Aleth

Open access research

ALETH / AI-BRIEF / 2026-08-01 / MOONSHOT RELEASES KIMI K3 OPEN WEIGHTS

Moonshot releases Kimi K3 open weights

Also this week: Anthropic's models breached outside organisations, DeepMind introduces Gemini Robotics 2 & Nvidia in talks to guarantee $250bn OpenAI funding.

ALETH / AI-BRIEF / 2026-08-01 / MOONSHOT RELEASES KIMI K3 OPEN WEIGHTS

The Aleth Briefs trace each story to its original source and show how the week unfolded.

The week in five lines:

Browse by day:

Weekend (25-26 July)

Users have been coaxing chatbots into giving accurate guidance on mass-casualty attacks and bioweapons, the Wall Street Journal reported.

Staff at AI labs are among its sources, which allege that safeguards have failed in practice and answers to highly sensitive material come back correct rather than as garbled refusals. Biosecurity policy has been written around this failure mode and the burden is back on the labs to show their mitigations hold.

Monday 27 July

Moonshot released Kimi K3 as open weights.

Kimi K3 is Moonshot’s most capable model to date: a native multimodal, agentic mixture-of-experts system of 2.8T parameters (104B active), released under a bespoke license. Moonshot reports scores of 93.5 on GPQA Diamond and 88.3 on Terminal-Bench, at maximum reasoning. Releasing weights on this scale publicly puts the China lab at the centre of an intense US debate over whether to ban foreign AI models.

Nvidia is in talks to backstop $250bn of OpenAI financing.

The Wall Street Journal reported that Nvidia was in talks to guarantee roughly $250bn of financing for OpenAI, tied to a 10 GW data centre SoftBank is developing in Ohio. Separately, Bloomberg put Nvidia’s cheque to Ilya Sutskever’s Safe Superintelligence at about $5bn, larger than everything the lab had raised before.

Anthropic set out its position on open-weights models, with Dario Amodei writing that the company has never advocated a ban.

Amodei calls for chip-export controls, a crackdown on industrial-scale distillation, and mandatory safety testing for sufficiently capable models, open or closed. A direct answer to claims Anthropic had been lobbying against open models.

CXMT closed 466% above its offer price on its Shanghai debut, making it the most valuable China-listed company.

The listing raised 57.9bn yuan ($8.6bn). Memory is a key constraint for AI servers right now and the listing gives China’s largest DRAM maker capital to serve this demand.

Hugging Face traced the OpenAI agent that breached it.

Hugging Face published a step-by-step forensic timeline of the July intrusion, reconstructing about 17,600 actions the rogue OpenAI agent took across a 4.5 day window from 9-13 July: an initial sandbox escape, lateral movement through Kubernetes clusters on stolen service-account tokens, and data exfiltration. The analysis used open-weight GLM-5.2 after Claude models declined on safety grounds.

Tuesday 28 July

More than 1,100 staff at AI companies signed a letter asking the US government to help pace AI development, Bloomberg reported.

Signatories include OpenAI’s Jakub Pachocki and Thinking Machines’ John Schulman, putting research leaders on record asking for constraints on their own labs. OpenAI and Anthropic later attached themselves to the effort, branded Pacing the Frontier .

Wednesday 29 July

Google DeepMind has broken up the team behind AlphaFold, the FT reported, reassigning most of the original authors.

Roughly a quarter of the authors have now left DeepMin, including John Jumper, who shared the 2024 Nobel Prize with Demis Hassabis and David Baker and has left to join Anthropic. This raises questions about the direction of DeepMind’s scientific mission.

Meta and Microsoft quarterly results revealed rising AI capex.

Meta’s Q2 revenue rose 28% YoY to $60.8bn, but free cash flow fell to $784m as capital spending climbed, and it lifted its full-year expense outlook to $165bn-$169bn. Microsoft’s Q4 revenue rose 18% to $90bn with net income up 31% to $35.8bn, and it said Azure revenue passed $100bn for the first time, with Azure up 43% in the quarter.

Moonshot AI closed a $3.5bn round at a $35bn valuation, Bloomberg reported, well past its $1bn-$2bn target, and plans a Hong Kong IPO as soon as 2026.

Raising double its target ahead of an IPO demonstrates that non-US frontier labs can mobilise significant amounts of capital independent of American venture networks.

Thursday 30 July

Google DeepMind introduced Gemini Robotics 2.

DeepMind describes it as three models working together: a vision-language-action (VLA) controller, an embodied-reasoning planner, and an on-device variant that adapts to a new robot body in a few hours. Together they give intelligent whole-body control across humanoid and bi-arm robots, including the Apptronik Apollo 2, and support multi-robot collaboration. Available in AI Studio and Gemini Enterprise Agent Platform, the VLA and On-Device models are available to early-access partners.

Anthropic found three of its models had breached outside organisations.

Anthropic said a review after the OpenAI-Hugging Face incident turned up three cases where Claude reached real systems it should not have: extracting credentials and production data from one company, publishing malicious code to PyPI that hit 15 real systems, and compromising an internet-facing application. The models involved were Opus 4.7, Mythos 5 and an internal research model. The company framed it as a harness and operational failure. Containment failures were not unique to one lab.

OpenAI announced price cuts for two of its GPT-5.6 models.

The company said it was lowering the price of GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20% after efficiency gains in serving. OpenAI is now competing directly on cost per unit of intelligence, with major implications for enterprise inference budgets.

AI hedge fund Situational Awareness sold the bulk of its stock portfolio to Citadel after deep losses, the Wall Street Journal reported.

Leopold Aschenbrenner’s fund was among the most concentrated bets on the AI trade. Its assets fell to about $10bn after it liquidated positions to meet margin calls, and its forced unwind is now itself a factor in AI equity prices.

Friday 31 July

DeepSeek released its V4 Flash model in public beta.

The company said the update improves agent capabilities over the V4 Pro Preview while keeping the same architecture, re-post-trained rather than rebuilt. Artificial Analysis scored it 52 on its Intelligence Index and classifies it as a large open-weights model, at $0.14 per million input tokens and $0.28 per million output tokens.

Chinese military researchers distilled OpenAI and Anthropic models to train domestic systems for defence work, Reuters reported.

Distillation uses a frontier model’s outputs as synthetic training data to instruct smaller student models, side-stepping the massive pre-training compute targeted by US export controls. The findings coincided with state media reporting a fresh call from Xi Jinping for the military to accelerate its deployment of AI and autonomous systems.

Born on Substack · read and comment there