Hugging Face published a tutorial on the blog showing how to use the TRL library to fine-tune LiquidAI/LFM2.5-350M with GRPO to improve structured output compliance, requiring only about 500 samples and 100 training steps, runnable on free-tier Colab or Kaggle GPUs.
Sentence Transformers v6.0 adds the MultiVectorEncoder model type, and this tutorial demonstrates how to train and finetune ColBERT-style multi-vector retrieval models.
Why it matters: The author backs every training choice with measured data, teaching a multi-vector model domain finetuning method reproducible on a single GPU.
Hybrid bonding is becoming a key interconnect technology for next-generation products such as HBM, NAND, 3D-stacked DRAM and AI accelerators. It eliminates bumps and uses direct Cu–Cu bonding, shrinking chip pitch from about 100µm with solder balls and about 20µm with micro-bumps to under 1µm.
Hugging Face explains how it built a hybrid search engine for Papers with Code using Inference Endpoints, Jobs, and Storage Buckets.
Hugging Face announced on its official blog that Sentence Transformers v6.0 adds a fourth model type, MultiVectorEncoder, supporting ColBERT-style late interaction retrieval; any PyLate and Stanford-NLP ColBERT checkpoint can be loaded directly, and colpali-engine models can also be used for visual document retrieval.
Why it matters: Official tutorials cover the full pipeline of loading, scoring, indexing, and compression, so readers can use them to decide whether multi-vector retrieval is worth adopting in their own search stack.
Hugging Face built a constraint-aware GPU allocator and benchmarked it against a FIFO scheduler across seven scenarios. On identical hardware and workloads, GPU utilization rose by up to 33 percentage points (for example, from 53.6% to 87.0% for an 8-GPU training-heavy workload), priority-weighted output increased in every scenario by up to 105%, averaging 52%. The key change is allocating real-time inference demand across time steps along its curve and placing batch-like tasks by priority across the whole scheduling horizon, rather than assigning by arrival order.
Hugging Face demonstrated a streaming data loop in Strands Robots (an AWS open-source SDK): an agent records robot demonstrations in LeRobot format and syncs them to Storage Buckets.
Hugging Face's ALTK-Evolve and ACE are both agentic memory systems that turn an agent's historical trajectories into reusable experience injected at inference time without weight updates; both refuse to compress experience, differing in how they deliver it.
Hugging Face has published a paper that enables offline distillation by caching the teacher's top-100 logits, and introduces a fused chunked KL loss that avoids generating full vocabulary-by-sequence-length matrices.
DharmaOCR scored 0.925 on a Brazilian Portuguese OCR benchmark, beating the newer Mistral OCR4 (0.798) and Unlimited-OCR (0.7587), confirming the value of domain specialization. The model uses supervised fine-tuning to align with Brazilian Portuguese vocabulary and document structure, then DPO to suppress repetitive or incoherent output, achieving the lowest degradation rate. The article argues that the reliability of generative OCR depends on how concentrated the parameters are within a single domain, not on a larger architecture.
Hugging Face published part three of its Profiling in PyTorch series, using the profiler on an NVIDIA A100 to compare traces of six attention implementations: naive attention, in-place masking, and SDPA math/efficient/flash/cuDNN.
Why it matters: The original benchmarks kernel-level differences across six attention implementations with the profiler, teaching readers to read from traces how in-place saves a copy and what each SDPA backend really costs.
Hugging Face details the data pipeline for its PRX image generation model: mixing public and internal datasets for pretraining, using a VLM to regenerate long captions, then converting them into streamable training data. The text encoder has been switched to Qwen3-VL, with text latents computed online during training at a throughput cost of only about 3–4%. Images are uniformly encoded as JPEG quality 92, and the data is built with Lance and streamed for training via MDS (Mosaic Streaming).
Mistral AI built an autonomous agent that reads Rails source files, generates or improves RSpec tests, and runs unattended in CI/CD pipelines. The agent is based on Mistral's open-source coding assistant Vibe, using a repo-level AGENTS.md, skill files divided by file type, and custom tools for RuboCop and SimpleCov.
The Mistral AI team investigated a memory leak of roughly 400 MB per minute in vLLM when running Mistral Medium 3.1 with graph compilation enabled in P/D disaggregated deployments.
Why it matters: The article documents the full process of localizing the leak layer by layer from pmap and BPFtrace to GDB, which can be applied to similar kernel-level debugging of inference services.
Mistral AI published a tutorial showing how to use the Mistral fine-tuning API and LaPlateforme UI to fine-tune Pixtral-12B with LoRA on the Aerial Image Dataset (AID) satellite image classification task.
Mistral explains how to evaluate RAG systems using “LLM As A Judge”: a judge LLM scores the answers of a generator LLM on numeric, binary, or qualitative scales.
Mistral AI has launched TranscriptToPRDTicket, an agentic workflow that automatically turns meeting transcripts into PRDs and actionable development tickets. The workflow consists of two components, PRDAgent and TicketCreationAgent, both powered by Mistral Large 2, and can automatically create structured tickets in project management tools such as Linear and Jira. The full implementation is available via a Google Colab notebook.