Hugging Face Daily Papers6060Claude Opus 5.5, combining Claude Code and GPT-6.1 Sol, completes 9 out of 46 real-world tasks under the HEELO framework alone, 39 tasks and 35 tasks respectively when connected to the MCP server, with time consumption reduced by 1.9 times and 2.3 times, and cost reduced by 1.55 times
llama-modchip releases a set of ggml/llama.cpp core patches and quantization transcoding tools, compressing the 284.33B DeepSeek-V4-Flash to 56.56 GiB (1.71 bpw), which can run on a pure CPU machine with 64 GB memory, with a peak memory of 57.61 GiB under 256K context and no need for swap.
Why it matters: The project provides a quantization scheme and core patches for running 284B MoE on 64 GB memory, along with measured data for each hardware, which can be used to determine the feasible boundary for pure CPU deployment.
Qualcomm showcased its edge-side MoE intelligent agents at the 9th CIIE, driving edge-side AI from single-point applications to intelligent agent experiences. MoE technology enhances the intelligence level of edge-side models through mixed expert modules, achieving higher complexity task completion in the 2 billion to 3 billion parameter scale. Qualcomm collaborated with Infris AI, Changelong, and Jieyu Xingchen to optimize models, inference, storage, and chip platforms, demonstrating a systematic solution for edge-side AI ecosystem synergy.
MarkTechPost6262Google AI open-sources the Elephant-Goldfish Model (EGM) Plugin, enabling AI coding agents to write specifications before writing code. The plugin includes 7 skills, 9 reviewer roles, and 6 brainstorming lenses, enforcing independent specification reviews through 5 zero-context review stages.
MarkTechPost3838DeepSeek-V4-Pro, Qwen3.6-35B-A3B, and ten other open-source MoE models differ in five aspects: expert count (8B-384B), top-k (1-8), shared experts (7/10 models retained), scoring function (Sigmoid/sqrtsoftplus), and load balancing (Aux loss/bias).
MarkTechPost3838MoE routing strategies determine which experts process each token, affecting model quality, load balancing, and deployment costs. Mixtral uses Softmax Top-K to select 2-8 experts, while DeepSeek-V4-Pro selects 1-8 out of 384. Load balancing is achieved through Sigmoid gating and auxiliary loss, with Sigmoid scoring combined with lossless bias balancing being a common approach for cutting-edge open models in 2026.
OpenAI Agents SDK releases2222OpenAI Agents SDK v0.24.0-alpha.7 released, supports classic workflows and native execution, adds voice and managed tools, introduces persistent session storage and context control. The new version defaults to strict mode execution, requiring updates to existing environments; bundled model default inference strength is increased to medium.
Sony globally launched the 'dual-path visual information fusion in-vehicle passenger dizziness detection' technology at the CIIE, which predicts passenger dizziness levels by analyzing the front road view and rear seat screen content. The evaluation results can be used to remind drivers to reduce hard braking or turning, or to recommend visually less changing content for content platforms. This solution reuses the ADAS front view camera and does not require a new rear seat camera.
MarkTechPost6565Use llama.cpp to run the NVIDIA Dory 1.0 4B model, which achieves inference enhancement through repeated intermediate layer calculations. Each additional loop increases computational load without increasing parameters. Build the specified branch of llama.cpp, download the 3.7 GB GGUF file, analyze the parameters and cache consumption per loop, test decoding speed and cache impact, and evaluate accuracy, answer length, and inference time changes with loop counts on GSM8K.
The Decoder6565Andy Hall researched the performance of GPT-5, GPT-5 nano, GPT-6 Astra, Opus 4.8, and Gemini 3.8 Flash in detecting improper agent behavior (such as reward attacks and sandbagging strategies).
Tencent Hunyuan · Hugging Face new models2222Qwen3.8-27B-DSpine-b16 pre-trained model released by Mixue, featuring 27B parameters, b16 precision, 5 layers of Draft layers, and an 8,192 token sliding window, designed to improve efficiency during inference and must be used with the Qwen3.8-27B model.
MarkTechPost6060Buoyant has released Kinference, an open-source Kubernetes and vLLM inference control plane, with the current version being v0.1.0-alpha.7. It distributes requests in front of vLLM worker nodes through routing policies, token rate limiting, KV cache-aware load balancing, and tenant-specific metrics. In benchmark tests, for shared prefix workloads, the median TTFT decreased from 6.88 seconds to 0.31 seconds (95%).
Hugging Face Daily Papers6565Drex open-source decision model returns the probability distribution of user options without autoregressive decoding, using supervised fine-tuning, RLAF, and consistency loss to improve accuracy, robustness, and latency. Its 35B parameter Max version achieves an industry-leading score of 67.52 on Decision Index 0.4.1, with a response time only 9% of GPT-6 Sol and 5% of Claude Opus 5.5.
Anthropic's yet-to-be-released AI model helps the physical research institution PSI prove the Allen-Arnold conjecture, which has not been resolved since it was proposed in 2008 and involves when quantum computers can surpass the computational speed of traditional computers. The research system developed by PSI based on this model completed the proof in just three days, confirming that quantum advantage depends on special laws of the problem. Previously, OpenAI has used AI to solve the Navier-Stokes equation problem.
vLLM Blog3838vLLM releases LibTorch ABI Stable CUDA since v0.33.0, allowing PyTorch upgrade without rebuilding.
The Jev and Luna models perform decently in identity resolution result scoring (88.8–97.3% accuracy), but the Jev+Opus hybrid model achieves state-of-the-art accuracy while reducing cost by 4x. Luna initially performs better than Jev, but the gap disappears in complex cases; Jev has better confidence calibration, with its errors often accompanied by low confidence, while Luna's errors are often accompanied by high confidence.
Baseten Labs6565Research indicates that on-policy self-distillation (OPSD) may still be valuable in specific scenarios, but more research resources should not be invested. OPSD trains through the outputs generated by the student model itself with context prompts, but research shows that this intensive supervision may introduce bias, especially for trained models, making it less effective.
Anthropic releases research stating that Claude AI agents can improve long-term task performance through a 'dreaming' mechanism, reflecting on historical records, correcting errors, and building knowledge structures during idle time. Claude Opus 5.5 and Sonnet 5.5, after using memory in CL-Bench database question-answering tasks, achieve average scores of 0.838 and 0.833 respectively, outperforming the memoryless baseline of 0.655 and 0.644.
Databricks Blog2222Databricks exposes shared Delta tables as real-time data sources through OpenSharing, allowing recipients to read incremental changes using Structured Streaming without local file tracking.
NVIDIA Technical Blog6565FLUX 3 Image Klein achieves a 3.5x speedup on NVIDIA Blackwell GPU. The model parameter size is 175B, scoring 95.3 on MMLU evaluation, supporting 8k context length, and providing closed-source services via API.
The DecoderSame story · 4 outlets6565NVIDIA Technical Blog3535NVIDIA has launched the Reshard Collective technology, which enhances the speed of reinforcement learning strategy model weight reshuffling by 3.5 times through cross-GPU resource coordinated refreshing. This solution supports parallel running of inference and training in distributed GPU architectures, achieving training efficiency optimization. Reshard Collective can be used in the post-training stage of reinforcement learning, reducing latency through dynamic weight synchronization.
MarkTechPost3838Zyphra has released ZYFA (Zyphra FlashAttention on AMD), a set of FP8 and BF16 pre-filled attention kernels optimized for the AMD Instinct MI355X GPU. On real model activations, its FP8 kernel is 1.16-1.76x faster than AMD's default AITER kernel library, and the FP8 error is 62-78% lower than AITER under 128-dimensional heads.
AWS Machine Learning Blog4242AWS launches serverless Model Context Protocol (MCP) server on Amazon ECS Fargate to address integration challenges of LLM agents accessing corporate data lakes in AI applications. The server achieves metadata-driven, configurable data lake access without the need to write integration code for each dataset, and supports elastic scaling in multi-tenant environments.
AWS Machine Learning BlogAWS Quick breaks down agent work into four primitives: 'When', 'What', 'Who', 'How', triggering tasks, agents execute and use capabilities (tools, skills, workflows) to complete work. Triggers are of four types: human-initiated, agent-delegated, scheduled, and event-driven; agents are persistent learners, including instructions, dynamically updated capabilities, and accumulated contextual memory; capabilities range from fully autonomous to fully deterministic, tools are atomic operations, skills provide structured guidance, and workflows are deterministic coordination.
AWS Machine Learning Blog5858Amazon SageMaker Inference now supports GPU partitioning, dividing physical GPUs into independent partitions via NVIDIA Multi-Instance GPUs (MIG), where each partition receives a fixed share of GPU computing units and dedicated memory blocks for isolation. Deploy more models in isolation and choose the appropriate size of GPU units on demand to reduce costs, with a maximum of 7 partitions per GPU, no need to build Kubernetes yourself.
Databricks Blog3838Databricks Feature Views achieve real-time fraud detection model training and online inference through a unified definition, preventing data leakage and training-service bias; support real-time feature calculation for Kafka streams, combining features like five-minute transaction count and thirty-day average spending, to implement risk scoring and real-time transaction monitoring in fraud detection applications.
Mastra Blog3838Mastra agents can use Ring signals to receive doorbell events, with each event arriving at the agent thread as a notification signal with configurable priority. Agents can retrieve snapshots or video clips from the Ring API and decide on actions when someone presses the doorbell, such as security agents detecting human motion, viewing snapshots, and marking visits requiring urgent attention, or delivery assistants notifying package delivery.
OpenClaw Blog5858OpenClaw has optimized its model integration mechanism, allowing new models to be accessed without updates. It dynamically retrieves model information through three methods: the self-updated model catalog catalog.openclaw.ai, the user account model list, and the local Codex version, solving the pain point in the old mechanism where models like GPT-6 Sol could not be selected due to version mismatches. The new mechanism reflects model availability in real-time and supports selecting models from the user account list and locally installed Codex, enhancing model availability and user experience.
Claude · YouTube4242Salesforce expanded Claude Code to thousands of engineers, doubling effective output and reducing the number of bugs per line of code. The 'effective output' scoring system, co-built by the company's engineering leadership and Stanford researchers, measures AI coding ROI, and the intelligent agent maturity curve guides engineers from simple prompts to fully autonomous agents.
O'Reilly RadarSame story · 5 outlets2222Google has released the Gemini 4 Carbon for cybersecurity and multi-step tasks, French Mistral has announced the weights of its 100 billion parameter Large 4 model, and DeepMind has previewed the Genie 4 world model for simulated environments.
Claude Code releases2222Claude Code v2.1.319 released, automatic mode is changed when telemetry or feature flags are disabled, server-side checks are disabled; server-side checks enabled after LLM gateway need to set CLAUDE_CODE_AUTO_MODE_SERVER=1.
AWS Machine Learning Blog5252Run Claude Code as a remote proxy on Amazon Bedrock AgentCore, with each task executing in an isolated AgentCore runtime within an AWS account.
AWS Machine Learning Blog3838Taxbit Assist embedded in the Taxbit dashboard, running via Amazon Bedrock AgentCore, requiring user authentication. Intelligently searches for candidate paragraphs in the Amazon Bedrock knowledge base and rearranges them on the server side. Answers must be source attributed and role permission compliant, without providing tax advice. AgentCore manages conversation state in memory, supports regional data residency, and complies with ISO/IEC 42001 standards.
AWS Machine Learning Blog3838Greenlight Financial autonomously processes over 90% of 4,000 security risk cases each month. Its solution, Luna, runs on Amazon Bedrock AgentCore, which includes over 30 specialized agents. Through AgentCore's runtime and memory support for hierarchical multi-agent systems, it achieves second-level security event triage in seconds without human intervention.
AWS Machine Learning Blog4848Okta Cross App Access manages access permissions for AI agents within identity providers, while Amazon Bedrock AgentCore verifies and transforms authorization decisions at the execution level.
