AI news — March 2026
Every AI story covered in March 2026, summarized. From the @rotative_ai daily feed.
stories
- Google releases Veo 3.1 Lite, cuts Veo 3.1 Fast pricing
google released veo 3.1 lite, a video generation model costing less than 50% of veo 3.1 fast while maintaining speed. it supports text-to-video, image-to-video, 720p/1080p resolution, and 4-8s durations. on april 7, pricing for veo 3.1 fast drops. available now via gemini api.
- IBM releases Granite 4.0 3B Vision with ChartNet
ibm launched granite 4.0 3b vision, a 3-billion parameter model optimized for enterprise documents. it features chartnet, a 1.7 million sample dataset for chart reasoning, and deepstack architecture for precise spatial understanding. available as a lora adapter on hugging face.
- CodonRoBERTa beats ModernBERT in mRNA optimization
openmed introduced codonroberta-large-v2, a transformer for mrna optimization that outperforms modernbert. trained across 25 species for $165 using 55 gpu-hours, the model achieves a perplexity of 4.10 and spearman cai correlation of 0.40.
- Hugging Face releases TRL v1.0 for post-training
hugging face released trl v1.0, a post-training library implementing over 75 methods including ppo, dpo, and grpo. the library handles 3 million monthly downloads and supports a chaos-adaptive design to manage rapidly shifting ai paradigms. it separates stable and experimental code under semantic versioning.
- Hugging Face outlines migration path for OpenClaw agents to open models
hugging face published a guide for migrating 'openclaw', pi, or open code agents to open-source models. users can choose between the hosted inference providers platform or running fully local models via llama.cpp. the guide specifically recommends the glm-5 model for its terminal bench scores or qwen3.5-35b-a3b for local setups.
- Google's James Manyika discusses AI and creativity with LL COOL J
google svp james manyika and ll cool j discussed the intersection of ai and creativity in a recent dialogue. manyika highlighted the potential for generative ai to democratize artistic tools. ll cool j reflected on his 40-year career and the need to protect the human element in art.
- Google Translate brings live translation to iOS headphones
google ai launched live translate with headphones for ios, expanding support to the uk, france, germany, italy, japan, spain, and thailand. the feature translates over 70 languages in real-time through any connected headphones. users must update the translate app and tap live translate to begin.
- Hugging Face releases EVA framework for voice agents
hugging face introduced eva, a framework that evaluates voice agents on both task accuracy and conversational experience using a bot-to-bot architecture. benchmarking 20 systems revealed a consistent accuracy-experience tradeoff. the release includes an airline dataset with 50 scenarios.
- Build domain-specific embedding models in under a day
nvidia and hugging face released a recipe to fine-tune embedding models on domain-specific documents in under a day using a single gpu. the pipeline uses llms to generate synthetic qa pairs and hard negative mining, eliminating manual labeling. internal tests on jira data improved recall@60 by 26% and nvidia docs data boosted recall@10 by over 10%.
- Hugging Face hits 2M models as China overtakes US in downloads
hugging face reports 13 million users, 2 million public models, and 500,000 datasets for spring 2026. the data shows china surpassed the us with 41% of downloads. industry contribution dropped from 70% to 37%, while 30% of fortune 500 companies now hold verified accounts.
- H Company releases Holotron-12B computer-use agent
h company released holotron-12b, a 12b parameter computer-use agent model post-trained from nvidia's nemotron-nano-2 vl. the hybrid ssm architecture achieves 2x higher throughput than holo2-8b on webvoyager benchmarks, scaling to 8.9k tokens/s on a single h100 gpu with 100 concurrent workers.
- Survey of 16 open-source async RL libraries
huggingface surveyed 16 open-source rl libraries solving the gpu idle time problem. synchronous rl training keeps gpus idle 60% of the time because data generation takes hours. the solution is disaggregating inference and training using rollout buffers and async weight transfers. ray dominates orchestration, nccl is the default for weight sync, and distributed moe support is emerging as the key differentiator.
- Hugging Face launches Storage Buckets
hugging face introduced storage buckets, an s3-like object storage solution on the hub backed by xet. this feature offers mutable storage, deduplication for ml artifacts, and pre-warming capabilities for aws and gcp. users can manage buckets via the hf cli or python.
- Hugging Face integrates Ulysses sequence parallelism for long context training
hugging face integrated ulysses sequence parallelism into accelerate, transformers, and trl. developed by snowflake ai research, this method distributes attention computation across gpus to support million-token context training. it reduces communication overhead compared to ring attention by leveraging all-to-all collectives.
- LeRobot v0.5.0 adds Unitree G1 humanoid support and Pi0-FAST policies
lerobot v0.5.0 released, adding full unitree g1 humanoid support with whole-body control. introduces pi0-fast autoregressive vlas, real-time chunking for responsive inference, and streaming video encoding. includes envhub, nvidia isaaclab-arena integration, and modernization to python 3.12+.
- NXP details embedded VLA robotics pipeline optimizations
nxp published a guide on deploying vision-language-action models on embedded platforms like the i.mx 95 soc. the report details best practices for recording datasets, fine-tuning act and smolvla policies, and optimizing for asynchronous inference to meet real-time latency constraints.
- Hugging Face releases Modular Diffusers framework
hugging face launched modular diffusers, a framework breaking diffusion pipelines into composable blocks. the system allows custom block creation, independent execution of components like text encoding or denoising, and integrates with the visual interface mellon. the community released a 14b parameter real-time video model and a 2.3b world model using the new architecture.
- Hugging Face trains competitive text-to-image model in 24 hours for $1500
hugging face trained a competitive text-to-image diffusion model in just 24 hours using a $1500 compute budget. the 'prx part 3' speedrun utilized x-prediction to train directly in pixel space without a vae, starting directly at 512px resolution. they also incorporated l1pips and dino-based perceptual losses to boost convergence speed and visual quality. the team open-sourced the training code and experimental framework on github.