Ling 3.0 Tiny hits OpenRouter: a reasoning MoE for $0.015/1M

A 1.3B-active reasoning model you can rent for cents or self-host in INT4 on a DGX Spark. Plus 4-bit DeepSeek V4.1 Flash for Blackwell and Hy4 at half price.

Nowline NOV 6 10:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • A reasoning agent you can rent for cents

    inclusionAI's Ling 3.0 Tiny — a 7.9B MoE with just 1.3B active per token, 262K context, and switchable thinking/instant modes — is now live on OpenRouter at $0.015 in / $0.045 out per 1M tokens, MIT-licensed. Weights ship in BF16, FP8 and INT4 and run on a DGX Spark or Apple Silicon, so you can prototype a cheap agent in the cloud and self-host the same model later.

  • DeepSeek V4.1 Flash, now 4-bit on Blackwell

    NVIDIA published an MIT-licensed NVFP4 (4-bit) build of DeepSeek V4.1 Flash, with native SGLang support as of Nov 4. The ~475 GiB checkpoint shrinks to ~400 GiB on four Blackwell GB300s and holds accuracy within about a point — GPQA Diamond ~93, Terminal-Bench 2.1 ~81 — so the frontier-class coder is finally practical to self-host on Blackwell silicon.

  • Update: Tencent Hy4 lands on OpenRouter, 50% off

    The official Hy4 — Tencent's agentic open model (Vibe Code Bench 81%, 1M-token context, open weights on Hugging Face) — is now callable on OpenRouter via Tencent Cloud at $0.417 in / $1.251 out per 1M, half the $0.834 / $2.501 list, with $0.021/1M cache reads. If you'd rather not host the weights, that's frontier-length context and strong coding as a cheap API call.