开源大模型 - AI News

DeepSeek V4 Open-Source Model Released: 1.6 Trillion Parameters, Million-Token Context – Can It Overthrow Closed-Source Dominance?

On April 25, 2026, Chinese AI company DeepSeek officially open-sourced its V4 series large models, with the Pro version boasting 1.6 trillion parameters and supporting a 1 million token context window, alongside a low-compute Flash variant and a 75% API discount until May 5, 2026. Winzheng.com's evaluation based on YZ Index v6 methodology reveals that it is the first open-source model to match closed-source leaders in key dimensions like code execution and grounding, while offering superior cost-effectiveness.

The Open Source Debut of 1.6T Parameter DeepSeek-V4 Matches the Performance of Top Closed Models at a Fraction of GPT-5.5's Cost

DeepSeek has launched the preview version of its V4 series of open-source large models, marking a significant milestone in the open-source AI community's challenge to closed-source leaders. The DeepSeek-V4 achieves performance comparable to top-tier closed models at a much lower inference cost.

Claiming to be the Third Globally, Supporting 8-Hour Long Inference: Can the Unofficially Announced GLM-5.1 Really Set a New Benchmark for Open Source Models?

This article analyzes signals from winzheng.com Research Lab, based on information from Z AI's internal test notifications. The status of these signals is unconfirmed, and we will continue to monitor official announcements and independent test results.

DeepSeek-V2 Open Source Release: 671B Parameters with Only 37B Activated, Performance Rivals GPT-4o

Chinese AI startup DeepSeek releases its latest open-source large language model DeepSeek-V2, featuring 671 billion total parameters but requiring only 37 billion activated for efficient inference, with performance metrics approaching OpenAI's GPT-4o.

DeepSeek-V2 Open Source Release: 236B Parameters Run on Just 16GB VRAM, Math Capabilities Surpass Llama3, Igniting Developer Community

DeepSeek team releases open-source LLM DeepSeek-V2 with 236B parameters requiring only 16GB VRAM for inference, outperforming Meta's Llama3 in mathematical benchmarks. The model has garnered over 150,000 reposts in Chinese communities, marking a major breakthrough for domestic AI in efficient large model development.

开源大模型 (5 articles)

DeepSeek V4 Open-Source Model Released: 1.6 Trillion Parameters, Million-Token Context – Can It Overthrow Closed-Source Dominance?

The Open Source Debut of 1.6T Parameter DeepSeek-V4 Matches the Performance of Top Closed Models at a Fraction of GPT-5.5's Cost

Claiming to be the Third Globally, Supporting 8-Hour Long Inference: Can the Unofficially Announced GLM-5.1 Really Set a New Benchmark for Open Source Models?

DeepSeek-V2 Open Source Release: 671B Parameters with Only 37B Activated, Performance Rivals GPT-4o

DeepSeek-V2 Open Source Release: 236B Parameters Run on Just 16GB VRAM, Math Capabilities Surpass Llama3, Igniting Developer Community