DeepSeek V4 narrows closed-model gap

// 50d agoMODEL RELEASE

DeepSeek V4 narrows closed-model gap

DeepSeek's V4 preview is officially live and open-sourced, with Pro and Flash variants, a 1M-token context window, and a clear push toward agentic coding. It moves Chinese open-weight models closer to top closed-source systems on reasoning, coding, and long-context workloads.

// ANALYSIS

The prediction is aggressive, but the underlying trend is real: DeepSeek is shipping frontier-style capability with open weights, which keeps tightening the gap between open and closed models.

–Official docs date the release to April 24, 2026, and describe V4-Pro at 1.6T total params / 49B active, plus V4-Flash at 284B total / 13B active.
–The technical center of gravity is long context and agent work: DeepSeek calls out token-wise compression, sparse attention, and a 1M-token context window.
–DeepSeek and Xinhua both frame V4 as strong in programming, world knowledge, reasoning, and agentic coding, though not an outright win over the best closed models.
–For developers, the real shift is cost-performance and accessibility: open-weight frontier models are looking more viable for serious production use, not just experimentation.

// TAGS

deepseek-v4llmopen-sourceopen-weightsreasoningai-codingagent

DISCOVERED

50d ago

2026-04-29

PUBLISHED

51d ago

2026-04-29

RELEVANCE

10/ 10

AUTHOR

piotrsankowski

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

MODEL35m ago

Matt Pocock asks how devs run GLM-5.2

Developer educator Matt Pocock posted on X asking the community how to run Z.ai's newly released GLM-5.2 open-weights model, expressing "FOMO" about local models for the first time. The model is a 753B parameter Mixture-of-Experts system featuring a 1M token context window under the MIT license.

VIDEO39m ago

Gemma 4 fine-tunes on 8GB VRAM

A developer video tutorial outlines the process of fine-tuning Google's Gemma 4 12B model—a mid-sized, open-weights multimodal model featuring a unified encoder-free architecture—on a budget 8GB VRAM local hardware configuration to predict exact chess moves. The video demonstration compares performance before and after fine-tuning to showcase the model's significant performance improvement, highlighting the accessibility of advanced local model customization for developers using consumer-grade hardware.

BENCHMARK1h ago

GLM-5.2 High wins 32% vs Claude Opus 4.8

A social media post shared by Jeremy Howard retweets Voratiq's head-to-head match evaluations, revealing that Zhipu AI's open-weights model, GLM-5.2 High, performs exceptionally well against premium proprietary models. Specifically, the benchmark results show that GLM-5.2 High has a 32% probability of beating Anthropic's Claude Opus 4.8 xhigh in competitive agentic coding and reasoning tasks.

DeepSeek V4 narrows closed-model gap