Hugging Face Daily Papers6565LongBanana is the first expert-designed and validated long-context multi-reference image synthesis benchmark, containing 90 carefully designed tasks, each task containing about 32 reference images, 20 constraints, and 50 evaluation items, a total of 4,670 expert-specified atomic evaluation items.
MiniMax · Hugging Face new modelsMiniMax-H3-Image-PE uses prompt expansion (PE) to convert short requests into detailed image descriptions and recommends aspect ratios. It supports two workflows: text-to-image (T2I) and image editing, using fine-tuned FP4 weights from Qwen3.8-27B. The online API can be accessed via platform.minimax.io or platform.minimax.cn, and the documentation provides support in both Chinese and English.
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.
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).
The Douyin App has integrated Apple's Smart function, allowing users to quickly retrieve information by taking photos of products with their iPhone camera. After taking a photo, users can search for related products on Douyin to view details, historical prices, AI interpretation, and user reviews.
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.
Hugging Face Daily Papers6565VisSearch-Bench evaluated the performance of 16 multimodal models on 204 human-written visual search tasks that require integrating visual context with external information, with the results showing that the best system only fully solved 26.47% of the tasks. The benchmark covers 7 major areas, including set-based visual search and sequential visual search, and evaluates task completion and the correctness of reasoning steps through accuracy and process scoring.
Apple introduces SynthID invisible watermark for some Apple Intelligence image tools (iPhoto, Image Magic Wand, and Clean Up, Extend, and Space Reconstruction in Photos) in iOS/iPadOS/macOS 27.2. It is invisible to the human eye but can be read by the SynthID detector. Users can check if an image has a watermark at synthid.com; this technology supports pixel-level annotation of real image data to prove photo authenticity.
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.
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.
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.
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.
Hugging Face Blog5858Hugging Face released ToyBlocks, a framework for robot learning and evaluation that provides a controlled simulation environment. The framework supports automatically generating tasks and training demonstrations, and includes metrics to measure changes in robot behavior over time, aiming to study how models improve with more training data and how to combine advanced reasoning with precise physical control.
Artificial Analysis Articles6565VIALS-AA tested 19 models and found that 38.3% of tool calls were used to view images, 47.3% for image preparation or measurement, while the image measurement call rates for GPT-6 Astra and GPT-6.1 Sol were only 5-6%, and Claude Opus 5.5 and Claude Sonnet 5.5 were 25-27%.
Hugging Face Daily Papers6565SLA model aligns pre-trained VLA strategy with frozen 3D perception model spatially through continuous pre-training and post-training, using feature-dependent gating and probabilistic state dropout to enhance spatial understanding across different robot datasets while retaining the inference architecture.
The Pangram Image 1.0 model locates AI-generated content using heatmaps, achieving an accuracy of 99.96% on a test dataset of 1200 photos. This model can resist the effects of compression, scaling, and artificial modifications, with an accuracy of 96.4% still maintained at 256×256 resolution. The simultaneously released Pangram Image Flash 1.0 video detection model achieves an accuracy of 98.7% on the DF26 dataset, leading the second place by 29 percentage points.
Developer releases Mac app Window Sweaters, adds knitted borders to 37 apps, color inspiration from app icons, supports custom styles and modes, installable via Homebrew or manual installation, requires macOS 13 or later.
QQ Harmony OS version App 9.2.80 (3592) official update released, Space Album adds AI repair function, adds support for simple mode
Huawei has rolled out the HarmonyOS 7.0.0.111 SP15 version upgrade for the Mate 90 series phones today, optimizing the shooting effect when paired with the Leica Z10 module camera, while also improving the overall system stability. The Leica Z10 module camera, code-named 'Vajra', integrates a Redwood sensor, supports soft and hard core AI collaborative focusing, and has an independent optical image stabilization module.
Hugging Face Daily PapersWeLM-Image is a unified image base model that integrates multimodal understanding, supporting text-to-image generation, instruction-guided editing, and multi-reference synthesis.
Hugging Face Daily Papers6565TailSteer, by maintaining early denoising dynamics to approach the base model and learning fine deviations in trajectory tails from within the policy through rolling out, significantly reduces accumulated visual drift in multi-round editing while preserving pre-trained generation and editing capabilities. Experiments show that TailSteer can greatly reduce drift on FLUX, SenseNova, and Qwen series models in 15-20 rounds of editing and largely maintains the performance of the base model on large text-to-image and single-image editing benchmarks.
AMD Next-Gen Mobile APU Medusa Point Scores 1560 Single-Core, 11575 Multi-Core in Lenovo Test System with Geekbench 7, Equipped with 10 Cores 20 Threads (4 Zen 6 Performance Cores + 6 Zen 6c Efficiency Cores), 16MB L3 Cache, Supports Instruction Sets like AVX-512
TechCrunch · AISame story · 2 outlets6565Meta announces that its consumer AI agent Muse's voice and video chat features Muse Realtime Avatar and Muse Realtime Voice are now available to all users in the United States. Users can engage in real-time conversations by interacting with the cute animated avatar Jolly, which can be customized according to user preferences and uses subtle expressions and gestures that match the mood of the conversation.
How the accounts differ:The Verge mentions the voice mode is available in iOS and will soon apply to Meta's AI glasses and upcoming Muse Charm device; TechCrunch does not mention these details.
The Verge · AISame story · 2 outlets6565Meta allows users to have real-time voice conversations with the Muse AI agent, CEO Mark Zuckerberg announces the feature has been rolled out to some users. In the iOS app, the Muse agent's avatar animates in real-time based on the conversation content. Users can adjust the voice tone by asking the agent or browsing the avatar settings options. This voice mode has now launched in the US region and will soon be applied to Meta's AI glasses and the upcoming Muse Charm device.
How the accounts differ:The Verge mentions the voice mode is available in iOS and will soon apply to Meta's AI glasses and upcoming Muse Charm device; TechCrunch does not mention these details.
SemiAnalysisGoogle TPUv7 Ironwood outperforms B200 and B300 by 50% per dollar in single inference performance on Qwen 3.5 397B FP8, and shows competitive or best results in the Multi-Model AgentX benchmark, supporting ecosystems like vLLM and SGLang, but NVIDIA reacts fiercely and accelerates R&D.
Microsoft ResearchThe PDD Kit translates user intent into formal specifications and verifies whether the code written by the agent for each input conforms to these specifications, supplemented by testing. The team officially validated the Rust-written Pet Store login logic using the Chopin environment on GitHub Next with the PDD Kit, and the Verus tool discovered and fixed existing bugs, proving no login bugs under the specification.
Hugging Face Daily Papers6565EmoReward adapts multi-reward weight allocation by extracting evidence reliability from reward trajectories collected during actual training runs to calibrate reward weights for emotional video generation. This mechanism first evaluates specific grounding cues based on emotional prompts, then scores videos during the generation phase using heterogeneous signal primitives, and calibrates their contributions through reflective reward gating.
Hugging Face Daily Papers6565NEO-VLA framework eliminates fixed visual encoders by allowing action strategies to directly access deep visual representations in the network, achieving stronger robustness in various manipulation scenarios and distribution shifts. The framework uses lightweight convolutional layers to tokenize images and allows visual representations to evolve together with language in the backbone network, enabling action experts to obtain information from different depths of instruction-conditioned visual representations. Layer exploration experiments show that action key signals often reach their peak before the final visual representation is formed.
MarkTechPostFeaturedSame story · 3 outlets7070OpenAI launches Images 2.6 on ChatGPT, Codex, and API, offering GPT-Image-2.6 Flare and Sunburst models, with the official claim that generation speed is 50% faster than the previous version
MarkTechPostSame story · 2 outletsGoogle AI Research releases Agile Deliberation framework, which transforms fuzzy concepts into structured definitions through human-in-the-loop, for gray-area image classification. The framework is based on Gemini 2.5 Pro and Gemini 2.5 Flash API, requiring no fine-tuning, with an average F1 score 10.5 points higher than zero-shot Gemini 2.5 Flash, reaching 0.59 in paid scenarios.
How the accounts differ:Google Research mentions user experiments reducing fatigue and anxiety; MarkTechPost focuses on F1 score improvements.
OrangePi released the OrangePi Zero 3W Pro, a small-sized SBC based on the Rockchip RK3576 processor, supporting 16GB LPDDR memory and various storage extensions, integrated Wi-Fi 6E & Bluetooth 5.3, suitable for local lightweight AI inference.
MIT News · AIThe UQ-DAAAM system can identify and correct errors in robotic memory descriptions, quantifying uncertainty by comparing multi-view images and optimizing object descriptions, thereby improving the accuracy of robots' perception of objects, locations, and time. The current correction success rate is approximately 92%, with an accuracy improvement of 5%. The system employs LLM to integrate multi-view descriptions and uses heuristic rules to select the best supplementary images, significantly reducing computational load without rerunning VLM.
FurMark 2 version 2.11.0.4 update, supports testing of new GPU devices such as RTX Spark, optimizes Windows GPU core hotspots and memory temperature display, and Linux Wayland synergy.
OpenAI's GPT-Image-2.6 has been fully launched on ChatGPT, Codex, and API, available to both free and paid users, with overall generation time reduced by 50%, producing images in as fast as 6 seconds.
OpenAI NewsFeaturedSame story · 3 outlets6262OpenAI has released Images 2.6, bringing faster image generation, GIF creation, more precise editing, and new API controls. This version covers three access points: ChatGPT, Codex, and API.
Why it matters: The original text lists four changes: image generation speed improvement, GIF creation, more precise editing, and API control, which can be used to judge the boundary of image capabilities for ChatGPT, Codex, and API.
