Kimi K3 is now available on Databricks
Kimi K3 is now available on @databricks!
All news · Today9 items
#Deployment and engineering
All collected and published items, ordered by time. Picks are a selected subset.
Kimi · X30 Hugging Face BlogPicks62 NVIDIA releases open-weights Magpie TTS multilingual model, adding Arabic, Korean, and Brazilian Portuguese
NVIDIA has released Magpie TTS Multilingual, a 364M-parameter open-weights speech-synthesis model supporting 12 languages. The latest release adds Modern Standard Arabic, Korean, and Brazilian Portuguese, and improves quality in several existing languages through updated training data.
Why it matters: The article reports the 364M open TTS model's language expansion and measured TTFA on different GPUs, helping readers estimate latency budgets for self-hosted voice pipelines.
Hugging Face Blog48 Hugging Face proposes efficient knowledge distillation with offline top-K logits and fused, chunked KL loss
A new Hugging Face paper proposes two systems improvements that sharply reduce the training cost of distilling LLMs: cache the teacher model's top-100 logits offline so teacher and student do not need to reside in GPU memory simultaneously, and use a fused, chunked KL loss to avoid generating a full vocabulary-by-sequence-length matrix.
Hugging Face Blog22 Baseten joins the Hugging Face Inference Providers ecosystem
Baseten has officially joined the Hugging Face Hub's Inference Providers ecosystem. Its initial support covers chat and text-generation tasks, including open-weights models such as Kimi K3, DeepSeek V4 Flash, and GLM-5.2.
Microsoft ResearchPicks72 Microsoft Research open-sources Orchard, a framework for scalable agentic AI
Microsoft Research has released Orchard, an open-source framework centered on Orchard Env, a Kubernetes-based environment service that can be reused for different tasks and agent systems, including software engineering, web navigation, and personal assistants.
Why it matters: Microsoft Research has open-sourced a reusable agent-environment service and three training recipes, offering insight into training smaller models within realistic deployment frameworks.
Hugging Face Blog22 GPU management: why idle GPUs are becoming the new grounded airplanes
A Hugging Face blog argues that GPU utilization, not model intelligence, is AI's next real constraint. GPU costs accrue by calendar hours, but output is produced only during compute hours; after buying GPUs, companies face the challenge of keeping their clusters busy. Using aviation as an analogy, the article says even fully loaded clusters may waste much of their potential because GPUs run all day while demand does not.
Berkeley AI ResearchPicks62 From CUDA to MLX: how K-Search brings kernel-optimization experience to Apple Silicon
Berkeley AI Research and IBM Research have extended the evolutionary kernel-search framework K-Search to MLX and added a structured CUDA-to-MLX translation layer that automatically adapts existing CUDA-kernel knowledge to Apple Silicon.
Why it matters: The article details the CUDA-to-MLX translation layer and reports two kernel benchmarks, showing how kernel knowledge can be transferred across platforms.
Hugging Face Blog52 NVIDIA releases Cosmos-H-Dreams, a real-time generative simulator for surgical robots
NVIDIA has released Cosmos-H-Dreams, a real-time, action-conditioned generative simulator for surgical robots. It distills Cosmos-H-Surgical-Simulator into a causal few-step student model and runs through the FlashDreams inference library on a single NVIDIA RTX PRO 6000 GPU.
Hugging Face Blog48 Model routing is more than a classification problem: it is a systems-optimization problem
Across 417 tasks in the AppWorld Test Challenge with the same CodeAct agent, Claude Sonnet 4.6 cost $79 in total ($0.19 per task), while GPT-4.1 cost $155 ($0.37 per task). Cache-hit rate was the main factor.
Hugging Face Blog36 Profiling PyTorch, part 3: a deep dive into attention
Hugging Face's “Profiling in PyTorch” series continues with a third post analyzing attention using an NVIDIA A100-SXM4-80GB profiler. It shows that a naive implementation triggers an extra memory copy. Changing `masked_fill` to its in-place form, `masked_fill_`, removes one GPU kernel per forward pass.
Mistral AI50 Mistral Studio adds system-level versioning and governance for Prompts and Skills
Mistral Studio now offers customers centralized management for Prompts and Skills, each with version history, clear ownership, and traceability. Published versions are immutable. The system supports version comparisons, rollbacks, classification tags, audit logs, and direct access through an MCP server. Assets can move from being visible only to their creator to a workspace or the entire organization; data remains within the enterprise boundary.
Hugging Face BlogPicks62 Hugging Face brings vLLM's Transformers modeling backend to native speed
Hugging Face says vLLM's Transformers modeling backend now matches or exceeds the throughput of vLLM's hand-written native implementations across several LLM architectures. The backend uses torch.fx for static-graph analysis and ast-based source rewriting to fuse inference-specific operators dynamically at runtime, matching the speed of custom implementations.
Why it matters: The article gives concrete architectures and configuration details for matching native performance with the Transformers backend in vLLM, helping readers assess migration costs.
Hugging Face Blog24 Hugging Face and Amazon SageMaker AI add one-click deep linking to SageMaker Studio
Hugging Face and Amazon SageMaker AI have introduced deep-link integration. On supported model pages, users can click “Customize on SageMaker AI” or “Deploy on SageMaker AI” to open SageMaker Studio with the model preloaded and the environment configured automatically.
Google Research36 Google Research uses Google Maps routing interventions to ease urban congestion
In the first large-scale real-world study of its kind, published in Nature Cities, Google Research modified the Google Maps algorithm to direct about 2% of trips to alternative routes across 10 U.S. cities for six months. Median travel speeds on target roads rose by about 2%, median fuel-consumption rates fell by 0.5% to 1.0%, and each city could cut thousands of tonnes of CO2e emissions annually.
Hugging Face BlogPicks62 Microsoft Foundry adds hosted compute for Hugging Face models
At Build 2026, Microsoft announced Foundry Managed Compute and a Hugging Face model collection, bringing a curated catalog of open-weights models into Foundry Model Catalog, refreshed weekly and deployable to managed GPU infrastructure with one click.
Why it matters: The article describes the full path from model listing through automatic runtime upgrades to one-click deployment, offering a view of how the barrier to production use of open models may change.
Berkeley AI Research40 What happens when intelligence becomes cheap? Data systems for agents, operated by agents, and built by agents
Berkeley's EPIC Data Lab argues that as inference costs have fallen from around $30 per million tokens for GPT-4-level models in early 2023 to under $1 today—and below $0.10 at some vendors—data systems face three new challenges: systems designed for agents, systems operated by agents, and systems built by agents.
Hugging Face BlogPicks71 SkyPilot and Hugging Face introduce storage mounts with zero egress fees
SkyPilot and Hugging Face have released the `store: hf` storage backend. A single `hf://` URL and an existing HF_TOKEN can mount a Hugging Face Bucket or any model, dataset, or Space repository into a SkyPilot task, which can run across more than 20 clouds, Kubernetes, Slurm, and local environments.
Why it matters: The source details configuration for mounting Hub data across clouds and the zero-egress-fee mechanism, helping readers assess storage costs for multi-cloud training and inference.
Import AI52 Import AI 464: Fable writes GPU kernels, AI automation, and analog computing
Import AI 464 summarizes several research developments. Fable wrote a megakernel on KernelBench-Mega that achieved an 18.71× speedup with CUDA on an RTX PRO 6000 Blackwell, compared with 14.4× for Claude Opus 4.8, 11.14× for GLM-5.2, and 4.34× for GPT 5.5.
Hugging Face Blog37 Major update to Hugging Face Kernels: new kernel repository type, trusted publishers, and code signing
Hugging Face has introduced a “kernel” repository type on the Hub for the Kernels project, along with two security measures: trusted publishers and code signing. By default, only kernels from trusted publishers are loaded; other sources require explicitly passing `trust_remote_code=True`.
Hugging Face BlogPicks62 Hugging Face and Cerebras bring Gemma 4 to real-time voice AI
Hugging Face and Cerebras demonstrated a real-time speech-to-speech pipeline using Cerebras to accelerate Gemma 4 31B inference, bringing voice-interaction latency closer to natural conversation.
Why it matters: The source presents a replaceable open-source speech-to-speech pipeline, helping readers assess engineering combinations for low-latency voice interaction.
Google ResearchPicks60 Google speeds up Gemini Nano on Pixel with frozen multi-token prediction
Google Research has introduced an architecture that attaches a multi-token-prediction (MTP) head to the frozen Gemini Nano v3 model, speeding up on-device generation without changing the backbone weights.
Why it matters: Google attaches an MTP head to frozen Gemini Nano v3 and reports concrete methods and gains under on-device memory and energy constraints.
Google Research31 Google lowers cloud costs with linear elastic caching: Spanner memory use drops 15.5%
Google Research has proposed linear elastic caching, modeling page eviction as a ski-rental problem and using lightweight machine learning to adjust cache size dynamically. After months of testing in Spanner production, memory use fell 15.5%, cache misses rose just 5.5%, total cost of ownership fell about 5%, and the impact on actual I/O costs was only 0.5%.
Mistral AI52 Mistral AI adds admin controls, API-key scoping, and a debugger to Connectors
Mistral AI has added several capabilities to Connectors. Admin controls are generally available, letting workspaces or organizations set access and toggle tools individually. Scoped API keys, also generally available, prevent automated AI workloads from impersonating a user. Multi-account connectors, also generally available, let one connector link multiple accounts.
Mistral AIPicks62 Mistral releases OCR 4 with 170-language support and single-container self-hosting
Mistral has released OCR 4, which returns bounding boxes, block-type classifications, and per-page, per-word confidence scores in addition to extracted text. It supports 170 languages across 10 language families and can be fully self-hosted in a single container.
Why it matters: OCR 4 returns bounding boxes, block classification, and confidence scores, helping readers assess how document parsing can fit into RAG and agent workflows.
Google Research36 Google supports UC San Diego's low-carbon compute platform built from 2,000 retired Pixel phones
With Google's support, UC San Diego is using motherboards from 2,000 retired Pixel phones to build a data center offering low-cost, low-carbon cloud computing for hundreds of students and faculty. SPEC benchmarks show that 25–50 phones roughly equal one modern server; the phones are organized into self-managing Kubernetes clusters of 25–50 devices. The system is expected to launch in fall 2026.
Google DeepMindPicks71 Google DeepMind releases DiffusionGemma, generating text up to 4× faster
Google DeepMind has released DiffusionGemma, an experimental open-source model that uses text diffusion to generate entire chunks of text in parallel. It generates text up to 4× faster on specialized GPUs, exceeding 1,000 tokens per second on a single NVIDIA H100 and 700 tokens per second on an RTX 5090.
Why it matters: Google DeepMind's open text-diffusion model comes with concrete speed and hardware figures, helping readers assess the feasibility of local interactive inference.
Mistral AIPicks67 Mistral launches industrial-engineering AI stack and Vibe agent, and announces a 10 MW data center in France
At AI Now Summit 2026, Mistral launched an AI stack for industrial engineering and announced work with Airbus, BMW, and ASML to optimize design, simulation, and production while retaining control of proprietary data and IP.
Why it matters: Mistral announced an industrial AI stack, the Vibe agent, and its own inference data center, showing the shape of its enterprise full-stack strategy.
Mistral AIPicks62 Mistral AI releases public preview of Search Toolkit
Mistral AI has released the public preview of Search Toolkit, a composable framework for building production-grade search pipelines for AI applications. It is open-source and supports cloud, on-premises, and edge deployments. Ingestion, retrieval, and evaluation share one interface, with built-in BM25 sparse retrieval, dense-vector retrieval, and hybrid configurations, plus metrics such as recall, precision, MRR, and NDCG.
Why it matters: Mistral brings ingestion, retrieval, and evaluation into one framework, offering a way to assess RAG retrieval quality separately.
Mistral AIPicks62 Mistral introduces physical AI, integrating Emmi AI into its enterprise platform
Mistral says it is bringing Emmi AI into the company and launching physical-AI capabilities for industrial engineering as a new foundational feature of its enterprise solutions. The models predict physical fields directly from geometry and boundary conditions in a single forward pass taking seconds on one GPU, enabling design iteration. Traditional solvers remain available for verification and edge cases.
Why it matters: Mistral is integrating physics-simulation models into its enterprise platform, offering a view of potential changes to compute requirements and iteration speed in industrial design.
Mistral AIPicks62 Mistral launches Connectors in Studio with built-in and custom MCP support
Mistral has released Connectors in Studio as a public preview. Built-in and custom MCP connectors can be used through the API or SDK with all models and agent calls. New direct tool calls let developers decide when tools are invoked, and human-in-the-loop approval lets applications confirm before a tool executes.
Why it matters: The source gives code paths for registering and calling connectors and for human approval, helping readers assess changes to enterprise tool integration.
Berkeley AI Research34 Adaptive parallel reasoning: the next paradigm for efficient inference scaling
A Berkeley AI Research team reviews progress in parallel reasoning and notes that most current methods—including Self-consistency, Best-of-N, Tree of Thoughts, MCTS, ParaThinker, GroupThink, and Hogwild!—
Google DeepMindPicks80 Google DeepMind reviews a year of AlphaEvolve, from mathematical discoveries to TPUs and commercial deployment
Google DeepMind has published a one-year impact review of AlphaEvolve, the Gemini-powered coding agent. It has expanded from exploring open problems in mathematics and computer science to becoming a core component of Google's infrastructure.
Why it matters: The official review summarizes AlphaEvolve deployments over its first year, helping readers assess how algorithmic agents are being used in science and infrastructure.
Mistral AIPicks71 Mistral AI releases public preview of Workflows
Mistral AI has released Workflows as a public preview of its enterprise AI orchestration layer. It offers persistent execution, observability, and fault tolerance to take AI processes from proof of concept into production. Organizations including ASML, ABANCA, CMA-CGM, France Travail, La Banque Postale, and Moeve already use it to automate key workflows.
Why it matters: The source describes Workflows' persistent execution, observability, and human approval features, helping readers assess the orchestration requirements for enterprise AI deployments.
Google DeepMindPicks62 Google DeepMind releases Decoupled DiLoCo distributed-training architecture
Google DeepMind has released Decoupled DiLoCo, a distributed-training architecture that divides large-scale training into compute islands communicating asynchronously, reducing bandwidth requirements and isolating hardware failures. In tests, training continued and reconnected after an entire learner unit went offline, achieving the same ML performance as traditional methods with Gemma 4 models.
Why it matters: The source reports distributed pretraining results across bandwidth, fault tolerance, and mixed-generation hardware, helping readers assess engineering tradeoffs.
Mistral AI36 Mistral AI launches Spaces, a CLI for humans and AI agents
Mistral AI has released Spaces CLI, a command-line tool for both human developers and AI agents. Commands such as `spaces init` and `spaces dev` can scaffold a multi-service project with hot reload, a database, and a Dockerfile.
Google ResearchPicks62 Google releases TurboQuant vector-compression algorithm, reducing KV cache to 3 bits without loss of accuracy
Google has released TurboQuant, a compression algorithm that quantizes KV cache to 3 bits without training or fine-tuning and without loss of model accuracy, reducing KV-memory use by at least 6×.
Why it matters: Google reports TurboQuant's algorithm details and benchmark results, helping readers assess the usefulness of KV-cache compression for long-context workloads.
Mistral AIPicks62 Mistral AI launches Forge, an enterprise model-training system
Mistral AI has released Forge, a system that lets enterprises build frontier-level AI models from their own proprietary knowledge. It spans pretraining, post-training, and reinforcement learning; supports dense and MoE architectures; and accepts multimodal inputs.
Why it matters: The source details the stages of training frontier models on enterprise data and names organizations already involved, helping readers assess the path and barriers to building models in-house.
Mistral AI34 Heaps do lie: investigating a memory leak in vLLM
The Mistral AI team investigated a memory leak in vLLM. In a Mistral Medium 3.1 deployment with graph compilation and disaggregated prefill/decode, system memory grew linearly by 400 MB per minute until the process ran out of memory hours later. Heaptrack showed stable heap usage, indicating the leak was outside the heap. The team traced it to KV-cache transfers through NIXL and UCX.
Mistral AIPicks65 Mistral releases OCR 3, with a 74% win rate over its predecessor on forms and handwriting
Mistral AI has released Mistral OCR 3, reporting a 74% overall win rate against Mistral OCR 2 on forms, scanned documents, complex tables, and handwriting. The company says it outperforms both enterprise document-processing and AI-native OCR solutions in accuracy.
Why it matters: The official post compares win rates with the previous version on forms, handwriting, and tables and gives pricing per thousand pages, helping readers assess document-parsing costs and use cases.
Mistral AIPicks82 Mistral releases Mistral 3 family, from 675B Mistral Large 3 to 3B, 8B, and 14B Ministral 3
Mistral has released the Mistral 3 family: dense small models at 3B, 8B, and 14B, plus Mistral Large 3, a sparse MoE model with 41B active and 675B total parameters. All are open-source under Apache 2.0.
Why it matters: Mistral has released Apache-2.0 weights ranging from 3B to 675B, giving readers a view of open-model options from edge devices to frontier scale.