DeepSeek V4 Flash: Public Beta Launches With Major Agent Performance Boost (2026)
What is DeepSeek V4 Flash?
DeepSeek V4 Flash is the latest addition to the DeepSeek model family, officially entering public beta on July 31, 2026. It's the smaller, faster, and more cost-effective counterpart to the flagship DeepSeek V4 Pro, designed specifically for agentic workflows and high-throughput applications.
The model is built on a Mixture-of-Experts (MoE) architecture with 284B total parameters, of which only 13B are activated per token. This design allows it to deliver strong performance while keeping inference costs and latency remarkably low.
Unlike the preview version released in April 2026, the V4-Flash-0731 release underwent additional post-training optimization that dramatically improved its agent capabilities—bringing benchmark results that in some areas exceed even the V4 Pro preview.
Architecture & Specifications
| Specification | Value |
|---|---|
| Architecture | Mixture-of-Experts (MoE) |
| Total Parameters | 284B |
| Active Parameters per Token | 13B |
| Context Length | 1M tokens (native) |
| Max Output | 384K tokens |
| License | MIT |
| Release Date | July 31, 2026 (public beta) |
| Version | DeepSeek-V4-Flash-0731 |
Attention Innovation
DeepSeek V4 Flash uses a hybrid attention mechanism combining:
- Token-wise Compression — Efficiently compresses information at the token level
- DeepSeek Sparse Attention (DSA) — Reduces compute and memory costs for ultra-long sequences
This architecture allows the model to maintain its full 1M context window while keeping inference costs dramatically lower than competing models with similar context lengths.
Agent Performance — The Headline Feature
The defining update in the V4 Flash public beta is its massively improved agent capabilities. The post-training optimization delivered benchmark results that significantly exceed the preview version:
| Benchmark | V4 Flash (Official) | Notes |
|---|---|---|
| Terminal Bench 2.1 | 82.7 | Agent terminal task performance |
| NL2Repo | 54.2 | Natural language to repository |
| Cybergym | 76.7 | Cybersecurity agent tasks |
| DeepSWE | 54.4 | Software engineering agent |
| Toolathlon verified | 70.3 | Tool use benchmarking |
| Agent Last Exam | 25.2 | Advanced agent challenges |
| Automation Bench (Public) | 25.1 | Automation task performance |
| DSBench-FullStack | 68.7 | Internal full-stack development |
| DSBench-Hard | 59.6 | Internal coding agent hard problems |
These benchmarks were tested using DeepSeek Harness minimal mode with max effort level, top_p=0.95, and temperature=1.0.
For context, the agent benchmarks of V4 Flash now exceed those of the V4 Pro preview in several categories, making it a compelling choice for developers building AI-powered agent systems.
Pricing & Value
DeepSeek V4 Flash continues DeepSeek's strategy of aggressive pricing:
| Billing Item | Price (per 1M tokens) |
|---|---|
| Input (cache hit) | $0.0028 |
| Input (cache miss) | $0.14 |
| Output | $0.28 |
Cost Comparison
Compared to other frontier models, V4 Flash offers exceptional value based on official pricing:
| Model | Input (per 1M) | Output (per 1M) | Context |
|---|---|---|---|
| DeepSeek V4 Flash | $0.14 | $0.28 | 1M |
| DeepSeek V4 Pro | $0.435 | $0.87 | 1M |
| GPT-4o | ~$2.50 | ~$10.00 | 128K |
Competitor pricing is approximate. Check official pricing pages for current rates.
⚠️ Note: DeepSeek will soon introduce peak/off-peak pricing. During peak hours (9:00–12:00 and 14:00–18:00 Beijing Time, UTC+8), prices will be 2x the regular rates.
Cache Hit Advantage
With the cache hit rate offering input at just $0.0028 per 1M tokens—a 50x discount over cache miss pricing—developers who structure their prompts to leverage context caching can achieve dramatically lower effective costs.
Key Features
Dual Mode (Thinking + Non-Thinking)
V4 Flash supports both modes in a single model:
- Thinking mode (default): Deeper reasoning for complex tasks, similar to the legacy deepseek-reasoner
- Non-thinking mode: Faster responses for simpler tasks, similar to the legacy deepseek-chat
Responses API & Codex Integration
V4 Flash natively supports the Responses API format and is specifically adapted for Codex integration. This makes it straightforward to use as the underlying model for agentic coding workflows. Note that the Responses API is currently only available for V4 Flash, not yet for V4 Pro.
Tool Use & Function Calling
The model supports:
- JSON output mode
- Tool calls (function calling)
- FIM (Fill-in-the-Middle) completion
- Chat prefix completion (Beta)
1M Context Window
With native 1M token context support and up to 384K tokens of output, V4 Flash can handle entire codebases, long conversation histories, and complex multi-turn agent tasks in a single session.
Use Cases
Agentic Coding
V4 Flash's strong agent benchmarks make it an excellent choice for AI-powered coding agents. Its DeepSWE score of 54.4 and Terminal Bench of 82.7 indicate robust performance on real-world software engineering tasks.
High-Throughput API Applications
The combination of low pricing ($0.14/M input) and fast inference makes V4 Flash ideal for applications that need to process large volumes of requests without breaking the budget.
Long-Context Analysis
With 1M token context and efficient attention mechanisms, V4 Flash can process entire codebases, long legal documents, or extensive research materials in a single pass.
Cost-Sensitive Deployments
For teams that need capable AI without the premium pricing of leading competitors, V4 Flash offers a compelling alternative—especially for workloads where cache hit rates can be optimized.
Limitations
- Public beta status: The API is still in public beta and may not be suitable for all production workloads
- V4 Pro pending: The full V4 Pro release is expected to follow, and Flash users may want to evaluate Pro for the most demanding tasks
- Peak pricing coming: The upcoming peak/off-peak pricing model could increase costs during business hours
- Ecosystem maturity: While growing, the DeepSeek ecosystem is smaller than OpenAI's or Anthropic's
- No multimodal support: V4 Flash is text-only; image or video understanding is not available
Competitor Comparison
| Aspect | DeepSeek V4 Flash | Equivalent Tier |
|---|---|---|
| Price (output/M) | $0.28 | Significantly lower than competitors |
| Context | 1M | Among the largest available |
| Agent performance | Excellent | Competitive with leading models |
| Open source | ✅ MIT | Rare at this capability level |
| Thinking mode | ✅ | Available in single model |
V4 Flash's main differentiator is its combination of strong agent performance, ultra-low pricing, and open-source availability—a combination that few competitors match.
Who Should Use DeepSeek V4 Flash?
- Agent developers: The strong agent benchmarks make V4 Flash a top choice for building AI coding agents and automation tools
- Cost-conscious teams: If you need capable AI at a fraction of the cost of premium models, V4 Flash is hard to beat
- Open-source advocates: The MIT-licensed weights allow self-hosting, fine-tuning, and customization
- High-volume API users: The low per-token pricing and cache hit discounts make V4 Flash economical at scale
Summary
DeepSeek V4 Flash represents a significant step forward for cost-effective AI. The public beta release delivers dramatically improved agent capabilities, maintaining DeepSeek's position as the leader in affordable open-source AI. While it's still in beta and the V4 Pro release is pending, V4 Flash is already a compelling option for developers who need strong agent performance without the premium price tag.
For teams already using DeepSeek's API, the transition is straightforward—simply set the model name to deepseek-v4-flash and the legacy deepseek-chat and deepseek-reasoner names will continue to route to V4 Flash until fully retired.
Frequently Asked Questions
What is DeepSeek V4 Flash?+
How much does DeepSeek V4 Flash cost?+
How does DeepSeek V4 Flash compare to V4 Pro?+
Is DeepSeek V4 Flash open source?+
What happened to deepseek-chat and deepseek-reasoner?+
Pros
- Exceptional agent performance — Terminal Bench 82.7, DeepSWE 54.4
- Extremely low pricing — $0.14/M input, $0.28/M output
- 1M native context window with 384K max output
- Open source (MIT License) — weights available for self-hosting
- Responses API support with native Codex integration
- Dual mode: thinking and non-thinking in one model
Cons
- Still in public beta, not yet production stable
- V4 Pro version not yet released
- Peak/off-peak pricing coming soon (2x during peak hours)
- Smaller ecosystem compared to OpenAI or Anthropic
- No multimodal capabilities in this release
Related Articles
Kimi K3: Moonshot's 2.8T Frontier AI Model — Complete Review 2026
In-depth review of Kimi K3, Moonshot AI's 2.8-trillion-parameter open-weight model. Features, benchmarks, pricing ($3/$15 per MTok), and how it compares to GPT-5.5, Claude Opus 4.8, and DeepSeek.
DeepSeek: Complete Guide to the Open-Source AI Reasoning Model (2026)
An in-depth review of DeepSeek—covering its R1 reasoning model, coding abilities, pricing (fraction of OpenAI), and real-world performance. Is it the best value AI in 2026?
Grok 4.5: Complete Guide to xAI's Coding-Focused Frontier Model (2026)
In-depth review of Grok 4.5—features, pricing, benchmarks, and real-world use cases. Learn how xAI's coding-focused model compares to Claude, GPT, and Gemini.
Claude Code: Anthropic's AI Coding Agent — Complete Review 2026
In-depth review of Claude Code, Anthropic's terminal-based AI agent for software development. Features, capabilities, pricing, and how it compares to Copilot, Cursor, and Devin.