Release 12.4

Release Date: May 7th, 2026

---

New Feature	Improvement	Bug Fix	Enterprise Only

Published Models

Status	Change	Details
	Published new models	Published Nemotron Nano V3 Omni (30B total / 3B active — MoE). It’s NVIDIA's multimodal LLM that unifies video, audio, image, and text understanding in a single model, with integrated reasoning support. Published Qwen3.6-35B-A3B, an efficient MoE LLM with 35B total but only 3B active parameters, delivering near-large-model performance at much lower compute cost.

Cached Prompt Tokens

Status	Change	Details
	Introduced cached prompt tokens in model responses	Model responses now include cached prompt token counts, providing accurate token usage reporting when prompt caching is active. Learn more here.

Pipelines

Status	Change	Details
	Introduced code-first Pipeline DSL with CLI support	Introduced a code-first Pipeline DSL that lets you define pipelines programmatically in Python and generate configs via CLI. Use clarifai pipeline init to scaffold the pipeline config and clarifai pipeline upload to deploy it to the platform. Learn more here.
	clarifai pipeline run --dev for local pipeline development	New --dev flag on clarifai pipeline run enables a local development loop, letting you iterate on pipeline steps without deploying to the cloud. See PR #1012 for details.
	clarifai pipeline local-run for Docker-based step testing	New clarifai pipeline local-run command runs individual pipeline steps locally inside Docker containers, matching the production runtime environment before you deploy.
	Auto-create compute resources via --instance flag	clarifai pipeline run now accepts an --instance flag that automatically creates the required compute cluster and nodepool if they don't already exist.
	Improved clarifai pipeline init UX and help text	Clearer help text, improved prompts, and a post-init next-steps message now guide users through the full pipeline setup flow after running clarifai pipeline init.

Model Deployment

Status	Change	Details
	Disabled deploy_latest_version for clarifai model serve	clarifai model serve no longer automatically promotes a newly uploaded version as the live deployed version, giving you explicit control over which version is active.

Local Runners

Status	Change	Details
	Local runner defaults to PRIVATE; --public flag patches all visibilities	Models and resources created via the local runner are now private by default. Pass the --public flag to make all associated resources public in a single command.

Bug Fixes (Python SDK)

Status	Change	Details
	Fixed deployment worker pinning on clarifai model serve	Re-pins the deployment's desired_worker to the current model version when running clarifai model serve, preventing stale version references after serving.
	Fixed Hugging Face private repo access validation	Corrected access validation for private Hugging Face repos that return a not_found response to anonymous requests, eliminating false access errors on valid private repos.
	Loosened pinned requirements and fixed Clarifai package detection	Relaxed overly strict version pins in the SDK's dependencies and fixed a bug in detecting the installed Clarifai package version.
	Fixed User.app() returning empty values	User.app() now correctly returns actual server-side app data instead of empty placeholder values.