DeepSeek V4 Flash: Official Release With Major Agent Performance Boost (2026)
Quick verdict
DeepSeek V4 Flash is a 284B MoE model with a 1M token context, 384K max output, and a large jump in agent performance over its predecessor. It is aimed at high-volume, cost-sensitive workloads that still need agentic capability.
What is DeepSeek V4 Flash?
DeepSeek V4 Flash is the latest addition to the DeepSeek model family, officially released 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 (official release) |
| 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 official release 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 (now publicly released) 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) | $0.007 off-peak / $0.014 peak |
| Input (cache miss) | $0.22 off-peak / $0.44 peak |
| Output | $0.66 off-peak / $1.32 peak |
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 | $0.22 | $0.66 | 1M |
| DeepSeek V4 Pro | $0.66 | $1.98 | 1M |
| GPT-4o | ~$2.50 | ~$10.00 | 128K |
Competitor pricing is approximate. Check official pricing pages for current rates.
⚠️ DeepSeek introduced peak/off-peak pricing on August 17, 2026. Peak hours are 01:00-04:00 and 06:00-10:00 UTC; peak rates are 2x off-peak rates.
Cache Hit Advantage
With cache-hit input at just $0.007 per 1M tokens off-peak—about 30x cheaper than cache-miss off-peak—developers who structure prompts to leverage context caching can still 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 and Codex support currently remain exclusive to V4 Flash — the official V4 Pro 0813 release does not have them yet.
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 off-peak pricing ($0.22/M input) and fast inference makes V4 Flash ideal for applications that can schedule workloads outside DeepSeek's peak hours.
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
- V4 Pro upgrade path: The official V4 Pro 0813 release (August 13, 2026) brings a major agent capability jump — Terminal Bench 2.1 at 87.9, DeepSWE at 62.7. Flash users with the most demanding tasks should evaluate Pro
- 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: Both V4 models are text-only per DeepSeek's official API documentation
Competitor Comparison
| Aspect | DeepSeek V4 Flash | Equivalent Tier |
|---|---|---|
| Price (output/M) | $0.66 off-peak | Lower than most 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 official release delivers dramatically improved agent capabilities, maintaining DeepSeek's position as the leader in affordable open-source AI. While V4 Flash remains text-only and the newly released V4 Pro 0813 is stronger, 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, retired on July 24, 2026, will not be needed anymore.
Related Articles
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
- Competitive pricing — from $0.22/M input and $0.66/M output (off-peak)
- 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
- V4 Flash is text-only — no image understanding (official docs confirm text-only for both V4 models)
- Smaller and less capable than the official V4 Pro 0813 release
- Peak/off-peak pricing coming soon (2x during peak hours)
- Smaller ecosystem compared to OpenAI or Anthropic
- Planned API price increase announced for the near future
Related Articles
GLM-5.3: Zhipu's Open-Source Programming Flagship — Complete Review (2026)
GLM-5.3 review: Zhipu AI's latest open-source flagship with 1M context, 128K output, +50% coding feel, CyberGym security leadership, and open weights expected in late August 2026.
DeepSeek: Complete Guide to the R1 & V4 AI Models (2026)
An in-depth review of DeepSeek—covering R1 reasoning, V4 Flash and V4 Pro (official release 0813), pricing, and real-world performance. Is DeepSeek still the best value AI in 2026?
Qwen3.8-27B: The 27B Open-Source Model That Packs Agentic Coding Into Your GPU
Qwen3.8-27B is a ~27B dense, Apache-2.0 multimodal model with strong agentic coding (DeepSWE 42.2, Terminal-Bench 73.0) and a ~17GB GGUF, made for consumer GPUs.
Qwen 3.8-Max: Alibaba's 2.4 Trillion Parameter Open-Source Flagship (2026)
Qwen 3.8-Max is here — 2.4T parameters, 1M context, open weights coming next week. Full review covering autonomous coding, agent benchmarks, pricing, and how it compares to GPT-5.6 and Fable 5.