Skip to content
Source · Daily Brief

AI Daily Brief — 29 January 2025

Lunar New Year’s Day delivered a multi-front escalation. Alibaba released Qwen2.5-Max — an MoE pretrained on more than 20 trillion tokens, claimed to beat DeepSeek V3 and to surpass Claude 3.5 Sonnet, GPT-4o, and Llama 3.1-405B on key benchmarks. Bloomberg reported Microsoft was probing whether a DeepSeek-linked group exfiltrated OpenAI API data in autumn 2024. Microsoft posted Q2 FY25 earnings with an AI run-rate of $13 billion (up 175% YoY). Meta’s Q4 print reaffirmed $60-65B 2025 capex; Zuckerberg said Llama 4’s smaller model had finished training and called for an “American” global open-source standard. Dario Amodei published “On DeepSeek and Export Controls.” Microsoft simultaneously added DeepSeek R1 to Azure AI Foundry and GitHub Models — a fast embrace play just as the distillation probe deepened.

Top stories

  • Alibaba releases Qwen2.5-Max on Lunar New Year Day. Large-scale MoE pretrained on 20T+ tokens, exposed as qwen-max-2025-01-25 on Alibaba Cloud. Claims to outperform DeepSeek V3 on Arena-Hard, LiveBench, LiveCodeBench, and GPQA-Diamond — and to surpass Claude 3.5 Sonnet, GPT-4o, and Llama 3.1-405B on key benchmarks. Timing on the first day of Lunar New Year was widely read as a direct response. via SiliconANGLE
  • Microsoft probing if DeepSeek-linked group improperly obtained OpenAI data. Bloomberg: security researchers observed individuals they believe linked to DeepSeek exfiltrating large amounts of data via the OpenAI API in autumn 2024. White House AI czar David Sacks publicly amplified the framing the prior day. via Bloomberg
  • Microsoft Q2 FY25: AI run-rate hits $13B, +175% YoY. Revenue $69.6B (+12%), net income $24.1B, EPS $3.23. Microsoft Cloud $40.9B (+21%). On the call Nadella praised DeepSeek’s software optimisations and emphasised compute-efficiency gains. via CNBC
  • Meta Q4: Zuckerberg says DeepSeek validates open-source, capex stays $60-65B. Q4 revenue $48.4B (+21%), net income $20.8B (+49%), EPS $8.02. Zuckerberg said a smaller Llama 4 has finished training and a larger version is “making great progress.” Called for an “American” global open-source standard. via CNBC
  • Dario Amodei publishes “On DeepSeek and Export Controls.” Anthropic’s CEO argues DeepSeek’s progress does NOT invalidate US chip export controls and in fact makes them more existentially important — framing RL-for-reasoning as the new scaling axis and warning that if China cannot access millions of chips, the world stays in a unipolar US-allied AI configuration. via darioamodei.com
  • Microsoft makes DeepSeek R1 available on Azure AI Foundry and GitHub Models. R1 added to the Foundry catalog (joining 1,800+ models) via a serverless endpoint, plus GitHub Models. Microsoft framed it as enterprise-grade access with SLA/security/responsible-AI guardrails. Future distilled R1 variants promised for local execution on Copilot+ PCs. Announced the same day as the distillation probe — a notably fast embrace play. via Microsoft Azure

Who shipped

Alibaba/Qwen‘s Qwen2.5-Max plus Microsoft‘s Azure-side R1 hosting dominated the lab side. OpenAI, Google DeepMind, and xAI stayed quiet on the product front. Anthropic‘s shipment was a policy essay, not a model.

Open-source pulse

DeepSeek apps were pulled from Italy’s Apple App Store and Google Play after Tuesday’s Garante information request; DeepSeek’s official response said it “had not entered the Italian market.” The full Garante processing ban would land Thursday. DeepSeek’s Janus-Pro (multimodal unified-encoder, 1B/7B) from earlier in the week continued spreading on Hugging Face — Simon Willison’s writeup remained the reference explainer.

Money, infra & hardware

Tesla Q4 missed: adjusted EPS 73¢ versus 76¢ expected, revenue $25.71B versus $27.26B expected; automotive revenue fell 8% YoY. The call leaned heavily into a 2025 robotaxi launch and 2026 Cybercab manufacturing. NVIDIA, TSMC, and Broadcom kept recovering from Monday’s wipeout as analysts argued cheaper inference would expand long-run GPU demand (Jevons paradox framing).

Quiet corners

The US Navy directive from earlier in the week was clarified as a reiteration of existing generative-AI policy rather than a DeepSeek-specific ban, but the episode kicked off a wider wave of US federal and state agency restrictions that would accelerate in the following days.

By the numbers

  • 20T+ tokens — Qwen2.5-Max pretraining
  • $13B — Microsoft AI annualized run-rate (+175% YoY)
  • $60-65B — Meta reaffirmed 2025 capex
  • $48.4B — Meta Q4 revenue (+21%)
  • Most-mentioned lab: DeepSeek (as subject); Alibaba (as actor)

Compiled by AI Feed’s editor from verified web sources for 29 January 2025.