Chinese lab DeepSeek releases two versions of its newest large language model, V4 Pro and V4 Flash, boasting 1.6 trillion parameters and improved p…
DeepSeek has launched two versions of its newest large language model, V4 Pro and V4 Flash, which the company claims have improved performance and efficiency compared to previous models. The V4 models use a mixture-of-experts approach, activating only certain parameters per task to reduce inference costs. This allow…
The DeepSeek V4 models boast an impressive 1.6 trillion parameters, with the smaller V4 Flash model having 284 billion parameters and the larger V4 Pro having a total of 49 billion active parameters. The company claims that this large parameter count enables the models to outperform previous versions on reasoning be…
While the DeepSeek V4 models show promise, they still lag behind frontier models such as OpenAI’s GPT-5.4 and Google’s Gemini 3.1 Pro in knowledge tests. However, DeepSeek suggests that their development trajectory is catching up quickly, with a gap of around 3 to 6 months between the two.
The affordability of the V4 models is another major advantage for operators, with prices starting at $0.14 per million input tokens for the smaller model and $0.145 per million input tokens for the larger model. This puts them below the costs of leading models such as GPT-5.4 Nano and Gemini 3.1 Flash.
What matters
- Mixture-of-experts approach reduces inference costs and enables large codebases to be used in prompts
- V4 models outperform DeepSeek V3.2 on reasoning benchmarks, but fall behind frontier models on knowledge tests
- More affordable than leading models, with prices starting at $0.14 per million input tokens
Why it matters
More affordable than leading models, with prices starting at $0.14 per million input tokens
This GenAI News article was prepared in original wording using reporting and materials published by TechCrunch AI. Source reference: https://techcrunch.com/2026/04/24/deepseek-previews-new-ai-model-that-closes-the-gap-with-frontier-models/.
Drafted by the GenAI News review pipeline.