Preview Versions of the DeepSeek-V4 Series Released
On July 3, DeepSeek released preview versions of its next-generation "DeepSeek-V4" model series on Hugging Face. The lineup consists of the following two models:
DeepSeek-V4-Pro**: 1.6T total parameters, 49B active
DeepSeek-V4-Flash**: 284B total parameters, 13B active
The standout feature of both models is support for a **1 million token context length**. On the architecture side, they employ a **hybrid attention mechanism** combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), cutting single-token inference FLOPs to 27% and KV cache usage to 10% compared to the previous model (V3.2) in the 1M token context configuration.
The models also received across-the-board updates, including enhanced residual connections via Manifold-Constrained Hyper-Connections (mHC), adoption of the Muon optimizer, and pre-training on more than 32T tokens.
In benchmarks, DeepSeek-V4-Pro-Max scored **93.5% (Pass@1) on LiveCodeBench** and reached a Codeforces Rating of 3206, coming close to closed-source state-of-the-art models on some metrics. The license is MIT.
A variant called DeepSeek-V4-Flash-DSpark, which adds a speculative decoding module, has also been released; it can be enabled in vLLM by adding a single flag.
Theorem-Proving AI "Leanstral 1.5" Finds 5 Unreported Bugs in Real-World Code
"Leanstral 1.5", an AI model that saturated the formal theorem-proving benchmark "miniF2F", has reportedly **discovered 5 unreported bugs in real-world open source code**.
One of the discovered bugs was an overflow issue lurking in the zigzag decoding logic of `datrs/varinteger`, a Rust data processing library. It was a serious bug — crashing in debug builds and silently corrupting values in release builds — yet nobody had ever filed a GitHub Issue for it.
It is a symbolic case worth noting: AI is beginning to step from "competition-style theorem proving" into "real-world software quality assurance".
Claude Fable 5 Is Back — But at Twice the Price of Opus
Anthropic has resumed offering its top-tier model "Claude Fable 5". Pro and Max plans include a bundled access period through July 7, but **API pricing is set at roughly twice that of Opus**.
The return of a flagship model is welcome news, but it calls for careful cost management. To avoid suddenly finding your API bill has doubled, it is recommended to restrict usage to targeted use cases.
Anthropic Reaches $30 Billion in Annualized Revenue
Anthropic has reportedly reached **$30 billion in annualized revenue**. That translates to a staggering 80x growth year over year, driven primarily by the rapid adoption of the AI coding tool "Claude Code".
Reaching SaaS-company revenue levels in under three years shows that enterprise adoption of AI agents has become a reality. The growth power of the AI coding space has once again been confirmed by the numbers.
---