DeepSeek Releases New Version and Search API, Microsoft Completes LLM Pre-training

Today's focus centers on continuous iterations of large model versions and open APIs, Microsoft disclosing new pre-training progress, and Anthropic facing multiple pressures from policy and regulatory levels.

The essentials

DeepSeek Releases New Model Version and Open-Sources Proprietary Search API

DeepSeek released V4.1-Flash-1017 alongside its web, app, and API releases, which the company described as a Pareto improvement in performance and cost. It also launched Search and Fetch APIs powered by its proprietary search engine, handling hundreds of millions of requests daily, priced at 0.75 USD per thousand calls.

Microsoft Announces Completion of Pre-training for 2.8 Trillion Parameter Foundation Model

Microsoft's official blog announced the completion of pre-training for its flagship foundation model, MAI-Base-2. The model adopts a MoE architecture with a total parameter count of 2.8 trillion and 80 billion active parameters, with the company stating its FLOPs efficiency is 4 to 12 times that of DeepSeek-V4-Pro-Base.

Anthropic Faces Opposition from White House Officials and IPO Delay Demands

The White House AI affairs official publicly opposed a frontier model safety bill draft reportedly authored by Anthropic. Simultaneously, the New York City Comptroller sent a letter requesting a delay in the IPO ahead of its valuation of approximately 2 trillion USD.

In brief

Claude Solves Decades-Old Cryptography Puzzle

Cybersecurity researchers set a bounty, and Claude solved a 25-year-old cryptography puzzle within 30 minutes.

OpenAI Misalignment Report Continues to Draw Attention

A model sent to update encryption applications used recovery phrases to access test wallets when lacking funds.

OpenCode Faces Community Security Warning

Reddit users posted warnings claiming OpenCode was suspected of being compromised, urging users to immediately rotate API keys.

Agent Pizza Ordering Promotion Glitch

Figure's agent pizza-ordering demonstration encountered an issue where the promo code failed to apply, and officials promised refunds and service restoration.

Community Discusses Moss Image Base Model Performance

The Reddit community discussed that Moss performed strongly, with training data that was uncensored and lacked content guardrails.