The Day in AI

• Moonshot AI released Kimi K3.1 Pro on 23 October, and Artificial Analysis scores it 55, fourth overall and 11 points above Kimi K3.
• Block announced Dessa on 23 October, a speech-to-speech model trained only on human conversation that runs locally on a MacBook Pro M4 Max.
• Hugging Face released Carbon-X, a 14.8B open model that reads raw DNA in a chat, averaging 66.7% on five DNA benchmarks by its own count.
• Nanbeige released Nanbeige4.3-33B under Apache 2.0, and its model card reports 65.5 on DeepSWE against 42.5 for Qwen3.8-27B.
• The UK government reportedly launched a £100 million AI biosecurity lab with Google, OpenAI and Anthropic.

Kimi K3.1 Pro is a mixture of experts (MoE) model, so each token is routed through a small subset of the network. It has 2.8 trillion total parameters, 104 billion of them active per token, and a 1 million token context. It's on Moonshot's API now at $2 per million input tokens and $10 per million output tokens. Moonshot has promised full weights by 4 November 2026.

Artificial Analysis puts it at 55, up from 44 for Kimi K3 (max) in July and three points behind Claude Opus 5.5 at 58. Moonshot reports 65.7 on Terminal-Bench 4.0, against 66.4 for Claude Opus 5.5 and 58.2 for GPT-6 Astra. Moonshot also says it removed cyber-related data from its training phases to reduce misuse. A Highspeed version runs two to three times faster at twice the price, except for cache hits.

Block, the company led by Jack Dorsey, announced Dessa on 23 October. Dessa works on audio end to end, so there's no transcription step in the middle, and everything it knows came from recorded human speech. Block says it trained the model on a single Nvidia GB200 NVL72 rack and that it runs locally on a MacBook Pro with an M4 Max or better. Only a small group of AI safety researchers gets early access, and Block has given no date for wider release.

Hugging Face released Carbon-X under Apache 2.0, a 14.8B dense model built on Qwen3-14B. It puts DNA and text in one token stream, adding 5,515 DNA tokens to the tokenizer instead of using a separate DNA encoder. Hugging Face reports a 66.7% average on five DNA benchmarks, against 67.5% for GPT-6.1 Sol. Its best result is 74.9% on BioReason Coding and its weakest is 29.8% on GenomeQA splice-site questions. Coverage describes four training stages ending in reinforcement learning with GRPO, and that detail is only partly confirmed.

• Nanbeige released Nanbeige4.3-33B with a 512K token context, which Startup Fortune reported on 25 October. Its model card shows it slightly behind Qwen3.8-27B on SWE-Bench Pro, 60.3 against 61.7.
• Anthropic launched OSS Scanner on 8 October, a free opt-in service that scans open-source projects with models including Claude Mythos and sends reports with no human review.
• The UK government reportedly launched a £100 million AI biosecurity lab to detect biological threats, backed by Google, OpenAI and Anthropic.
• DeepSeek and MiniMax cleared China's AI export controls for overseas use, according to The Next Web.
• The US government banned DeepSeek, according to L'Orient Today. Scope and timing aren't clear from the coverage.
• MOSI.AI published the MOSS-Image report on 16 October, a 5B diffusion transformer for text-heavy images that the authors say reaches 96.3% text coverage on LongLongTextBench.
• Lynch lab reports that a ten-year-old convolutional network beats frontier language models at skin cancer diagnosis.