DeepSeek drops 158B V4 Flash, V4 Pro open-source models

// 45d agoMODEL RELEASE

DeepSeek drops 158B V4 Flash, V4 Pro open-source models

DeepSeek has launched its new V4 model family, introducing a massive 158B-parameter Flash model geared toward local execution alongside a V4 Pro version. Both open-weight models are available for immediate download on Hugging Face.

// ANALYSIS

DeepSeek continues to push the boundary for massive open-weight models with this V4 release, offering a colossal 158B parameter model squarely aimed at local deployment.

–The 158B Flash model specifically targets developers running massive local hardware setups, cementing DeepSeek's appeal to open-source enthusiasts.
–Releasing a high-end V4 Pro tier indicates a strong commitment to pushing state-of-the-art boundaries against closed AI labs.
–Hugging Face availability ensures immediate access for the developer community to benchmark and fine-tune these new heavyweights.

// TAGS

deepseek-v4deepseekllmopen-weightsopen-sourceself-hosted

DISCOVERED

45d ago

2026-04-24

PUBLISHED

45d ago

2026-04-24

RELEVANCE

9/ 10

AUTHOR

Bijan Bowen

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

BENCHMARK11m ago

Structured docs boost coding agent precision

Mintlify published results showing that connecting Claude Code to structured documentation via MCP on a million-line monorepo improved answer precision by 64%. The experiment also demonstrated a 50% reduction in token consumption and a 1.5x speedup in task completion.

OPEN SOURCE11m ago

Xiaomi MiMo-V2.5-Pro-UltraSpeed tops 1,000 TPS

Xiaomi's MiMo team and TileRT have released MiMo-V2.5-Pro-UltraSpeed, achieving decoding speeds over 1,000 tokens per second on a trillion-parameter MoE model using a single 8-GPU node. This ultra-fast serving mode is enabled by DFlash speculative decoding and MXFP4 quantization, and the model checkpoints are open-sourced on Hugging Face.

RESEARCH15m ago

Chris Potts introduces AI coding CPI

Stanford researcher Christopher Potts and Moritz Sudhof created a "Consumer Price Index" for AI coding using developer interaction data from the SWE-chat dataset. Examining Anthropic's Opus 4.6 model, they found that engineering output per token dropped by up to 85% in two months, though code survival rates rose from 90% to 95%.