's code and fixing it. Here the jump is sharper: from 7.3 to 54.4 points, roughly 7.5 times higher.
  • Agents' Last Exam is a demanding test for AI agents. DeepSeek-V4-Flash-0731 scored 25.2 points, while the closed Opus-4.8 scored 25.7. The difference is 0.5 points.

Gains of this size from a single update suggest the team deliberately strengthened agentic behavior: the ability to carry a multi-step task through to the end, not just produce a good chat reply. For an agent working in a terminal, that matters most.

What to do about it, and where to be careful

If you run a coding agent, a CI bot or an internal DevOps assistant, assemble a small set of your own real tasks and run DeepSeek-V4-Flash-0731 alongside your current model. If the gap on your tasks is small, savings on API costs and control over your data may outweigh it.

The caveats are these:

  • All figures come from a post on X, not from an independent audit. Results depend heavily on how the runs were configured: the toolset, the number of attempts, time limits.
  • A nearly equal score on Agents' Last Exam does not mean parity on every task. On your stack and your repositories, the picture may differ.
  • 304 billion parameters is a lot, even for a MoE. Running the model yourself requires a serious GPU cluster, not a single laptop.

The fact itself is still telling. The gap between open and closed agentic models has shrunk to the point where price, privacy and ease of deployment, not just quality, determine the choice. For those building agents, it is a reason to revisit the budget and check whether a paid closed API is needed for every task.

Story

Key Questions Answered

What exactly was released?

DeepSeek-V4-Flash-0731 is an updated version of V4-Flash with a MoE (mixture of experts) architecture and a total size of 304 billion parameters. In a MoE model, only some of the "experts" inside the network process each token, so such a model is cheaper to run than a dense model of the same size. The source does not say how many parameters are active in this specific version. For practitioners, the MIT license matters more.

What the numbers mean?

The comparison is with the preview version of the same model, and the difference is significant. Terminal-Bench 2.1 measures how reliably a model handles command-line tasks: installing packages, editing configs, running programs. The score rose from 61.8 to 82.7 points. DeepSWE is a set of complex software development tasks that require understanding someone else's code and fixing it. Here the jump is sharper: from 7.3 to 54.4 points, roughly 7.5 times higher. Agents' Last Exam is a demanding test for AI agents.

What to do about it, and where to be careful?

If you run a coding agent, a CI bot or an internal DevOps assistant, assemble a small set of your own real tasks and run DeepSeek-V4-Flash-0731 alongside your current model. If the gap on your tasks is small, savings on API costs and control over your data may outweigh it. The caveats are these: All figures come from a post on X, not from an independent audit. Results depend heavily on how the runs were configured: the toolset, the number of attempts, time limits.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…