Tinfield 1 is an open-weight coding and long-horizon software-engineering model released by Bad Theory Labs (Nigeria) that targets terminal work and multi-step tool use. It reports 33.0 on Terminal-Bench 4.0 and 62.0 on DeepSWE v1.1, exceeding the published scores for Claude Opus 4.8. The model has 177 billion total parameters with 6.6 billion active parameters per token and a massive 256K context window. It ships in three builds - Base, Compact and Mini - so users can trade off footprint and performance.
Benchmarking and distribution details are explicit: Tinfield’s published improvements over its base run are Terminal-Bench 29.0 → 33.0 and DeepSWE 58.7 → 62.0, evaluated with a mini-swe-agent at k=5 across 66 Terminal-Bench and 113 DeepSWE tasks. Compact and Mini builds are sized at 72.0 GiB/45.2 GiB GPU-resident weights and 61.5 GiB/34.7 GiB GPU-resident weights respectively, with an engram table memory-mapped on CPU. Weights and instructions are available under the Qwen Community License 1.0 on Hugging Face. The developer is also soliciting pre-seed investment.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
A recent report indicates that nearly half of young people prefer trusting AI over humans for fact-checking. This shift highlights growing confidence in artificial intelligence for verifying information. (news.sky.com)
TypeSafe AI introduced Jev, a new category of decision models that output floating point numbers instead of text, enabling fast and cost-effective classification and scoring tasks. Jev processes unstructured data and answers yes/no, choice, or score questions, making it suitable for applications like spam detection and search ranking. (simonwillison.net)
MiMo-v2.6-Pro is a new AI model focused on balancing intelligence, performance, and cost. It aims to provide advanced capabilities at a competitive price point. (artificialanalysis.ai)
Multiple models of Claude, including Mythos, Fable 5.1, and Opus 5, are experiencing elevated error rates. The incident affects various Claude services and is currently under investigation. (status.claude.com)
Xiaomi MiMo v2.6 is an open-source language model that demonstrates diverse capabilities, including use in scientific environments and multimedia tasks. The release is seen as a significant step forward in lightweight models at an affordable cost. (mimo.xiaomi.com)
Transformers are neural network architectures that use self-attention to process sequences and capture long-range dependencies, revolutionizing AI since 2017. They power models like GPT-2, LLaMA, and Gemini across various domains, with key components including embeddings, attention mechanisms, and MLP layers. (poloclub.github.io)
Today's best Hacker News stories, summarized and screenshotted, one email a day.