DeepSeek-V4-Flash Outshines Pro, The Biggest GitHub Crawl Yet, Engineering System Prompts for Safer Code
The Batch Analytics DeepLearning.AI ● Covered by 2 sources
DeepSeek’s new Flash model beats its bigger Pro sibling after another fine-tune. It’s also cheap enough to make always-on agents a lot more practical.
Based on reporting by The Batch, Analytics DeepLearning.AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
DeepSeek has done the annoying thing rival labs love to do: make the smaller model look smarter. The company’s new DeepSeek-V4-Flash-0731, built on the same architecture as the April preview, now beats the larger DeepSeek-V4-Pro on independent tests after another round of fine-tuning.
The model is still a mixed bag of hard numbers and careful engineering. It uses a mixture-of-experts transformer with 284 billion total parameters and 13 billion active per token, plus an optional speculative decoding module that brings the checkpoint to 304 billion parameters. DeepSeek says it can handle text inputs of up to 1 million tokens and outputs of up to 384,000 tokens, with a top output rate of 122.7 tokens per second.
On Artificial Analysis’ Intelligence Index, the max-reasoning setting scored 50 points. That put it ahead of the April preview’s 40 and above DeepSeek-V4-Pro’s 44, tied with Gemini 3.6 Flash at 50, and just behind GPT-5.6 Luna and GLM-5.2, both at 51. It still trailed Kimi K3 at 57. But the cost story is the louder one: DeepSeek says an Intelligence Index task at max reasoning costs $0.03 through its API, versus $0.05 for GPT-5.6 Luna under the comparison cited here.
The gains show up where agents actually spend money. On GDPval-AA v2, it reached 1,558 Elo, second among open-weight models. On Terminal-Bench 2.1 it solved 82.7 percent of problems, up from 61.8 percent in the preview. On τ³-Bench Banking it hit 31.1 percent, also ahead of preview. And on CodeArena WebDev, its high-reasoning setting debuted at 1,577 points, good for seventh overall.
DeepSeek says the boost came mainly from fine-tuning, not a new architecture. It also leaned on specialist models for areas like math, coding, and agentic work, then merged them with on-policy distillation. The max reasoning mode now uses a system prompt that pushes the model to break problems down and test edge cases, while tool-using runs keep the reasoning history in context instead of tossing it away. That’s the sort of plumbing that matters when a model is meant to sit in the loop all day, not just ace a demo.
My take — AI-written commentary, not fact-checked reporting
This is the real AI race now: not who can make the biggest model, but who can make a competent one cheap enough to leave running. The tokenmaxxing crowd can keep chanting; finance will eventually ask for a bill. Open weights are winning the part that actually matters to builders: control, price, and the option to walk away from one provider without a melodramatic migration saga.
Read more about this at: The Batch
