Key Takeaways
- DeepSeek V4 Flash (build V4-Flash-0731, released July 31, 2026) is a 284B-parameter Mixture-of-Experts model that activates only 13B parameters per forward pass. It uses the same architecture as the V4-Flash-Preview from April 2026. Only post-training weights were updated in the 0731 build.
- V4-Flash-0731 outscored DeepSeek’s own V4-Pro-Preview on all nine published agent and coding benchmarks. Terminal Bench 2.1: 82.7 (Flash) versus 72.1 (Pro-Preview). DeepSWE: 54.4 (Flash) versus 7.3 (Flash Preview). DSBench-FullStack: 68.7 (Flash) versus 37.0 (Flash Preview). No price change accompanied the post-training update.
- Pricing is $0.14 per million input tokens and $0.28 per million output tokens. Cache-hit input drops to $0.0028 per million tokens, a roughly 50x reduction for repeated-prompt workloads. Both V4 Flash and V4 Pro support a 1 million token context window at no surcharge.
- Output speed on DeepSeek’s own API averages 117 to 122.7 tokens per second. Third-party providers vary widely: Makora reaches 267 tokens per second while slower providers fall below 75 tokens per second. Time to first token on DeepSeek’s API is approximately 300 to 500 milliseconds.
- V4-Flash-0731 does not beat Claude Opus 4.8 on any of the nine benchmarks, trailing by an average of 5.7 points at roughly one fifty-fourth of Opus 4.8’s blended token cost. On HLE (PhD-level reasoning), V4 Flash scores 8.1% compared to Claude Sonnet 5’s 57.4%, showing the model’s practical ceiling on open-ended reasoning tasks.
- Open weights are available on Hugging Face under a staged regional release, subject to local regulation. Organizations with sufficient GPU capacity can self-host and eliminate per-token fees at high volumes. The model is also available via DeepSeek’s own API, OpenRouter, DeepInfra, Fireworks AI, and Vercel AI Gateway.
- V4 Flash is purpose-built for high-throughput agentic pipelines, code generation at scale, classification workloads, and any production application where per-token cost is a primary constraint. It is not the right model for tasks requiring deep open-ended reasoning, complex multi-step analysis on ambiguous inputs, or applications where HLE-class performance matters.
DeepSeek V4 Flash arrived on July 31, 2026 as the official release of a model that had been in preview since April. The headline story was not a new architecture or a larger parameter count. It was a post-training rerun on an existing model that pushed it past DeepSeek’s own more expensive flagship on nine agent and coding benchmarks simultaneously. That result, at $0.28 per million output tokens, changes the calculation for teams building production pipelines where cost and agentic performance are both constraints.
This review covers the full technical picture: architecture, benchmark results with third-party context, pricing and caching, provider options, and an honest assessment of where V4 Flash excels and where its limits show up.
What Is DeepSeek V4 Flash?
DeepSeek V4 Flash is a Mixture-of-Experts (MoE) language model developed by DeepSeek, a Chinese AI research lab. It is the lighter sibling in the V4 family, with 284 billion total parameters and 13 billion parameters activated per forward pass. The larger sibling, V4 Pro, has 1.6 trillion total parameters with 37 billion active per pass. The MoE architecture is what makes Flash inference fast and inexpensive: the model routes each token through a small subset of experts rather than processing all parameters on every forward pass.
The V4-Flash-0731 build was released to public API on July 31, 2026. It is functionally a post-training upgrade to the V4-Flash-Preview released in April 2026. The architecture, parameter count, and context window did not change. DeepSeek ran a new post-training cycle on the same base, and the resulting model substantially outperformed the preview across agentic and coding benchmarks. The build is available in two modes through the DeepSeek API: a standard non-reasoning mode and a reasoning mode that runs extended chain-of-thought before producing output.
The model is also known as part of the Hailuo product line through MiniMax’s consumer-facing branding, though the technical model family is DeepSeek’s own. V4 Flash is positioned for high-throughput, cost-sensitive production workloads, specifically agentic pipelines, code generation, classification, and applications where 117-plus tokens per second output matters more than frontier-tier reasoning depth.
DeepSeek V4 Flash Features
MoE Architecture and Activation Efficiency
The MoE design activating 13B of 284B parameters per forward pass is the foundational reason V4 Flash is both fast and cheap. Traditional dense models (where all parameters participate in every computation) at 284B total would be far slower and more expensive to run than Flash. The expert routing means the model can handle a 1 million token context window without the memory and compute overhead that a dense model of equivalent parameter count would require.
The same MoE efficiency that makes Flash fast creates its reasoning ceiling. Dense models at equivalent effective compute budget tend to have stronger cross-domain generalization. Flash’s routing trades raw reasoning breadth for speed and cost, which is an explicit design choice rather than a limitation of the overall V4 family.
Post-Training Rerun: What Changed in V4-Flash-0731
The April 2026 preview (V4-Flash-Preview) and the July 2026 official release (V4-Flash-0731) share identical architecture and weights at the pre-training stage. What changed is the post-training pipeline: reinforcement learning from human feedback, instruction tuning, and the agent-specific fine-tuning that governs how the model handles multi-step tool use, code execution, and agentic task completion.
The results of the post-training rerun are documented in DeepSeek’s own benchmark publication. Terminal Bench 2.1 moved from 61.8 (preview) to 82.7 (official), a 33.8% improvement. DeepSWE moved from 7.3 to 54.4, a 645% improvement on that specific benchmark. DSBench-FullStack moved from 37.0 to 68.7. These gains, all on the same base model, demonstrate that agent-specific post-training is currently one of the highest-return investments in model improvement, separate from scaling compute or architectural changes.
Reasoning Mode
V4-Flash-0731 is available in both standard and reasoning modes through the DeepSeek API. Reasoning mode runs extended chain-of-thought before producing the final response, which improves performance on tasks requiring multi-step logic but increases latency and token consumption. Third-party benchmarks place the reasoning mode (max effort) at an AA Intelligence Index score of 50, 10 points above the April Flash preview and 6 points above V4 Pro at non-max effort, according to Artificial Analysis.
The reasoning mode is relevant for applications where Flash’s standard mode falls short on complex tasks. For applications where low latency is critical (real-time chat, streaming completions, high-frequency agentic loops), the non-reasoning mode at 117-plus tokens per second is the practical choice.
1M Token Context Window
Both V4 Flash and V4 Pro support a 1 million token context window at no additional surcharge. For agentic workflows that need to hold large codebases, long conversation histories, or extended documents in context simultaneously, this eliminates the need to implement chunking or retrieval-augmented generation for most real-world input sizes. Competing models charge premiums for extended context or cap context at 128K-200K tokens, making Flash’s 1M context at its price point a practical differentiator for long-context workloads.
Tool Use and Function Calling
V4 Flash supports structured tool use and function calling, which is what makes it suitable for agentic pipelines rather than just text generation. The agent benchmark improvements in V4-Flash-0731 are specifically on tasks that require tool invocation, state management across turns, and multi-step execution sequences. DeepSWE (a software engineering benchmark requiring actual code execution and debugging) is particularly relevant here, and the move from 7.3 to 54.4 on that benchmark reflects genuinely improved agent competency rather than instruction-following improvements alone.
DeepSeek V4 Flash Pricing
| Input type | Price per million tokens |
|---|---|
| Standard input | $0.14 |
| Cache-hit input | $0.0028 |
| Output | $0.28 |
The standard input and output rates are $0.14 and $0.28 per million tokens respectively. The cache-hit input rate of $0.0028 per million tokens applies when the input matches a previously processed prefix, reducing effective input cost by roughly 50x for workloads with repeated system prompts, shared context, or fixed preambles. For agentic loops that repeatedly send a large system prompt with a tool manifest, cache-hit pricing significantly reduces total spend.
DeepSeek V4 Pro, the flagship sibling, is priced at $0.50 input and $1.50 output per million tokens (non-cache). V4 Flash’s output is $0.28 versus V4 Pro’s $1.50, making Flash roughly 5.4x cheaper on output than its own more expensive sibling, at a point where Flash now outperforms Pro on agentic benchmarks.
DeepSeek V4 Flash Pros and Cons
Pros:
- Cheapest capable model for agentic and coding tasks at $0.28 per million output tokens
- Outperforms V4-Pro-Preview on nine agent and coding benchmarks after post-training update
- 117-plus tokens per second on DeepSeek’s API, making it one of the faster models at this capability level
- 1 million token context window at no surcharge
- Open weights available on Hugging Face for self-hosting and eliminating per-token costs
- Cache-hit pricing of $0.0028 per million input tokens for repeated-prompt workloads
- Available across multiple providers (DeepSeek, OpenRouter, DeepInfra, Fireworks AI, Vercel AI Gateway)
Cons:
- Scores 8.1% on HLE versus 57.4% for Claude Sonnet 5, showing a wide gap on open-ended reasoning
- Trails Claude Opus 4.8 on all nine published benchmarks by an average of 5.7 points
- DeepSeek’s origin as a Chinese company creates data residency and compliance concerns for some regulated industries
- Reasoning mode increases latency and token consumption, reducing the cost advantage for tasks requiring extended chain-of-thought
- Provider speed varies significantly: a 356% difference between fastest and slowest third-party providers means choosing the wrong provider matters
- Benchmark scores are self-reported by DeepSeek; third-party Terminal-Bench 2.1 run by Artificial Analysis shows 79% versus the self-reported 82.7%
DeepSeek V4 Flash vs Alternatives
DeepSeek V4 Flash vs V4 Pro: V4 Flash now outperforms V4 Pro on nine agent and coding benchmarks at roughly one-fifth of V4 Pro’s output price. For agentic workloads specifically, there is no longer a quality argument for using V4 Pro over V4 Flash. V4 Pro retains an advantage on non-agentic reasoning tasks and on benchmarks that are not in the nine DeepSeek published, where the larger active parameter count (37B vs 13B) provides more capacity.
DeepSeek V4 Flash vs Claude Sonnet 5: Claude Sonnet 5 scores 57.4% on HLE versus V4 Flash’s 8.1%, and 85.2% on SWE-bench Verified versus no published equivalent for Flash. Sonnet 5’s output costs $10 per million tokens at launch pricing versus Flash’s $0.28, a 35x difference. For tasks where reasoning quality and instruction-following on ambiguous inputs matter, Sonnet 5 is clearly stronger. For high-volume agentic pipelines on well-specified tasks, Flash’s cost advantage is decisive.
DeepSeek V4 Flash vs GPT-5.6 Luna: GPT-5.6 Luna costs $6 per million output tokens versus Flash’s $0.28, a 21x difference. Luna’s first-token latency of under 800ms is competitive with Flash’s 300-500ms. For teams in the OpenAI ecosystem who cannot or do not want to route through DeepSeek’s API, Luna is the comparable option; for teams where cost is the primary constraint and model-provider flexibility exists, Flash is significantly cheaper.
Who Is DeepSeek V4 Flash Best For?
DeepSeek V4 Flash is best suited for teams building high-volume agentic pipelines, code generation systems, and classification workloads where per-token cost is the dominant operational constraint. The model’s agent benchmark performance after V4-Flash-0731 makes it competitive with frontier models on well-specified multi-step tasks at a price point that makes large-scale agentic deployment economically viable.
It is also well-suited for organizations with GPU infrastructure who want to self-host via the open weights and eliminate per-token costs entirely, subject to data residency review given DeepSeek’s Chinese origin. Teams building long-context applications that would otherwise require chunking or RAG benefit from the 1M token context window at no additional cost.
V4 Flash is not well-suited for open-ended reasoning tasks requiring PhD-level analysis, nuanced instruction-following on ambiguous inputs, or applications where Claude Sonnet 5 or Opus-level quality is needed. Regulated industries with strict data sovereignty requirements should conduct compliance review before routing production data through DeepSeek’s hosted API.
Our Verdict
DeepSeek V4 Flash-0731 makes a strong case for being the default model for high-volume agentic and coding pipelines in the second half of 2026. The post-training rerun that pushed it past its own Pro-Preview on nine benchmarks without an architecture change or price increase is a meaningful result, and the $0.28 per million output tokens rate makes it the cheapest capable agentic model available through a public API.
The caveats are real: it trails Claude Opus 4.8 on all nine benchmarks, its HLE score shows significant reasoning limitations compared to frontier models, and compliance review is warranted for regulated use cases. But for the specific workload it is designed for (agentic pipelines, code generation, high-throughput classification), V4 Flash-0731 is currently the best value model available.
Frequently Asked Questions
What does “retrained” mean for DeepSeek V4 Flash?
The V4-Flash-0731 release uses the same model architecture and pre-trained weights as the V4-Flash-Preview from April 2026. What changed is the post-training stage: a new reinforcement learning and instruction-tuning cycle was run on the existing base model, specifically targeting agentic and coding tasks. This produced large benchmark improvements (Terminal Bench 2.1 went from 61.8 to 82.7, DeepSWE went from 7.3 to 54.4) without any change to architecture or parameter count. The price also did not change.
Does DeepSeek V4 Flash really beat V4 Pro?
On the nine specific agent and coding benchmarks DeepSeek published on July 31, 2026, yes. V4-Flash-0731 outscored V4-Pro-Preview on all nine, including Terminal Bench 2.1 (82.7 vs 72.1) and DeepSWE (54.4 vs approximately 7-9 for Pro-Preview). Third-party benchmarks from Artificial Analysis show an AA Intelligence Index of 50 for Flash-0731 reasoning mode versus lower scores for Pro at non-max effort. However, V4 Pro retains advantages on non-agentic reasoning tasks not included in the nine published benchmarks. The claim that Flash beats Pro is accurate for the agentic task set; it is not a universal quality claim.
What are DeepSeek V4 Flash’s limitations?
V4 Flash scores 8.1% on HLE (a PhD-level reasoning benchmark) compared to Claude Sonnet 5’s 57.4%, showing significant limitations on open-ended reasoning. It trails Claude Opus 4.8 by an average of 5.7 points on the nine agent benchmarks where it beats V4 Pro. Benchmark scores are self-reported by DeepSeek; third-party reproductions show a roughly four-point gap between vendor-reported and independently measured Terminal-Bench scores. For teams with data residency requirements, DeepSeek’s Chinese origin requires compliance review before using the hosted API.
How do I access DeepSeek V4 Flash?
V4 Flash is available through DeepSeek’s own API at platform.deepseek.com, through OpenRouter (which provides fallback routing across providers), through DeepInfra, Fireworks AI, and Vercel AI Gateway. Open weights are also available on Hugging Face for self-hosting. Provider selection matters: output speed ranges from 75 to 267 tokens per second across providers, with Makora and DeepSeek’s own API providing the fastest throughput. DeepInfra is generally recommended for most production deployments due to its latency and reliability profile.
What is cache-hit pricing for V4 Flash?
Cache-hit input pricing is $0.0028 per million tokens, approximately 50x cheaper than the standard $0.14 input rate. Cache hits occur when the input matches a previously processed prefix within a session or across sessions within the cache window. For agentic applications that repeatedly send a large system prompt, tool manifest, or shared context prefix, cache-hit pricing significantly reduces effective input cost. The cache window duration for DeepSeek’s API follows standard caching behavior; high-frequency repeated-prefix workloads benefit most.
Can DeepSeek V4 Flash be self-hosted?
Yes. Open weights are available on Hugging Face under a staged regional release subject to local regulation. Self-hosting requires sufficient GPU RAM for the 284B total parameter model, though the MoE architecture means only 13B parameters are active per forward pass, which reduces peak memory requirements compared to a dense 284B model. Organizations self-hosting eliminate per-token API fees at scale, which makes V4 Flash economically attractive for high-volume workloads. Infrastructure and operational overhead are the trade-offs versus using the hosted API.
