DeepSeek previews new AI model that ‘closes the gap’ with frontier models

Source: TechCrunch

Overview

Chinese AI lab DeepSeek has launched two preview versions of its newest large language model, DeepSeek V4, a much‑awaited update to last year’s V3.2 model, along with the accompanying R1 reasoning model that took the AI world by storm.

Model Details

• Architecture: Both DeepSeek V4 Flash and V4 Pro are mixture‑of‑experts (MoE) models with context windows of 1 million tokens, enabling large codebases or documents to be used in prompts. The MoE approach activates only a subset of parameters per task, lowering inference costs.
• Parameter Counts:
  • V4 Pro: 1.6 trillion total parameters (49 billion active) – the largest open‑weight model currently available, surpassing Moonshot AI’s Kimi K 2.6 (1.1 trillion), MiniMax’s M1 (456 billion), and more than double DeepSeek V3.2 (671 billion).
  • V4 Flash: 284 billion total parameters (13 billion active).

Both models are text‑only, unlike many closed‑source peers that also handle audio, video, and images.

Performance

• Efficiency: Architectural improvements make the V4 models more efficient and performant than DeepSeek V3.2.
• Reasoning Benchmarks: DeepSeek claims the V4‑Pro‑Max model outperforms open‑source peers across reasoning benchmarks and surpasses OpenAI’s GPT‑5.2 and Gemini 3.0 Pro on certain tasks.
• Coding Benchmarks: In coding competition benchmarks, both V4 models are described as “comparable to GPT‑5.4.”
• Knowledge Tests: The models lag slightly behind frontier models such as OpenAI’s GPT‑5.4 and Google’s Gemini 3.1 Pro on knowledge‑focused evaluations, indicating a developmental trajectory that trails state‑of‑the‑art frontier models by roughly 3–6 months.

Pricing

• V4 Flash:

  • $0.14 per million input tokens
  • $0.28 per million output tokens

• V4 Pro:

  • $0.145 per million input tokens
  • $3.48 per million output tokens

Both pricing tiers undercut comparable offerings from GPT‑5.4 Nano, Gemini 3.1 Flash, GPT‑5.4 Mini, Claude Haiku 4.5, Gemini 3.1 Pro, GPT‑5.5, and Claude Opus 4.7.

Controversy

The launch follows a U.S. accusation that China is stealing American AI labs’ IP on an industrial scale using thousands of proxy accounts. DeepSeek itself has been accused by Anthropic and OpenAI of “distilling”—essentially copying—their AI models.

• Accusation details: BBC report
• Anthropic’s claim: TechCrunch article

References

• DeepSeek V4 collection on Hugging Face: https://huggingface.co/collections/deepseek-ai/deepseek-v4
• R1 reasoning model coverage: https://techcrunch.com/2025/01/27/deepseek-claims-its-reasoning-model-beats-openais-o1-on-certain-benchmarks/
• Accusation of IP theft: https://www.bbc.com/news/articles/cpqxgxx9nrqo
• Anthropic’s distillation allegation: https://techcrunch.com/2026/02/23/anthropic-accuses-chinese-ai-labs-of-mining-claude-as-us-debates-ai-chip-exports/