DeepSeek V4 Flash 0731: Open-Weight MoE Beats Bigger Model

DeepSeek V4 Flash 0731: How Better Training Beat a Bigger Model

On 31 July 2026, DeepSeek released DeepSeek-V4-Flash-0731, publishing the weights on Hugging Face under an MIT licence and moving the official V4-Flash API into public beta. The model is a 284-billion-parameter sparse Mixture-of-Experts architecture that activates 13 billion parameters per token. Its context window extends to 1,048,576 tokens — one million tokens — with a maximum output of 65,536 tokens. The pricing enters the public beta at 0.14 US dollars per million cache-miss input tokens, 0.28 US dollars per million output tokens, and 0.0028 US dollars per million cache-hit input tokens — a 98 percent discount for cached input that has direct implications for agentic workflows that re-read the same system context across many steps.

What DeepSeek V4 Flash 0731 Is Not

V4 Flash 0731 is not a new model architecture. It is a re-post-trained version of the April 2026 preview release, built on the same 284B MoE base but refined through an entirely new post-training pipeline focused on four domains: coding, agentic task execution, multi-step reasoning, and tool use. DeepSeek did not train a larger model to achieve the V4-Pro benchmark position; it trained the existing Flash-tier model better. That decision matters because it demonstrates a consistent pattern in frontier model development in 2026: targeted post-training on narrowly defined capability domains can shift benchmark performance as significantly as increasing parameter count, and at a fraction of the compute and cost required to scale parameters.

Benchmark Results: Outperforming V4-Pro

DeepSeek reports that V4 Flash 0731 surpasses the larger DeepSeek-V4-Pro (Preview) on every agentic benchmark the company has published. On Artificial Analysis's Intelligence Index, the model scores 50, placing it well above the median of 25 across comparable models. The output speed at DeepSeek's own API reaches 113 tokens per second, with SiliconFlow at 107.9 and Fireworks at 77.9. For reference, V4-Pro Preview is a substantially larger model in terms of active parameters. The result — a smaller active-parameter model outperforming a larger one after post-training — extends a trend that DeepSeek established with V3 in late 2024 and continued with R1 in early 2025: architecture and training efficiency matter as much as scale.

The Cache Pricing Mechanic

The 0.0028 US dollar per million cache-hit input tokens price deserves specific attention. At 0.14 US dollars per million for a cache-miss call and 0.0028 US dollars for a cache-hit call, a developer whose agent loop re-reads a 100,000-token system prompt on every step pays 98 percent less per re-read after the first. For agentic applications that process long tool outputs, conversation history, or document corpora across hundreds of sequential steps, the difference between 0.14 and 0.0028 US dollars per million tokens shifts the unit economics of the entire application. A session that would cost 28 US dollars in input tokens without caching costs approximately 0.56 US dollars if the first read is the only full-price call. This pricing model is now the baseline for how Indian AI product teams should structure long-context agentic applications: front-load the expensive reads, cache aggressively, and run subsequent steps at the cache-hit rate.

The MIT Licence: Open Weights, Commercial Use

The weights released on Hugging Face are under an MIT licence, which permits commercial use, modification, and redistribution without royalty or usage fee obligations. This is the same licence structure DeepSeek used for R1 in January 2025, and it has the same practical effect: any team with the inference infrastructure can run V4 Flash 0731 as a self-hosted model, eliminating per-token API costs entirely for teams at sufficient scale. For Indian software agencies and product companies that have built internal GPU infrastructure — or that use cloud GPU providers such as RunPod, Lambda Labs, or DigitalOcean GPU Droplets — self-hosting the model is now an option at MIT licence terms.

What This Means for Indian AI Product Teams

For Indian AI product companies, V4 Flash 0731 raises the capability ceiling of what is achievable at sub-dollar-per-million-token pricing. At 0.14 US dollars per million input tokens for cache misses and 0.28 US dollars per million for output, V4 Flash 0731 sits in a highly accessible pricing tier while delivering performance that DeepSeek's benchmarks indicate exceeds the larger V4-Pro Preview on agentic tasks. For teams building coding assistants, document analysis tools, long-context RAG pipelines, or autonomous workflow agents, that means a higher-performing model is now available at a price point that makes production deployment economically rational at most scales. For teams with self-hosted GPU infrastructure, the MIT licence makes it possible to eliminate inference costs entirely for on-premise deployments.

The Bottom Line

DeepSeek released V4 Flash 0731 on 31 July 2026, publishing MIT-licensed weights on Hugging Face and moving the API into public beta. The model is a 284B-parameter sparse MoE architecture activating 13B parameters per token, with a 1,048,576-token context window, output speed of 113 tokens per second at DeepSeek's API, and pricing at 0.14 US dollars per million cache-miss input tokens, 0.0028 US dollars per million cache-hit input tokens, and 0.28 US dollars per million output tokens. It is not a new architecture — it is a re-post-trained version of the April preview, improved through a new pipeline focused on coding, agents, reasoning, and tool use. Despite being a Flash-tier model, DeepSeek reports it outperforms the larger V4-Pro Preview on every published agentic benchmark. The 98 percent cache-hit discount makes it particularly cost-effective for agentic applications that re-read large contexts across sequential steps, and the MIT licence makes self-hosting a viable option for teams with GPU infrastructure.