AMD has released the first Vulkan-supported FSR 'Redstone' version FSR SDK 2.4, bringing supersampling and frame generation features to the graphics API that are the same as DirectX 12.
ByteDance Seed · Hugging Face new modelsByteDance-Seed/GenFirst trains a latent VAE and combines it with a generative prior to avoid latent collapse, which usually makes end-to-end training unstable. The two-stage plan of this method is used for prior and entropy weights: the first stage lets the prior shape the latent geometry, and the second stage reduces both weights tenfold, allowing the decoder to recover details.
Hugging Face Daily Papers6565DTCA performs label-level credit assignment through confidence enhancement and entropy calibration, improving the mathematical reasoning ability of diffusion language models. The framework, when tested on LLaDA-8B-Instruct, shows significant performance improvement in mathematical reasoning under different generation budgets, with the MATH500 metric increasing by 9.0 points and the GSM8K metric improving by 5.3 points.
Tencent announces the upcoming release of the Hunyuan Hy4, which offers a comprehensive upgrade compared to the Hy4 preview, introducing visual understanding capabilities for the first time, and enhancing long-range and productivity tasks; the Hy4 flash AI model will also be launched, providing excellent cost performance, inheriting the visual and Agent capabilities of Hy4.
OpenRouter BlogOpenRouter has released a new supplier routing algorithm that selects the best cost-performance solution for each request by evaluating price, latency, throughput, and error rate in real time, significantly improving reasoning speed (TTFT improved by 16.9%), throughput (improved by 16.3%), and reliability (error rate reduced by 67%), with only a slight increase in cost of 6.0%. This algorithm rewards suppliers that provide high-quality reasoning without requiring users to modify their code, ensuring a continuously optimized experience.
OpenRouter BlogThe Anthropic/claude-haiku-5.5 model paired with a strict JSON schema can extract tables from PDFs and return typed rows. Tested five PDFs (including scans) successfully extracted all 644 cells; failure cases were due to PDF processing paths, not model capability. Note empty results and output length limits.
The Nemotron 3.5 Transcribe open-weight speech model now supports RTX Spark, enabling speech transcription, speaker recognition, and translation, supporting 40 languages. The Liquid AI LFM3 series models have also been supported and can be run locally and fine-tuned. The RTX Spark laptop starts at $2599, with the highest configuration of 128GB unified memory, and FP4 AI computing power up to 1 petaflop.
Google Developers BlogLiteRT has received a major upgrade in Google Play services, with GPU inference speed about 1.4 times faster than traditional GPU accelerators, achieved through the newly introduced CompiledModel API, driven by the open-source GPU engine ML Drift
Latent SpaceAfter becoming the preferred backend for many cutting-edge AI models, Supabase reduces development friction through practical features like starting a local backend without Docker, managing database schemas with Git, and running computations alongside the database.
Replit BlogViBench 2.0 Released, used to evaluate the ability of models in long-term software construction, including 17 products, 241 feature requests, and 544 test plans, top model GPT-6.1 Sol.
AWS Machine Learning BlogQA Studio organizes tests and reports health metrics, allows team-level user management to control access, monitors test coverage through health scores and trend charts, and implements application-level permission management with five roles (admin, team lead, editor, executor, viewer).
Apple released its first touchscreen MacBook Pro equipped with a PROLED XDR display, featuring the M5 Max chip version with up to 128GB unified memory, capable of locally running large language models with billions of parameters, supporting applications like LM Studio, starting at 22,999 yuan.
MarkTechPostMiniMax Open-Sources MCode Stack, including 17 AI agent skills and 5 lists used by its coding team to build MiniMax Code, under MIT license, non-prompt set, each skill is a SKILL.md file that can be directly installed into MiniMax Code
Google Developers Blog5858Agents CLI v1.8.0 achieves Hill-Climbing optimization of Agent quality through commands like eval generate, eval grade, etc., using the custom_response_quality metric to rate 4 cases from 1 to 5, finding that changing the system prompt to a chef role reduced the average score from 4.0 to 2.75. This tool supports model backend experiments and automatic alerts, improving Agent performance through iterative cycles.
Artificial Analysis ArticlesFeaturedSame story · 5 outlets8585Google's Gemini 4 Carbon ranked first with 59 points on the Artificial Analysis Intelligence Index v5, leading in four out of nine evaluation items: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, and AutomationBench-AA
Baseten BlogBaseten achieved the lowest cost per minute for batch speech transcription by optimizing Qwen3-ASR 1.7B (175B parameters, 5.4% WER) on Baseten. Its vLLM stack supports 7 times the throughput of standard vLLM on RTX PRO 6000, with a maximum real-time factor of 3,128, enhanced by native audio decoding, FP8 KV cache quantization, and reduced host-device synchronization.
LangChain BlogOpenWiki makes agents about 5 times more accurate on cross-source connection issues than single search, adding evidence images increases accuracy by another 27% and doubles citation consistency.
PyTorch Blog5858PyTorch® 2.15 released, defaulting to the rewritten NCCL backend, supporting fault logging and rollback; torch.distributed adds NIXL backend for unilateral transmission; FSDP2 and pipeline parallel optimize memory communication control.
Hugging Face Daily PapersThe EvidenceConnect framework generates comparative, aggregated, or chained visual evidence queries by synthesizing image region thumbnails, enhancing the evidence connection capability of VLMs. The ZwZ2 model is trained based on this framework, without targeting any downstream benchmarks but improving fine-grained and spatial perception, and achieving progress in embodied spatial reasoning, multimodal long-context understanding, multimodal agent tasks, and general visual reasoning
Meta EngineeringMeta Andromeda's Next-Generation Ad Retrieval Engine Achieves Near-Ranking-Level Model Performance at the Retrieval Layer, Processing Over 1 Billion Ads Per Card Per Second, System Throughput Reaches 100 Billion Times Per Second, 100 Times Higher Than the 2024 Version, and Brings a 2.4% Conversion Rate Increase for Advertisers.
OpenRouter Blog5858When you reach the Claude Code usage limit, you can continue coding using an OpenRouter API key, which adopts a token-based pricing model and has no five-hour window limit. Setting three environment variables can switch you back to the original plan. Typical short conversations on Claude Sonnet 5.5 cost 0.16 USD, and on Opus 5.5, 0.31 USD.
Hugging Face Daily Papers6565Multiscale Policy Optimization (MPO) improves the training of long-horizon LLM agents by aligning credit assignment and policy updates at the episode scale. Experiments on search, tool use, and automated tasks using 4B and 30B-A3B models show that MPO has sustained improvements over strong RL baselines.
MooreThread open-sources Megatron-MUSA-Titan project, supporting trillion-parameter MoE model training for multi-thousand node clusters, covering pre-training and reinforcement learning scenarios, achieving high MFU and long-stable training. The project supports blockwise FP8 training, combining DeepEP and ACE asynchronous communication engines to accelerate token distribution and result merging, improving end-to-end training efficiency.
The all-new Jietu Freedom 7 will go on sale on November 4, equipped with Lingxi Smart Cabin TOUR OS, featuring a dual-supported AI voice large model with DeepSeek+ and iFlytek Spark, a 2.5K ultra-clear screen, and a Qualcomm Snapdragon 8155 chip. Additionally, the car supports APA all-scenario automatic parking.
Liquid AI releases the LFM3 model family, with the first batch of multiple models based on LFM3-1.7B-Base, among which LFM-1.7B has a 256K context window, outperforming the previous generation LFM2.5-2.6B while reducing parameters by 35%
Survival LLM is a free AI assistant and emergency knowledge base that runs entirely on a personal computer, providing answers based on trusted sources without the internet and displaying the answer sources.
MarkTechPost6060FlashMap replaces the discrete diffusion drawer with a continuous flow drawer, achieving up to 6.14 times lossless speed improvement on Qwen3-8B for GSM8K. Its drawer weights (Qwen3-8B and Qwen3-4B) have been open-sourced to Hugging Face, supporting vLLM and SGLang plugins.
WuWen CoreVault ranks first among all participating teams in the NVIDIA Research GPU operator performance evaluation list SOL-ExecBench with 88 single-item first places, covering four major tracks: L1 basic operators, L2 fusion operators, Quantization operators, and FlashInfer-Bench inference operators. Through containerized isolation architecture, modular Skills, and shared memory mechanisms, combined with refined optimization strategies for different computing scenarios, it achieves the ability of AI to autonomously complete complex operator optimization.
Developers found three model identifiers—deepseek-v5, deepseek-v5-open, and deepseek-v5-next—in the DeepSeek Harness GitHub repository, currently serving as framework placeholders.
Hugging Face Daily Papers6565Large-scale evaluations show that dedicated memory systems are often weaker than simple retrieval-enhanced generation. The Memory Benefit and Memory Tax framework is introduced to evaluate the net value. Memory is becoming increasingly common in long-term intelligent agents, but the net value of building memory compared to directly retrieving from raw history remains unclear.
Hugging Face Daily PapersPDRVLA improves robot action generation by integrating a lightweight physical dynamics module that predicts future visual features under the supervision of a physical world model, avoiding latency caused by online video generation. PDRVLA achieves success rates of 99.3% (LIBERO), 89.9% (LIBERO-Plus), 93.6% (RoboTwin~2.0), and 81.5% (SimplerEnv) on four simulation benchmarks.
Hugging Face Daily Papers6565Looped Transformers train better than truncated backpropagation with the same computational cost through full backtracking, extended regularity shows that when the model size is adjustable, two loops are the best. Looped Transformers increase computational depth by reusing shared layers without increasing the number of trainable parameters, usually using backtracking training (BPTT), but are often truncated to the last few loops to save training computational cost, which may lead to gradient bias and weaken the loop advantage.
Ant Group's Baoling large model has open-sourced Ling-3.1-flash-1019, scoring 1524 Elo in the intelligence index AA-Briefcase, an increase of 3 points compared to the previous version; the basic model Ling-3.1-flash-base adopts a hybrid linear MoE architecture.
Hugging Face Daily PapersHarness Self-Evolution improves LLM agents by guiding skill adaptation without parameter updates, enhancing the performance of weak models. The study found that skills are beneficial for strong models but potentially harmful for weak models. In seven models and three benchmark tests, 103 task-skill pairs improved GPT-5.6-luna by 31.1 points but reduced Gemma-4-E4B by 8.8 points.
Hugging Face Daily Papers6565The Test-Time Memory Evolution (TTME) method guides scientific discovery more efficiently than Test-Time Training (TTT) by updating external storage without changing model weights. On five mathematical tasks, TTME outperformed TTT while reducing FLOPs computational cost by 82.8-84.6% and runtime by 47.7-58.4%.
Browser UseGPT-6 Astra performs best in the Odysseys2 long-term web intelligence agent benchmark, completing 47 tasks with an average score of 84.6. The benchmark includes 100 real web tasks, each with 8-12 pass/fail checks, 15 agents participated in the test, among which Opus 5.5 efficient mode completed 18 tasks with an average score of 71.2.
Hugging Face Daily Papers6565The ThinkWorld model enhances robot generalization by representing actions as dynamic chains and learning from unlabeled human videos using world imagination. This VLA framework represents each action reasoning as a dynamic reasoning chain (CoD), combining reinforcement learning with world imagination to reinforce reasoning, allowing unlabeled videos to provide direct supervision for state transitions, achieving a shared reasoning space for human and robot experiences. Experiments prove its effective implementation of human-robot transfer and generalization to unseen tasks and environments on both simulation and real robots.
Hugging Face Daily PapersOvisOCR3 is a 0.8B visual language model that unifies multiple OCR tasks through multi-mentor distillation and reinforcement learning, achieving SOTA results on document parsing, general OCR, key information extraction, image-text description, and cross-page table reconstruction tasks. It scored 96.74 on OmniDocBench v1.6, ranked first in OCRBench v2 English and Chinese split, and surpassed baseline models in multiple benchmark tests.
Hugging Face Daily PapersChronoReplay training framework organizes heterogeneous spatiotemporal supervision into time-ordered interaction segments, preserves specific task answer formats, and balances learning from different supervision signals, enabling a single model to learn complementary spatiotemporal capabilities while maintaining temporal consistency. It achieves improvements on OVO-S-Bench, OST-Bench, and UCS-Bench benchmarks, with a 13.41-point increase on OST-Bench, while optimizing streaming inference performance through adaptive frame selection.
