Ant's Ling-3.2-flash: a 560B multimodal model, free for 2 weeks

The 25B-active MoE takes text, image and video in a 1M-token window via OpenAI- and Anthropic-compatible APIs. Plus Claude Code 2.1.319 and OpenLake.

Nowline NOV 7 10:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • A 560B native multimodal model, free for two weeks

    Ant's Ling-3.2-flash is a mixture-of-experts model (560B total, 25B active) that natively ingests text, images and video across a 1M-token context, served through OpenAI- and Anthropic-compatible APIs — point an existing client at it by swapping the base URL and model ID (Ling-3.2-flash), with image_url/video_url content blocks for media. It's free for the first two weeks, then original pricing resumes; vendor-reported evals edge past Claude Opus 5 on vision suites (OmniDocBench 92.21 vs 90.56, Vision2Web 85.86 vs 80.20). Weights aren't on Hugging Face, so it's API-only — but it's an easy weekend swap for video-understanding agents or image-to-frontend tools.

  • Claude Code can now deny tool calls instead of prompting

    v2.1.319's new permissions.denyInsteadOfPrompting setting rejects any tool call that would otherwise block on a prompt and lets the turn continue — the missing piece for unattended runs, subagents and teammates. MCP tool calls that run as tasks now move to the background immediately, /clear no longer kills background agents or workflows, and several Bash permission bypasses (a crafted env prefix, a sed trick) are patched. If you run Claude Code behind an LLM gateway, auto mode now prompts or denies calls the server can't verify instead of silently falling back to the local classifier.

  • OpenLake serves open models that keep post-training themselves

    The YC launch lets you deploy an open model (DeepSeek V4, GLM 5.3), wire in your datasets, data sources and RL environments, and have it continuously post-trained in place as new data arrives — then A/B each version against the last and export the checkpoints, which are yours to keep. The company claims models cut serving costs roughly in half on average (no baseline disclosed). Live now at openlakehq.com, with a Fall 2026 YC batch behind it.