Skip to content
Only Models
Oct 19Mon
  1. 智源研究院 · 微信公众号68

    BAAI's FlagOS 2.2 released: over 1,454 operators, 27% cross-chip performance gain

    BAAI presented FlagOS 2.2 progress at the 2026 AI Open Computing Conference and FlagOS Technical Conference; the version was released on September 30. Public operators in the nine operator libraries rose from 564 to 1,454 (890 added). In tests covering over a thousand operators and 10 heterogeneous AI chips, cross-chip operator performance improved by about 27% over the previous version. The FlagOS-Compressor weight conversion and quantization tool was released for the first time, enabling a new model to be adapted to as many as 12 chips within one day.

    Why it matters: The original text provides specific figures for operators, quantization, and serving throughput, which readers can use to assess the performance and cost potential of existing chips for running new models.

  2. 36Kr Embodied Intelligence42

    MiniMax M3.1 Flash launches with across-the-board performance gains, 3x faster speed, and 70% lower price

    MiniMax has officially released its new-generation Flash model, M3.1 Flash, with online inference speed nearly 3 times faster and usage costs reduced by 70%. The Space Bunny model, which previously handled over 200T tokens anonymously, was in fact an early version of M3.1 Flash. The model delivers generational improvements in agentic tasks, coding, Computer Use, and cybersecurity capabilities. MiniMax has teased that a Pro version pursuing a higher intelligence ceiling will be released soon.

  3. 深蓝AI · 微信公众号62

    Daxiao Robotics releases ACE-Brain-0.6: 9B model beats GPT-6 Astra on six Grounding benchmarks

    Daxiao Robotics released the embodied spatial perception model ACE-Brain-0.6, built on Qwen3.5-9B with a frozen vision encoder and supervised fine-tuning of the language model parameters. It surpasses GPT-6 Astra on six of eight Grounding benchmarks, leads with an average of 74.53 points across seven public benchmarks versus Astra's 73.68, and scores 94.13 on RobotGrounding versus 91.10.

  4. 36Kr Embodied Intelligence36

    MetaDecision enters JevBench open-source top three: 31B version has median single-decision latency under 0.2 seconds

    MetaDecision, the structured decision model from MetaDecision Lab, has entered the top three on the JevBench decision model evaluation leaderboard: as of October 17, among 129 open-source systems, its 31B version ranks 3rd, and its 12B version ranks first among models of the same size. According to the leaderboard's GPU measurements, the 31B version has a median single-decision latency of under 0.2 seconds (excluding network). MetaDecision is an AI lab focused on Agent Infra and decision models, offering an embeddable Agent runtime foundation as well as knowledge and model services.

  5. arXiv cs.RO64

    GroundingPI releases 4B grounding foundation model, averaging 73.68% across 34 benchmarks

    GroundingPI is a 4B-parameter grounding foundation model that outputs points and boxes as quantized coordinates in a shared vocabulary; its training combines multimodal and spatial pretraining, supervised fine-tuning, and GRPO reinforcement learning. It averages 73.68% across 34 grounding benchmarks and 44 baselines, above GPT-6 Astra's 71.54%.

  6. 仙工智能58

    Seerbotics demonstrates VLA model completing Tower of Hanoi task: still able to continue solving after base is pushed and discs are scrambled

    Seerbotics demonstrated its self-developed VLA model driving a robotic arm to complete the seven-move sequence of a three-disc Tower of Hanoi. When an observer slides the base to change its overall position or rearranges the discs into an arbitrary intermediate state, the robotic arm can still re-identify the current configuration, determine the next move, and continue completing the remaining task. Trained on only 60 demonstration trajectories, the model can generalize across changes in object position, color, and shape. The Seerbotics robot "brain" consists of three layers—model, cerebellum, and nerve—responsible respectively for scene understanding and decision-making, motion planning and execution, and a unified interface connecting sensors and joint actuators.

  7. HuggingFace Daily Papers62

    ACE-Brain-0.6 released: scaling embodied foundation models with verifiable spatial grounding supervision

    ACE-Brain-0.6 combines robot manipulation recordings with complementary visual data from multiple sources through a unified data pipeline; after automatic annotation, the data is converted into verifiable grounding supervision, and Qwen3.5-9B is fine-tuned on about 1.5 million samples to produce the model. It outperforms GPT-6 Astra on four of seven public grounding benchmarks, with an overall average score of 75.41 versus Astra's 73.68. Under a fixed fine-tuning setup, aggregate grounding performance improves steadily as the dataset expands to cover more grounding scenarios.

  8. 机器人全球资讯 · 微信公众号62

    Westlake Robotics WR1 demonstrates full-body coordinated autonomous clothes folding and cross-area household chores

    Westlake Robotics has upgraded its general-purpose brain WR1, adopting a dual pretraining architecture of a general-purpose large brain and cerebellum, integrating a world model with VLA to generate action sequences, and executing full-body control through its self-developed general action model GAE. In the demonstration, WR1 completed a continuous task sequence: taking clothes from a drying rack, folding pants with two machines working together, and across the kitchen, living room, and bedroom, fetching corn, storing dolls, and tidying a quilt. The system introduces long- and short-term motion state review and a real-time action chunking mechanism to reduce pauses and abrupt changes in long-horizon tasks.

Oct 18Sun
  1. 量子位74

    VAST open-sources real-time world model EdenLive, over 24 FPS at 720P on a single RTX 5090

    VAST has released and open-sourced its real-time world model EdenLive, opening up the model weights, inference code, technical report, and an end-to-end demo. According to VAST, on a single consumer-grade Nvidia RTX 5090 GPU at 720P, the 5B model generates at over 24 FPS, can generate continuously for minutes at a time, and has end-to-end latency in the hundreds of milliseconds.

  2. 36Kr Embodied Intelligence47

    StartLux-Decision-v1.1-27B Instant Tops Hugging Face Decision Index 0.3.2

    StartLux-Decision-v1.1-27B Instant, from Chinese AI startup StartLux (Yuandian Xinghui), ranked first worldwide on the Hugging Face Decision Index 0.3.2 with a score of 69.1, placing it at the top among 173 publicly reproducible models. It leads the second-place model by nearly 3 points and is the only model on the list to exceed 69 points. After releasing its first-generation model in late September, StartLux set a new best result in just over ten days, maintaining its position as the global SOTA in decision models.

  3. Guyuehome46

    Light-MER: Compressing an 8B Emotion Teacher into an 855M Student Using Hidden-State Geometry Rather Than Answers

    Light-MER is a multimodal emotion recognition compression work by the University of Glasgow, the Institute of Computing Technology of the Chinese Academy of Sciences, and Shandong University. Accepted to the ACM MM 2026 main conference, it compresses an 8B teacher model into an 855M student model, with the training signal drawn from hidden-state geometry rather than answers. The work targets the deployment cost of generative multimodal emotion recognition, with the goal of enabling the model to output paragraphs that explain emotions.

  4. Guyuehome58

    τ₀-VLA breakdown: Turning subtask decisions into a scalable reasoning problem

    The article breaks down τ₀-VLA, a hierarchical robot foundation model open-sourced on July 27, 2026 by Luo Jianlan's team at the Shanghai Innovation Institute (SII). The model targets long-horizon mobile manipulation on the order of ten minutes in home scenarios, such as meal prep, folding and storing clothes, and tidying across rooms. Its core idea is to turn "what should the next subtask be" into a reasoning problem that can be scaled with more compute.

  5. Guyuehome51

    TD-MPC2: Model-Predictive Reinforcement Learning Control Practice with Latent World Models

    The author uses official pretrained weights to drive a quadruped canine model through real running inference on a local CPU, reconstructing item by item the mathematical construction of the latent world model, the consistency learning objective, symlog two-hot encoding, and MPPI-style sampling-based planning, and records observed phenomena such as planning distribution contraction and the gap between estimates and realized outcomes.

Oct 17Sat
  2. 机器人产业应用 · 微信公众号74

    How far has Generalist AI GEN-1.5's physical prompting paradigm progressed?

    Generalist AI released GEN-1.5 on August 19, 2026; it can attempt new tasks from a single 3-12 second task demonstration without gradient updates, achieving a one-shot average success rate of 59%±10% across 10 short manipulation tasks, and this can be raised to 83%±9% by adding about 5 minutes of data and performing 10 gradient updates.

Oct 16Fri
  2. AI前线 · 微信公众号78

    Google releases Gemini 4 Argon physical reasoning model, connecting the full robotics R&D chain

    Google has released Gemini 4 (Argon), positioned as a general physical reasoning model focused on interaction with the physical world, covering the full robotics lifecycle from control execution and simulation reconstruction to data annotation and self-improvement. On the ERQA embodied reasoning benchmark, Argon scored 87%, surpassing Claude Opus 5.5 (82.3%) and GPT-6 Astra (81%); in the AA-Omniscience test, its hallucination rate was 15%, lower than the 51% and 59% of the latter two.

    Why it matters: The article breaks down the model's specific uses and measured figures across the robotics R&D chain in four layers: control, simulation, data quality inspection, and self-improvement.

  3. 量子位70

    Google releases Gemini 4 Argon, demo shows transfer from human video to Apptronik Apollo humanoid robot

    Google has officially opened Gemini 4 Argon to developers, enterprises, and paid subscribers worldwide; the core upgrade is stronger long-horizon task execution and multimodal capabilities. In the official demo, after watching a video of a person playing beanbag toss, Argon analyzed spatial relationships, calculated full-body joint motion, autonomously wrote simulation code, and transferred the plan to Apptronik's Apollo humanoid robot. Google says this video-to-simulation-to-reality pipeline can compress simulation modeling that previously took weeks of manual work into a few hours.

Oct 15Thu
  3. Xbotics具身智能实验室 · 微信公众号76

    Shanghai AI Lab evaluates GPT-6-Astra's capability boundaries as a robot policy on EBench

    Shanghai AI Laboratory's embodied evaluation team used EBench to evaluate GPT-6-Astra as a robot policy directly, without task-specific post-training, across 26 tasks and 510 total executions. Its average task success rate of 46.7% ranked third, and its average task score of 0.654 ranked second; OpenWAM-α led with a 55.3% success rate.

    Why it matters: The benchmark quantifies the capability gap between general large models and post-trained policies, helping readers judge which task lengths and manipulation precisions each approach suits.

  4. GitHub Robotics Topic41

    BrickSort v0.1.0 released: model recognizes about 600 LEGO parts

    The LEGO part recognition project BrickSort has released its first public version, v0.1.0, along with a trained model, lego_model.pt, that can recognize the type, color, category, and whether a part is printed for about 600 LEGO parts. The code is under the MIT license, and the model file uses the same license because its training data (the B200 LEGO Detection Dataset) is CC BY-SA 4.0. The author notes that the accuracy printed by training is optimistic, and an independent test on photos of unseen parts has not yet been conducted.

  5. 36Kr Embodied Intelligence61

    StepFun opens Step 5 Preview weights

    StepFun announced the official release of the Step 5 Preview weights, including BF16 and FP8 versions, along with a technical blog, supporting open-source development tools, and a deployment guide. The model weights are now available on Hugging Face and ModelScope.

Oct 14Wed
  4. Xbotics具身智能实验室 · 微信公众号66

    1.7B-parameter 3D world action model UNITAS open-sourced, surpassing 8B-parameter models on LIBERO and other benchmarks

    CrossDim Intelligence, together with The Chinese University of Hong Kong, Shenzhen, Foshan University, Sun Yat-sen University, and South China University of Technology, proposed and open-sourced UNITAS, a robot task-native 3D world action model. The code and usage instructions are publicly available in the GitHub repository DexForce/UNITAS.

  6. 甲子光年 · 微信公众号60

    Huanwang and Pyromind co-build AITV solution, raising focus recognition accuracy from 45% to 99.6%

    Huanwang Technology and Pyromind have co-built an AITV intelligent solution, performing scenario post-training on Qwen3-VL-4B. Focus recognition accuracy rose from 45% to 99.6%, the success rate of "see it, say it" reached 99%, and the success rate of "search and play" exceeded 90%. It has been deployed in "Xiaoyou Casting - Movie Search" and covers TV apps including Cloud Video Aurora, Kiwi, Mango, Youku, and Bilibili.

Oct 13Tue
  1. 银河通用机器人 · X42

    Galbot releases PASSAGE humanoid traversal framework

    The Galbot team proposes the PASSAGE framework, which combines environmental perception, whole-body motion planning, and closed-loop control so that humanoid robots can autonomously plan and execute agile traversal motions based on their surroundings. The framework was trained on 100 hours of human data covering 1,500 scenes and achieves a 98.67% success rate on the Clutter Scene Traversal benchmark.

  2. Xbotics具身智能实验室 · 微信公众号74

    UniWAM open-sources unified world and action model, leading multiple benchmarks with 99.2% on LIBERO and 68.32% on RoboTwin C2R

    The UniWAM team has open-sourced the code and weights of UniWAM, a unified world and action model, achieving success rates of 99.2%, 92.6%, and 68.32% on LIBERO, LIBERO-Plus, and RoboTwin 2.0 C2R, respectively, the highest values among the methods listed in the corresponding tables of the paper.

  3. 深圳市机器人协会 · 微信公众号70

    HKU MMLAB releases PhysicalRSI 1.0, tops RoboDojo and sharply cuts token costs

    The University of Hong Kong's MMLAB jointly released PhysicalRSI 1.0, a self-evolving embodied intelligence system baseline, achieving a score of 36.27 and an average task success rate of 31.38% on RoboDojo, ranking first.

    Why it matters: The material provides the specific mechanisms of the self-evolving baseline and comparison data with general large models, allowing readers to understand the causes of shortcomings in precision manipulation and paths for improvement.

  4. ZhenFund · 微信公众号63

    Natural Will releases first physical general foundation model Natura-1, 50-billion-parameter MoE drives 20-DoF dexterous hand

    Natural Will has released Natura-1, its first general foundation model for the physical world. It is a 50-billion-parameter mixture-of-experts (MoE) model pretrained on 1 million hours of first-person human experience data. Taking language and vision as input, it can drive a 20-DoF dexterous hand within a 100-millisecond control loop and follow complex natural-language instructions zero-shot.

  5. Dexmal 原力灵机 · 微信公众号56

    Force Machine DM0.5 showcases three task types at the CIIF: connector insertion quality inspection, industrial insertion, and flexible cutting

    Force Machine presented its self-developed DM0.5 general-purpose embodied brain at the AXERA booth at the 2026 China International Industry Fair, completing three task types: connector insertion quality inspection, Insert Anything general-purpose industrial insertion, and dexterous-hand flexible cutting, with model inference performed on the AXERA high-compute platform.

Oct 12Mon
  2. Shanghai AI Laboratory - Embodied Intelligence73

    Shanghai AI Lab launches decision model Intern·Mingjue, outperforming Jev and releasing 0.8B, 2B, and 4B weights

    Shanghai AI Laboratory has launched the decision model Intern·Mingjue (Decision), which outperforms Jev and is available in 0.8B, 2B, and 4B parameter versions. In RTX 4090 tests, the 0.8B and 2B models had an average end-to-end latency of about 33 milliseconds and about 30 decisions per second, while the 4B model was 44.16 milliseconds and Jev was 109.70 milliseconds.

    Why it matters: The original text provides measured latency, benchmark scores, and a mechanism for routing between large and small models, which readers can use to judge the decision model's suitability for interactive scenarios.

Oct 11Sun