The Day in AI

• DiffuSpace open-sourced Clover-35B-A3B-Preview, a block diffusion model for coding and agents, and reports 73.6 on SWE-bench Verified.
• Mistral AI released the weights of Mistral Large 4, its 1-trillion-parameter multimodal model aimed at agentic coding and cybersecurity.
• Tencent said at its Global Digital Ecosystem Conference that Hy4 and Hy4 Flash will be released soon and will support visual understanding.
• Rime launched Prism, a fine-tunable speech-to-speech model that writes a checkable text version of its answer before it speaks.
• DeepSeek and Moonshot AI used OpenAI's GPT-6 Astra to train their latest models, according to The Information.

Clover uses block diffusion. It writes text in blocks, one after another, and inside each block it refines a noisy draft of all positions at once instead of producing one token at a time. To make this work, the team restructured a Gated DeltaNet backbone and added a learned calibration of block size. It's a mixture-of-experts model with 35 billion parameters, of which about 3 billion are active per token, and it has a thinking mode.

DiffuSpace reports 51.7 on Terminal-Bench 2.1 alongside the SWE-bench score. Speed is the point of the design. The team measured 515 tokens per second on agent tasks on one H200, against 365 tokens per second for a DFlash2 speculative-decoding setup. No independent evaluation of these results has been published yet.

Mistral released the weights of Mistral Large 4, nicknamed Le Chonk, on 29 October. The model entered public preview on 6 October with 1 trillion total parameters and 49 billion active (Mistral's documentation lists 52 billion), and Mistral had promised weights by the end of October. During the preview, Mistral red-teamed a version with reduced moderation with cybersecurity partners. Mistral reports 19.7% on Cognition's FrontierCode 2.0, a private benchmark that scores code quality and whether a maintainer would merge the change, and says this leads open-weight models.

Mistral also says the model matches the best on the Artificial Analysis Coding Agent Index while using fewer tokens, and it offers three reasoning modes to trade cost against performance. Two further claims appear in Mistral's post only in truncated passages. These are that the model received three more weeks of reinforcement learning after the preview, and that it leads open-weight models on Terminal-Bench.

Tencent's Hy4 preview came out in August with 770 billion total parameters, 49 billion active, and a 1 million token context window, all company-reported. Tencent says the full Hy4 will do better on long reasoning and execution tasks, including real productivity workflows. Hy4 Flash is the cheaper version and keeps visual understanding and agent features. Tencent gave no release date.

• Rime reports 56.6 on the τ-Voice benchmark for Prism and says it can be deployed in a private cloud or fully on-premises.
• DeepSeek and Moonshot AI reportedly made GPT-6 Astra the top target for distillation among Chinese labs, meaning training on its outputs, according to The Information.
• Anthropic is "Customer B" in a SpaceX filing, its largest customer at 34.5% of third-quarter revenue, according to Bloomberg.
• Fluidstack is reportedly raising at a $40 billion valuation, up from $18 billion in September, according to Bloomberg.
• Sebastian Thrun co-founded Dulo, which builds foundation models for hardware design and reportedly raised a $130 million seed, according to The Information.
• OpenAI reorganized its Model Spec around five values, from faithfulness to human intent through to care.
• OpenAI now gives eligible ChatGPT Pro and Business subscribers up to $500 a month in free API credits.
• Google now lets all Gemini users preview and edit PDFs, Docs, Sheets, slides and code files inside a chat.
• Manohar Paluri, a founding member of FAIR, reportedly left Meta on 23 October to become CTO of Kevin Weil's drug discovery startup Warp Sciences.
• Deepgram acquired Daily, the maker of the Pipecat voice agent framework.