Baidu opens ERNIE 5.5 Flash Base: a 280B MoE to fine-tune yourself

Pre-trained on 33.5T tokens, 14B active, released before instruction tuning. Plus: Cursor ships prompt-to-deployed web apps; Copilot hits Slack and Teams.

Nowline NOV 7 6:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • A 280B open base, yours to post-train

    ERNIE 5.5 Flash Base is a Mixture-of-Experts model — 280B total params, 14B active per token — pre-trained on 33.5T tokens and shipped as a raw end-of-pretraining checkpoint with no instruction tuning, so the post-training is yours to do. Weights are on Hugging Face in BF16 and FP8 under Baidu's own license, serve on vLLM and SGLang, stretch to a 1M-token context for fine-tuning, and self-report MMLU 93.3 and HumanEval+ 82.3 against open peers like DeepSeek V4 Flash and MiMo V2.5.

  • Cursor ships prompt-to-deployed web apps

    Cursor's cloud agent now turns a plain-language brief into a running web app: it scaffolds the project, provisions a Postgres database, wires auth behind Cursor SSO, runs security checks, then publishes to a live, shareable link. It's on every paid plan with cloud agents enabled — so a weekend build can go from prompt to hosted, authenticated app without you touching infra.

  • GitHub Copilot lands in Slack and Teams

    @GitHub is now generally available in Slack and Teams, out of preview for everyone. Reference issues and PRs, get live previews, and kick off and triage work from the chat window your team already sits in — one less tab-over to github.com mid-conversation.