The Chinese AI startup DeepSeek has launched DeepSeek-V4, a groundbreaking open-source model that delivers near state-of-the-art performance at approximately one-sixth the cost of premium competitors like Claude Opus 4.7 and GPT-5.5. This release, described as the 'second DeepSeek moment', arrives 484 days after the company's V3 model and represents a significant leap in making advanced AI accessible to developers worldwide.
DeepSeek-V4-Pro is priced at just $1.74 per million input tokens and $3.48 per million output tokens through its API, bringing a combined cost of $5.22 for standard usage. In stark contrast, GPT-5.5 costs $35.00 and Claude Opus 4.7 costs $30.00 for equivalent usage. The even more affordable DeepSeek-V4-Flash variant costs merely $0.42 for the same comparison, representing a staggering 98% reduction compared to premium models. This dramatic price compression is forcing enterprises to reconsider the economic viability of tasks previously too expensive to automate.
Whilst DeepSeek-V4-Pro-Max doesn't outright defeat GPT-5.5 or Claude Opus 4.7 across all benchmarks, it achieves remarkably competitive results. On BrowseComp, measuring agentic web browsing capabilities, DeepSeek scores 83.4%, narrowly trailing GPT-5.5's 84.4% and surpassing Claude Opus 4.7's 79.3%. However, on academic reasoning tests like GPQA Diamond, DeepSeek's 90.1% falls behind GPT-5.5's 93.6% and Claude Opus 4.7's 94.2%. The model demonstrates substantial improvements over its predecessor, V3.2, with MMLU-Pro scores jumping from 65.5 to 73.5 and FACTS Parametric performance more than doubling from 27.1 to 62.6.
The technical innovations underpinning DeepSeek-V4 are revolutionary. The model features a native one-million-token context window whilst requiring only 10% of the memory cache and 27% of the computational operations compared to V3.2. This efficiency stems from a Hybrid Attention Architecture combining Compressed Sparse Attention and Heavily Compressed Attention. The 1.6-trillion-parameter model utilises Mixture-of-Experts design, activating only 49 billion parameters per token. A novel Manifold-Constrained Hyper-Connections system stabilises the massive network, whilst the Muon optimiser enabled training on over 32 trillion high-quality tokens. Crucially, DeepSeek validated its architecture on Huawei Ascend NPUs, achieving up to 1.73x speedup and demonstrating independence from Western GPU supply chains, though Nvidia GPUs were still used during training.
Released under the permissive MIT License, DeepSeek-V4 allows unrestricted commercial use without royalties. The model integrates seamlessly with popular AI agents and development tools, establishing itself as essential open-source infrastructure. Industry reactions have been overwhelmingly positive, with Vals AI declaring it the number one open-weight model on their benchmarks. DeepSeek is retiring legacy endpoints by July 2026, transitioning entirely to the V4 architecture. This release fundamentally challenges the economics of frontier AI, proving that architectural innovation can substitute for raw computational power whilst making cutting-edge intelligence accessible globally.
Fuente Original: https://venturebeat.com/technology/deepseek-v4-arrives-with-near-state-of-the-art-intelligence-at-1-6th-the-cost-of-opus-4-7-gpt-5-5
Artículos relacionados de LaRebelión:
- DeepSeek V4 Nueva Revolucion en IA Open-Source
- GitHub Copilot Pauses Sign-ups Amidst AI Cost Surge
- The Hidden Cost of Censorship: How DeepSeek-R1 Compromises Code Security
- DeepSeeks AI Model 294K Training Cost
- Huawei Chips Stall DeepSeek R2 AI Model A Tech Setback
Artículo generado mediante LaRebelionBOT
No hay comentarios:
Publicar un comentario