Transformers v5.23.0 adds Inkling 2 Small, a 594B model with 17B active
Library users also get the experimental JEV decision head, Liquid AI’s LFM3, and Dots 3 Note Preview. Vercel serves Inkling 2 Small on AI Gateway for coding agents and image checks.
Transformers v5.23.0 adds Inkling 2 Small from Thinking Machines Lab, so library users can run that multimodal model, along with JEV, LFM3, and Dots 3 Note Preview.
Inkling 2 Small
Inkling 2 Small accepts text and images and generates text. Thinking Machines Lab built it for English plus other languages, and for coding in multiple languages. Its language backbone contains 48 decoder layers that use a mixture of experts. Every layer sends each token to 8 experts drawn from 384. The release puts the model at 594B parameters in total, with 17B active.
The network keeps the Inkling decoder: relative attention that is learned, causal convolutions that stay short, and mixture-of-experts layers. Added on top are residuals around block attention, gated attention that includes a learned sink, and dynamic convolutions over keys and values. A vision transformer based on the Perception Encoder turns each image into language-model embeddings, one embedding for every 40x40 pixel patch.
The other new models
JEV scores labels that occupy a single token. A linear layer on the last token of the prompt produces the score, and that score is soft-capped in the same fashion as logits from a language-model head. JevForDecision joins the checkpoint to this head and loads head.safetensors and judge_config.json stored beside the weights. GEV uses the same head. The head can read 255 answer codes tied to pplx-decider and to d3. Callers may use the experimental decision pipeline or invoke decide directly. Transformers warns that JEV and the pipeline are experimental: programming interfaces, configuration, and behavior may change as development continues, and production environments should treat them with caution.
LFM3 is the third generation of the Liquid Foundation Models from Liquid AI. As with LFM2, the architecture pairs gated short convolutions and grouped-query attention. Attention is split between full-attention layers and sliding-window layers. RMSNorm is applied to query heads and key heads, and a sigmoid on each head gates the output. Each layer places a sandwich norm on the operator output and the feed-forward output before the residual is added. The family added here is LFM3-Text.
Dots 3 Note Preview is a causal language model built as a mixture of experts. It takes native text, image, video, and audio. Images and video share one vision encoder, while audio uses an encoder in the style of Whisper. Each encoder projects the representation it produces.
On AI Gateway
Vercel offers Inkling 2 Small on AI Gateway and describes it as a multimodal reasoning model for coding agents, visual understanding, and professional knowledge work. It can stay on long coding jobs, call tools to collect information, and shape that material into documents and spreadsheets, among other deliverables. Visually, it can read text inside images, interpret charts and infographics, and inspect interfaces. In visual development, screenshots let it judge code it has written and steer the next revision.
AI Gateway exposes one API for model calls, usage and cost tracking, retries, failover, and performance optimizations aimed at uptime above the provider’s. Built-in pieces include custom reporting, Zero Data Retention support, budgets on API keys, and routing rules. Prices follow the provider, with no markup and no platform fee on inference, including Bring Your Own Key requests. The model name is thinkingmachines/inkling-2-small. For Claude Code, Codex, or Cursor, install the newest Vercel CLI, run setup, and select that model in the agent.