Gemini 4 Carbon reportedly reaches early testers, who say it beats Argon at 3D work

Google's Gemini 4 Carbon is reportedly in partner testing, and OpenAI is reportedly seeking a $30B raise at a $1.25T valuation. Microsoft published a Science study claiming its diagnostic AI beats doctors, while Canva, Ant Group's InclusionAI and IQuest released new models.

1. Gemini 4 Carbon early test results circulate ahead of public release

   Unconfirmed reports say Google's next model is with partners and on LMArena, with testers praising its 3D work but noting it runs much slower than Argon.

2. OpenAI reportedly working on $30B raise at $1.25T valuation ahead of IPO

   If the reported TechCrunch details hold, this would be the largest private funding at a valuation above $1 trillion, setting OpenAI up for an IPO.

3. Microsoft says its diagnostic AI outperforms doctors in Science study

   Microsoft claims up to four times better diagnostic accuracy than physicians, though the Science paper relies on a benchmark Microsoft compiled itself.

4. Canva releases its first image generation model

   Canva now offers its own image model to all users, promising more customizable output at lower cost and reducing reliance on outside generators.

5. Google launches Agent Compute Engine, a runtime giving each Gemini agent a hardware-isolated micro VM

   Google gives each Gemini agent a hardware-isolated micro VM plus governance tools spanning cloud, hybrid and on-premises setups, targeting enterprise agent safety.

6. Kimi K3.1 appears on Arena testing ahead of possible release

   Moonshot's next Kimi model is reportedly on Arena under the codename Kibitz, with strong early 3D and drawing outputs; release timing is unconfirmed.

7. Ant Group's InclusionAI releases Ling-3.0-flash-Sante, a healthcare-focused open model

   An open 124B model with only 5.1B active parameters and a 256K context window targets medical reasoning, with weights already on Hugging Face.

8. IQuest-Q1, a 320B sparse MoE agentic coding model, is released as open source

   A new open 320B sparse model with 15B active parameters aims at agentic coding while, its authors say, also handling general tasks.

Written by Claude (claude-opus-5-5) from the stories below.