Qoder Blog7070Qoder Fast reduces the first response time from 8 seconds to 3 seconds through client optimization, parallel tool execution, model routing, and network improvements, supporting smoother AI collaboration.
Databricks BlogDatabricks has integrated the RFD4-Proteina model into Genesis Workbench, supporting execution in AI runtime environments and connecting research data and workflows. Developers can build and adjust pipelines using Genie Code and the NVIDIA BioNeMo Agent Toolkit, leveraging serverless NVIDIA GPU computing for protein design.
OpenHands Blog6363Recursive Flint is released, a hybrid expert model with 4 billion parameters, designed for efficient local inference, activating only about 250 million parameters per token, achieving 125+ tokens/second on M4 Pro MacBook Pro, and reaching 150+ tokens/second on NVIDIA RTX Spark. OpenHands, as its first partner, supports the model's operation on local hardware, providing developers with an open-source coding model option.
vLLM BlogAnt Group and the vLLM community open-source Humming, a high-performance quantized inference kernel library. Humming supports NVIDIA GPUs starting from the Turing architecture and CUDA-compatible devices like the T-Head Zhenwu 810E, providing extensive quantization support covering various integer and floating-point types, bit widths, and scaling schemes. It significantly outperforms existing implementations such as vLLM Marlin, FlashInfer, and DeepGEMM on multiple representative workloads.
Google Developers BlogAgents CLI 1.5.0 version introduces an extension system, allowing developers to override or add commands without modifying the tool, supporting cross-framework usage; the official LangChain extension replaces framework-bound commands while retaining other lifecycle functionalities; developers can create custom extensions, such as adding deployment commands for software bill of materials (SBOM) scanning, enabling enterprise platform teams to enforce governance policies.
Baseten LabsSame story · 2 outlets6060Baseten Labs releases the world's top decision models, Elephant-1 (9B parameters) and Mammoth-1 (27B parameters), both ranking first in the Jev Decision Index and Artificial Analysis Decision Index evaluations.
Browser Use3838Browser Use Cloud MCP servers support proxies to submit complete web tasks in plain text form, which are processed by their cloud browser agents. The cost is $0.02 per hour, and up to 10 projects can run simultaneously. This service uses API V3, offering 11 model selections. The browser is a reinforced version of Chromium, with incognito mode and Cloudflare proxy support enabled by default.
Apple Machine Learning ResearchTEAL has for the first time efficiently applied active learning to the token space of ViT for low-budget pixel-level semantic segmentation, boosting mIoU by +24.36 on datasets like CamVid and Cityscapes with only one pixel labeled per image, and achieving full-supervised performance of 97% with 10 pixels labeled.
a16z News7070OpenAI has acquired Orchestra Bio. Current cutting-edge models like GPT-6 Astra, Fable 5.1, and Grok 4.6 can generate CAD, design PCBs, and describe assembly, but hardware design benchmarks like CAD Arena, EE Bench, and LuxoBench show that improvements are still needed. AI in hardware design faces the lossy problem of the interface between the digital and physical worlds, even if it can generate designs, a lot of work is still required.
Databricks BlogDatabricks introduces Agent Memory, which improves agent quality, reduces costs and latency by storing key facts, preferences, and processes, and supports running on Lakebase. It is currently in Beta version. Agent Memory is based on Databricks' latest memory research and uses the MemCurrent architecture. In internal benchmark tests, it has tripled the pass rate of agent tasks while reducing token usage by 31%.
OpenAI Alignment Blog5050OpenAI has introduced the complementary metric g-mean² to improve model monitorability evaluation. g-mean² calculates TPR and TNR through behavioral labels and monitoring predictions, but cannot solve the problems of mixed differences in different model tasks and deployment risks. The new metric includes a paired matching metric for aligned shared instances and a post-control integrated metric simulating monitoring rejection of outputs, to more comprehensively evaluate model deployment risks.
Hugging Face Blog5858ALTK-Evolve uses GPT-4.1 to learn and apply customer service scripts. In the τ²-bench customer service agent benchmark test, it surpasses all retail entries in the retail sector, achieving a score of 67.5% Pass^4, and does not require updating model weights. The model is based on a basic ReAct agent and GPT-4.1. It attempts to analyze tasks and use GPT-4.1 to distill success and failure into short behavioral rules, then adds these guidelines to the agent's system prompts without updating weights or manual annotation.
Tessl BlogTessl Blog stops line-by-line code review and starts verifying intent. The team's merged PR volume is four times that of last year and almost all AI-generated, making review a bottleneck. The team shifts to reviewing intent before code generation (acceptance criteria, constraints, and edge cases), and transforms repetitive review comments into machine-executable rules (invariants), thus shifting the focus of review from the code itself to decision-making and rule ownership.
Tessl BlogTessl Blog published key design decisions for building cross-company AI agent memory systems, proposing to store statements rather than data blocks, explicitly express time and missing states, and determine permissions before model execution.
Tessl Blog7575Engineering teams' trust in AI code assistants determines their scope of use, ranging from viewing code differences to full authorization for merging. Trust is built through rapid validation mechanisms. Development platform costs increase with the intensity of AI agent work, with about 80% of the budget consumed by 20% of heavy users, requiring an evaluation of the effective cost per unit of work. Successful teams start with auxiliary tasks and gradually expand the scope of AI, rather than applying it directly to core systems.
Cohere BlogCALIBER estimates success probability before inference through training reasoning models and evaluates the confidence of specific answers after inference, significantly reducing calibration error. This 7B model on the BigMathDigits benchmark reduces ECE from 0.059 to 0.028 (a reduction of 52.5%), while achieving an AUROC of 0.806 and a Brier score of 0.130, proving that confidence evaluation is more reliable and can effectively distinguish between correct and incorrect answers.
Mistral AISame story · 2 outletsMistral AI has launched Voxtral 3, which includes an advanced voice model capable of real-time multi-step conversations in over 50 languages, along with a voice agent SDK. Its voice-to-text model, Voxtral Small Transcribe 3, achieved a Forced WER of 0.9% on the FLEURS 45 benchmark.
How the accounts differ:The Decoder mentions a text-to-speech model (Voxtral Mini TTS 3) that is not detailed in the first report.
IBM Research BlogThe llm-d project announces the release of version 1.0, which is a Kubernetes-native distributed inference service network that supports deploying any model's inference stack on any accelerator in any cloud. This version enhances multi-tenant quality of service, zero-downtime operations, and agent-aware services to achieve continuous performance in common production operations. Through techniques such as prefix cache-aware routing and KV cache offloading, llm-d 1.0 can capture 97% of available cache reuse, reducing the cost of large-scale AI services.
Hugging Face Blog5858AnyLanguageModel 1.0 is released, supporting local model advancements on Apple platforms for a year. New support added for cloud models such as OpenAI Chat Completions, Anthropic Messages, and Google Gemini. Local model support includes llama.cpp, MLX, and Core ML. Tool calling and guided generation features are expanded. Cloud models have the same functionality as local models, and API keys support dynamic acquisition through closures.
Google Developers BlogGoogle has released the native PyTorch backend for vLLM and SGLang inference engines, supporting TPU and optimizing production-level services; vLLM-TorchTPU achieved a cost of 2.01 cents per million tokens for 175B parameter models in the SemiAnalysis InferenceX benchmark, leading the industry.
Hamel Husain6060When using a decision model for evaluation, it is necessary to adjust the threshold according to the task to optimize performance. On the log dataset of an apartment rental agent, Jev 1.13, Perplexity Decider 1.1 27B.
vLLM BlogvLLM TorchTPU (vllm-torchtpu) released, enabling vLLM to run on Google Cloud TPU, supporting Pallas core's attention, Mixture-of-Experts, and quantized matrix multiplication while maintaining the vLLM experience, supporting six-dimensional parallelism, and recommending TPU v7x (Ironwood).
Modal Blog5858Modal introduces Sticky Sessions functionality, implemented through the @modal.sessioned() decorator, to achieve strong affinity routing at the session level, ensuring all requests within the same session always reach the same container. It is suitable for workloads requiring state continuity, such as speech-to-text, voice proxying, and multiplayer games. This feature is based on the Modal Server architecture, authorizes by returning a signed session token, implements load balancing through container concurrency probing, and automatically scales sessions based on active session count.
Anthropic ResearchGoogle Developers BlogCoral NPU has released a complete 32-bit architecture, including new matrix execution units, supporting up to 1 TOPS performance and 10mW average power consumption; launched the Coral Compiler v1.0 based on IREE/MLIR.
Google Developers BlogArm leveraged the hierarchical graph visualization and plugin architecture of Google AI Edge Model Explorer to develop open-source adapters supporting formats such as TOSA, VGF, and PTE, helping developers more easily inspect edge model conversion processes.
Browser UseBrowser Harness achieves a score of 60.4 in matching evaluation using 2.1M median input tokens, matching Playwright MCP and Chrome DevTools MCP, becoming the browser control tool with the lowest cost per task at $0.042. Playwright MCP and Chrome DevTools MCP are faster but use more tokens.
Hugging Face BlogHugging Face has released TorchTPU, making TPU a native backend device for PyTorch. Transformers, Diffusers, and Kernels can now run directly on TPU. Users only need to select a model from the Hugging Face Hub and specify device_map="tpu".
Meta EngineeringGenRecs Tokenization enhances conversion rates by 0.35% for advertisers by representing items like ads and products with a stable semantic vocabulary, combined with Meta's Generative Ad Model (GEM) and adaptive ranking model. This technology leverages the world knowledge of pre-trained large models to enable recommendation systems to understand preferences of new items or users with no historical interaction, accelerates learning through shared concepts, achieves cross-activity transfer, and improves performance as large model scale increases, forming a feedback loop for performance improvement.
O'Reilly Radar6565The Monday Morning Agent project uses the NFL public dataset to analyze a controversial American football topic every two weeks, gradually equipping AI agents with analytical capabilities. The project uses 250,000 game clips and current season data, each episode having AI explain a specific tactic (such as 'establishing the offensive line') and verify its analysis results. The AI must state data limitations during analysis and refuse requests beyond the task scope (such as predicting game outcomes).
Meta EngineeringMeta engineers built a lean router with 1,300 lines to assign over 20 workflows to 14 specialized AI agents, with the router itself not performing any domain-specific tasks, significantly improving the efficiency and accuracy of data pipeline agents.
Sakana AI5555Sakana AI proposed 'QD-Rec', recommending candidate-hypothesis pairs as units, with generative AI writing new hypotheses based on past excellent hypotheses in a cycle to accumulate high-quality and diverse recommendations. Experiments were conducted in M&A, customer targeting, and paper recommendations, achieving high average evaluations in many settings and building large sets of recommendations.
LangChain BlogNu company, while developing AI health coach Nu, prioritized building evaluation criteria to avoid boundary-crossing medical diagnoses. Using the LangGraph framework and LangSmith tools, it ensures the system meets health guidance needs while complying with regulatory requirements through LLM judges scored by category and open-ended behavior tests. The evaluation framework is designed separately for single responses and complete conversations, supporting rapid migration between two leading model providers. Safety probes and guard nodes review conversation content in real-time, ensuring the product stays within compliance boundaries.
Modal Blog7575NVIDIA has delivered its first Vera CPU server to Modal, an Arm-compatible CPU that, through emulation capabilities testing, has increased the throughput of the emulation capability sandbox by 2x, suitable for heterogeneous workloads with peak resource usage up to 10x the baseline, helping to reduce costs and accelerate reinforcement learning deployment.
n8n Blogn8n workflows can run locally on NVIDIA DGX Spark without calling cloud APIs. By deploying the 35B hybrid expert model Qwen3.6-35B-A3B (NVFP4 format) and 8 other models on DGX Spark, tasks such as transcription, OCR, and image generation are supported. All models are accessed by n8n through OpenAI-compatible endpoints. Deployment requires caching models and using version 0.30.0 of vLLM to ensure offline operation and license compliance.
The Qwen3-4B-deslop model eliminates common AI writing habits (such as dashes, 'not X, but Y' sentences, three-element lists, etc.) through fine-tuning, reducing the SLOP score from 121 to 12.5 on 40 unseen prompts, with 68% of responses reaching human writing level. This 4B model, after being quantized by Ollama, is only 2.5GB and can run locally while still maintaining a 30/30 fact question accuracy rate comparable to the original version.
IBM Research Blog2222Coursera's new course 'AI Data Stream Architecture: From Principles to Practice' teaches engineers how to work across GPUs, TPUs, Cerebras WSE, Groq LPU, and IBM Spyre accelerators. The course is taught by Kaoutar El Maghraoui, a senior research scientist at IBM Research, covering execution models, compilation, memory systems, debugging, and performance modeling through six courses and hands-on experiments, emphasizing how different architectures affect model performance and the key role of compiler optimization.
