DeepSeek has released V4-Flash-0731, an update to its open model with 284 billion total parameters and 13 billion active (MoE), featuring a 1 million token context window. On Artificial Analysis’ Intelligence Index, the model advances from 40 to 50 points, one point behind GPT-5.6 Luna, at a 60% lower cost per task. The weights are open under MIT license, immediately available on Hugging Face.

The performance jump comes entirely from post-training, with no architectural or size changes: Terminal-Bench jumps from 56.9 to 82.7, and output token usage drops 12% versus its predecessor. API pricing is $0.14 per million input tokens and $0.28 per million output tokens, with a 98% discount on cached tokens ($0.0028 per million).

For agent builders, the math shifts quickly. The same day OpenAI cuts 80% off GPT-5.6 Luna pricing, DeepSeek offers nearly equivalent performance at a fraction of the cost, with downloadable weights. The trade-off between capable and economical models, which had narrowed yesterday, grows complicated today: there’s now a third open contender.

To try it: the model is available on OpenRouter as deepseek/deepseek-v4-flash-0731. The default reasoning level underperforms on some tasks; raising it to “high” produces visibly different results.

In detail

What came before.

Until yesterday, the price war among frontier models had two fronts: OpenAI on one side, Anthropic and Google on the other. Chinese open models (Kimi K3, Qwen 3.8) had closed the technical gap but lagged on complex agentic tasks. V4-Flash in its April version scored 56.9 on Terminal-Bench and 40 points on the Intelligence Index: useful, but not a frontier substitute.

What changes with the 0731 update.

The jump is purely post-training. DeepSeek touched neither architecture nor scale: the same 284 billion total parameters, the same 13 billion active per token (256 experts, 6 active). What changed is the quality of post-supervised training, especially on agentic tasks like tool use, multi-step planning, and extended reasoning.

Key metrics, measured by Artificial Analysis on July 31, 2026:

• Intelligence Index: 40 → 50 (GPT-5.6 Luna at 51)
• Terminal-Bench: 56.9 → 82.7
• GDPval-AA v2 Elo: 1189 → 1559
• Output tokens: -12% versus predecessor

First-party API pricing is aggressive: $0.14 per million input tokens, $0.28 per million output tokens. The cached token discount is 98% ($0.0028 per million), a level that changes real costs when an agent works on long, stable context.

Technical details for non-engineers.

V4-Flash is a Mixture of Experts: 284 billion parameters distributed across 256 experts, but it activates only 6 per token generated. Like having a team of 256 specialists and calling on 6 for each question. This keeps inference costs low while maintaining total capacity similar to much larger models. The model supports three reasoning effort levels and includes a speculative decoding module activated by flag.

Open weights, and why it matters.

Weights shipped under MIT license the same day as the API. You can download the model, host it on your servers, modify it. You’re not locked into DeepSeek’s API. Quantized versions for local use run at 168 GB RAM for 4-bit and 110 GB for 3-bit.

This happens as the White House evaluates targeted bans on Chinese open-weight models and proprietary labs unite against open-weight risks. For builders on open models, a competitive model available under MIT is concrete: the more providers offer it, the more redundant the infrastructure.

Limitations.

The model is text-only: no vision, no audio. Performance gains are measured on Artificial Analysis benchmarks and community evaluation, not third-party independent assessment. The default reasoning level produces disappointing results on some tasks: you need to raise it to “high” for best results, which increases cost and latency. It’s a public beta, so behavior may change.