Agent capability upgraded
Official scores rose across terminal, repository, tool-use, automation, and full-stack Agent evaluations.
Official release notes
English
Release desk · July 31, 2026
Use DeepSeek V4 Flash in Tabbit
The official V4 Flash API is now in public beta with a major Agent upgrade. Put that speed to work inside a browser with page context, model switching, and Agent Mode.
What changed
A stronger Agent model, not a new architecture
DeepSeek says V4-Flash-0731 keeps the preview architecture and size. The leap comes from re-post-training focused on Agent work.
- Context
- 1M tokens
- Maximum output
- 384K tokens
- Modes
- Thinking + non-thinking
Responses API is native
V4 Flash now speaks the Responses API directly and is specifically adapted for Codex-style Agent clients.
Same model name
Existing API callers keep using deepseek-v4-flash; the official endpoint now serves the latest release.
Fast path
From release note to working browser in three steps
Skip the local runtime and wiring. Model availability can vary by Tabbit edition, region, and rollout.
- 01
Install Tabbit
Download the macOS or Windows build that matches your region and open a new Tabbit window.
- 02
Choose DeepSeek V4 Flash
Open the model selector and choose DeepSeek V4 Flash when it appears in your edition.
- 03
Add context or run Agent
Reference pages, tab groups, screenshots, or files with @, or hand off a multi-step browser task to Agent Mode.
Browser-shaped work
Agent capability matters when the model can act
A benchmark score is abstract. In Tabbit, Agent Mode can turn a request into visible browser steps—reading the page, operating a web app, and keeping the work in its own tab group.
- Current pages and files become explicit context
- Execution steps stay visible for review
- Switch models without rebuilding the workflow
Official 0731 results
The Agent jump, in numbers
Selected scores reported by DeepSeek for the official V4 Flash release. Higher is better within each benchmark.
82.7
Terminal Bench 2.1
70.3
Toolathlon verified
68.7
DSBench FullStack
54.4
DeepSWE
54.2
NL2Repo
25.2
Agent Last Exam
DeepSeek tested public Code Agent tasks with its forthcoming Harness minimal mode, max effort, top_p 0.95, and temperature 1.0. DSBench FullStack is an internal test.
Choose your path
Tabbit, official API, or self-hosting?
| Decision | Tabbit | Official API | Self-host |
|---|---|---|---|
| Setup | Install and select a model | Create a key and configure a client | Prepare large weights, hardware, and runtime |
| Web context | Pages, tabs, screenshots, and files via @ | You build the context pipeline | You build the context pipeline |
| Agent surface | Built-in browser Agent Mode | Bring or build an Agent client | Bring or build an Agent stack |
| Best fit | Use the model while browsing | Build a product or integration | Control infrastructure and weights |
Release boundaries
What the July 31 announcement does—and does not—change
API release only
DeepSeek says the V4 Pro API and its own App/Web models were unchanged at announcement time.
Public beta
Treat behavior, limits, and availability as release-stage details that can still change.
Tabbit rollout varies
The models shown in Tabbit depend on edition, region, account, and rollout. Check the in-product selector.
FAQ
DeepSeek V4 Flash in Tabbit: common questions
01What did DeepSeek release on July 31, 2026?
+
DeepSeek put the official V4 Flash API into public beta as version DeepSeek-V4-Flash-0731, with stronger Agent post-training and native Responses API support.
02How do I use DeepSeek V4 Flash in Tabbit?
+
Install Tabbit, open its model selector, and choose DeepSeek V4 Flash when available in your edition. Then chat with page context or start Agent Mode.
03Do I need a DeepSeek API key in Tabbit?
+
Tabbit exposes supported models through its own model selector, so you do not configure the official DeepSeek API inside the normal browser flow. Availability and access terms can vary.
04Is the DeepSeek web app also running the 0731 model?
+
Not according to the July 31 release note. DeepSeek explicitly said its App/Web models were unchanged at announcement time.
05Is V4-Flash-0731 a different architecture from the preview?
+
No. DeepSeek says the architecture and size are unchanged; the official release was re-post-trained.
06Why use Tabbit instead of the API?
+
Use Tabbit when your work starts in the browser: pages, tab groups, screenshots, and files can become context, while Agent Mode can perform the web steps. Use the API when building your own product.
Tabbit × DeepSeek V4 Flash
Put DeepSeek V4 Flash inside a real browser workflow
Download Tabbit for macOS or Windows, check the model selector, and turn the new Agent capability into work you can see.