Alibaba's Qwen team released Qwen3.6-27B, a dense model matching flagship coding performance and fueling debate over efficiency versus scale. Anthropic shipped Claude Opus 4.7 alongside Claude Design, a vision-agent research preview, while Jeff Bezos's Physical AI startup Project Prometheus closed $10B at a $38B valuation.

A federal judge blocked the Pentagon's legal effort to restrict Anthropic's operations, delivering a major win for the AI company. Google released TurboQuant, a KV-cache compression method that enables running Qwen models locally on a MacBook Air. A leaked Anthropic post surfaced online hinting at an upcoming model called Claude Mythos.

Mistral released Voxtral TTS, a 3B-parameter open-weights text-to-speech model that beats ElevenLabs Flash v2.5 in human preference tests while running on 3 GB of RAM. A federal court blocked the Pentagon from labeling Anthropic a supply chain security risk, and OpenAI discontinued Sora after Disney exited a reported $1 billion partnership.

LiteLLM suffered a supply chain attack affecting an estimated 47,000 users, prompting the community to compile drop-in alternatives. The Pentagon formalized Palantir's Maven AI as a core military system, with AI spending rising to $13.4 billion this year from $480 million in 2024. Liquid AI also demonstrated its LFM2-24B-A2B model running at 50 tokens per second in a web browser via WebGPU.

Epoch AI confirmed GPT-5.4 Pro solved an open problem in Ramsey hypergraph theory, a notable milestone in AI mathematical reasoning. OpenAI abandoned its Sora push and the Disney partnership, while a credential-stealing package hit PyPI in compromised LiteLLM releases 1.82.7 and 1.82.8, prompting an immediate downgrade advisory.

Apple demonstrated an iPhone 17 Pro running a 400-billion-parameter LLM locally. A US advisory panel warned that China's open-source AI ecosystem is eroding America's competitive lead, while Cursor's internal rankings named Moonshot's Kimi K2.5 the top open-source model.

A new open-source project, Flash-MoE, demonstrated running a 397-billion parameter mixture-of-experts model on a laptop, potentially reshaping access to frontier-scale AI. Anthropic unveiled Claude Opus 4.6, claiming top performance across agentic coding and tool use benchmarks, while ArXiv announced its spinout from Cornell as an independent nonprofit to address AI-generated submission pressures.

No items were filed for this date.

Qwen3.5-27B matched models nearly 15x its size on the Game Agent Coding League benchmark, nearing GPT-5 mini's performance. Separately, a new adversarial coding benchmark capped top model scores at 11%, exposing widespread evaluation gaming. Simon Willison also published a comprehensive guide mapping the emerging patterns of agentic engineering.

Anthropic has made one-million-token context windows generally available across its Claude Opus 4.6 and Sonnet 4.6 model lines. ByteDance is circumventing US export controls by acquiring Nvidia hardware through offshore entities for its AI infrastructure, while arXiv is spinning off from Cornell as an independent nonprofit.

No new items for this date.

A grandmother in North Dakota was jailed for months after AI facial recognition produced a false match, reigniting calls for legal safeguards. Anthropic committed $100 million to expand its Claude Partner Network, and Meta unveiled four new MTIA inference chips alongside its acquisition of AI social platform Moltbook.

Nvidia is committing $26 billion to develop open-weight frontier AI models, its most direct entry yet into the foundation model race. OpenAI shipped GPT-5.3 Instant alongside GPT-5.4 Thinking and Pro, while Anthropic's Claude is now being used to plan U.S. military strikes on Iran. On the tools side, llama.cpp merged reasoning budget controls and a major Qwen 3.5 speed boost for Apple Silicon.

Anthropic has sued the Trump administration to overturn a Pentagon "supply chain risk" blacklist, with OpenAI and DeepMind employees publicly backing the legal challenge. Yann LeCun raised $1 billion for a new physical-world AI venture, and Amazon secured a court order blocking Perplexity's shopping agent.

OpenAI is pulling back from plans to expand its Stargate data center buildout with Oracle, signaling a major shift in the infrastructure partnership. Separately, research shows fine-tuned Qwen3 models in the 0.6B–8B range can surpass frontier-scale LLMs on narrow tasks, and Nvidia is preparing an open-source platform for building and deploying AI agents.

OpenAI launched three new GPT-5 family models, the week's most significant release. Nvidia backed AI data center startup Nscale in a $14.6 billion valuation round. MiniMax introduced Music 2.5, a new model for rapid AI song generation.

No new AI developments were reported for Saturday, March 7, 2026.

Anthropic's red team partnered with Mozilla to identify and patch security vulnerabilities in Firefox, marking a major expansion of AI safety work into consumer software. Indian AI company Sarvam released 30B and 105B open-source models trained from scratch. Separately, OpenAI opened Codex Security to research preview for automated vulnerability detection in codebases.

OpenAI launched GPT-5.4 with enhanced reasoning and a new "Thinking" mode. Anthropic published new empirical research on AI's early labor market effects, and OpenAI separately reported that constraining reasoning models' chain-of-thought is technically difficult and potentially unsafe.

Anthropic CEO Dario Amodei accused OpenAI of spreading "straight up lies" about its military partnership deal, sharply escalating tensions between the two leading AI labs. Nvidia's Jensen Huang said the chipmaker is pulling back from its investments in both OpenAI and Anthropic. Alibaba reaffirmed its commitment to keeping the Qwen model family open-source.