DeepSeek has released the V4-Flash model, which reportedly outperforms the Pro version. The article also highlights the largest GitHub data crawl to date and efforts in engineering system prompts for safer code generation. These developments underscore the rapid pace of AI innovation and the data-driven nature of the open-source community.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
With a fine-tuning run that left the architecture unchanged, DeepSeek's smaller V4-Flash model outperformed its own flagship Pro in independent tests, landing at the frontier of intelligence and cost. The event challenges the intuition that bigger is always better, and highlights the untapped potential of fine-tuning and inference optimization.
DeepSeek released V4-Flash-0731 in August 2026, the official version of the V4 series small model, replacing the April preview. According to The Batch, the model outperformed its own flagship V4-Pro in independent tests, achieved solely through a new round of fine-tuning with no changes to architecture or parameter count. This breaks the conventional wisdom that larger models are necessarily stronger, highlighting the critical role of fine-tuning in unlocking a model's potential.
The specific mechanisms behind the performance gain have not been fully disclosed. DeepSeek only stated that the improvement comes mainly from fine-tuning, but did not explain how the new fine-tuning differs from the preview version, nor did it disclose training data or knowledge cutoff dates. This lack of transparency leaves uncertainty for external evaluation of its reproducibility.
V4-Flash uses a Mixture-of-Experts (MoE) architecture with 284 billion total parameters, but only 13 billion are activated per token. An optional speculative decoding module can increase the checkpoint to 304 billion parameters. Its attention layers alternate between two compression schemes, compressing 4 tokens and 128 tokens into a single entry respectively, significantly reducing computation and memory requirements for long inputs.
At full input length, V4-Flash requires only 10% of the computation of V3.2 and 7% of the memory for storing key-value pairs. This architectural advantage translates directly into cost savings: API pricing is $0.14 per million input tokens, $0.0028 for cached input, and $0.28 for output, far below comparable proprietary models.
On Artificial Analysis's Intelligence Index, V4-Flash scores 50, just one point below GPT-5.6 Luna at its maximum reasoning setting. It also sits on the evaluator's Pareto frontier of intelligence versus cost per task, meaning no other tracked model offers equal or higher intelligence at a lower cost.
The model supports reasoning modes (low, high, maximum) and non-reasoning mode, with tool calling and context caching. Its weights are freely available under the MIT license, allowing both non-commercial and commercial use, further lowering adoption barriers and potentially attracting developers and enterprises.
DeepSeek pre-trained V4-Flash on over 32 trillion tokens, followed by two-stage fine-tuning. In the first stage, expert models are built for each domain (e.g., math, coding, agent tasks) using supervised fine-tuning and reinforcement learning based on Group Relative Policy Optimization (GRPO). The second stage uses online policy distillation to merge experts: the merged model generates responses, which are then corrected according to the selected expert's style.
Reasoning levels are trained as separate behaviors, applying different length penalties and context windows. The maximum reasoning level preloads a system prompt that requires the model to fully decompose problems and test boundary conditions. In tool-calling tasks, the model retains the entire reasoning history, including user messages, to ensure contextual coherence.
Despite the impressive performance, key details remain undisclosed: the differences between the new fine-tuning and the preview version, training data composition, and knowledge cutoff dates are all unknown. This limits external assessment of the model's capability boundaries.
Furthermore, the performance data comes from independent tests by Artificial Analysis, but the specific methodology, test sets, and error margins are not detailed in the source. Therefore, the one-point gap between the 50 score and GPT-5.6 Luna may be influenced by test conditions and should be interpreted with caution.
This report is based on a retelling by The Batch (DeepLearning.AI), which sourced information from DeepSeek's official release and Artificial Analysis's evaluation. DeepSeek's official statements are source claims, and while independent test results are cited, raw data is not provided, so overall credibility is moderate. Key performance figures (e.g., the 50 score, costs) should be treated as claimed by the source rather than independently verified conclusions.
The upset by DeepSeek-V4-Flash demonstrates that, with an unchanged architecture, careful fine-tuning and inference optimization can significantly boost model performance and achieve the frontier of intelligence per dollar at very low cost. This case may prompt the industry to reconsider the 'bigger is better' paradigm, but undisclosed training details and evaluation methods still require further verification.
Primary report
Primary source