Step 5 Preview is a long-context, agent-focused language model from StepFun built as a sparse Mixture-of-Experts system (27B active parameters, 600B total) and positioned for extended, tool-driven workflows across large codebases and documents. It targets software engineering and professional knowledge work with particular strength in finance, supports function-calling and structured JSON outputs, and accepts text, images, and video as inputs while returning text. The model is designed to refine results over multiple steps and to integrate with external tools, making it suitable for agentic automation and multi-stage reasoning tasks.
Key operational details: a 1,000,000-token context window with up to 64,000 completion tokens; pricing at $1.00 per 1M input tokens, $2.70 per 1M output tokens, and $0.05 per 1M cache-read tokens; and OpenAI-compatible API access via OpenRouter. Measured performance shows 107 tokens/sec throughput and a P50 latency of 3.49s, with recent three-day uptime around 99.94% and availability ~99.43%. The model is hosted by a single provider on OpenRouter, which offers monitoring, routing, and app-usage signals. Related StepFun offerings include Step 3.7 Flash and Step 3.5 Flash, which emphasize multimodal understanding and efficient MoE activation with smaller active parameter counts and 262K context windows.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Artificial intelligence is used by a developer to build free alternatives to Adobe software following dissatisfaction with Adobe's cancellation policies. The project offers accessible design and editing tools for users seeking free options. (techspot.com)
Superwhisper S1 mini is a text and tone normalizer designed for speech-to-text applications, available on Hugging Face. It aims to improve transcription accuracy by standardizing speech output. (huggingface.co)
Frontier AI models have surpassed human experts in predicting earnings surprises, marking a significant breakthrough in financial forecasting. The latest models outperform analyst consensus in accuracy, especially in predicting company revenue reports. (samaya.ai)
Anthropic has updated its usage policy to prohibit 'abusive or cruel behavior' towards its AI model Claude. The new rules also address propaganda, surveillance, and weapon development concerns. (theverge.com)
Whistle is an open-source speech recognition model that runs entirely on devices, with a 16.9 MB size and no dependencies. It transcribes speech in seven languages, provides word timestamps, and generates speech embeddings, all on the device itself. (cactuscompute.com)
Jevman is a benchmark where AI decision models play Pac-Man against arcade ghosts, with six models competing in 100 games each. The platform allows users to benchmark, rank, and submit their own models via an open-source API interface. (opper.ai)
Today's best Hacker News stories, summarized and screenshotted, one email a day.