Mistral Eyes €3B Raise at €20B Valuation

Google researchers propose 'faithful uncertainty,' a technique that allows large language models to express qualified guesses rather than either confidently hallucinating or refusing to answer. The approach reframes hallucinations as 'confident errors' and enables models to hedge responses appropriately, preserving utility while maintaining trustworthiness. This addresses a core tradeoff in LLM deployment where eliminating factual errors typically forces models to abstain from answering questions they actually know.

Moonshot AI released Kimi K2.7-Code, an open-source coding model claiming 30% lower thinking-token usage and double-digit performance gains over K2.6. Independent practitioners testing the model on public benchmarks report it produces more honest code implementations but with weaker actual performance, and have challenged Moonshot to submit results to independent benchmarks like DeepSWE rather than relying on proprietary test suites. The efficiency gains are immediately deployable via OpenAI-compatible API, but real-world capability claims remain unverified.

Researchers from NYU, Columbia, Princeton, University of Maryland, Harvard, and Lawrence Livermore National Laboratory published a paper introducing Latent Context Language Models (LCLMs), a compression technique that reduces LLM input by 16x while maintaining accuracy better than existing methods. Unlike KV cache compression, LCLMs compress tokens before decoder processing, delivering 8.8x faster output on long-context benchmarks. The models are open-sourced on HuggingFace and designed to integrate into existing LLM stacks.

Anthropic has released Claude Fable 5, making its Mythos-class model available to the public for the first time. The model includes built-in guardrails that restrict responses in high-risk domains including cybersecurity and biology. This release marks a significant step in bringing advanced AI capabilities to broader audiences while attempting to manage safety concerns.