FlashState is an inference system that flexibly manages the prefix cache of hybrid LLMs, and on GLM 5.3 Flash it more than doubles vLLM's throughput. It cuts latency by 3.7×, reduces user cost by 2.5×, and raises provider profit.
Key points
Throughput: more than doubles vLLM's throughput on GLM 5.3 Flash
Latency: cuts latency by 3.7×
Cost: reduces user cost by 2.5× and raises provider profit
Sources
Anchor 1 controller launches, lowest tracking errors
Anchor 1 whole-body controller for humanoid manipulation turns sparse human targets into precise, stable robot motion. It separates motion encoding from feedback correction for accurate tracking and smooth movement. On a real robot it posts the lowest tracking errors and highest task success against baselines. Weights and deployment code are expected late 2026.
Key points
Sparse human targets (pelvis, wrists, ankles) become precise, stable robot motion, skipping online joint-space retargeting.
Lowest tracking errors and highest task success against MimicLite, ScaleBFM, and SONIC baselines on a real robot.
Weights and deployment code expected late 2026.
Sources
Alibaba Qwen agent drives 140M first purchases, 4B catalog
Alibaba reports 140 million users made a first agent-driven purchase through Qwen, which now accesses over 4 billion products. The post also notes Alipay AI Pay reached 100 million users and processed more than 300 million agent transactions by late May.
Key points
140 million Alibaba users had their first agent-driven purchase via Qwen, catalog now over 4 billion products.
Alipay AI Pay passed 100 million users and completed over 300 million agent transactions by late May.
Sources
Falcon-OCR 270M ranks top small OCR model in private eval
Falcon-OCR (270M) is among the top two models in a private OCR evaluation on 20th-century documents, judged by a VLM across 10,730 pairwise comparisons. The evaluation covered 338 scanned government pages and nine open-weight OCR models.
Key points
Falcon-OCR size 270M parameters.
Evaluation used 338 scanned government pages, nine open-weight OCR models, and a VLM judge performed 10,730 pairwise transcription comparisons.
Sources
- Original post
- huggingface.co - Linked in the post
- vision.falcon.aidrc.tii.ae - Linked in the post
CS-4 documentary released – first engineering film
CS-4 is the subject of the first engineering documentary from the company. The film follows its creation with lab footage and engineer interviews, covering topics such as wafer-scale history, cooling, and deployment.
Key points
Introducing the engineering team at 0: 00
Covering the cooling problem at 4: 00
Detailing CS-4 vs. CS-3 performance at 20:01
Showing 3D Stacked DRAM at 20: 49
Sources
Fable 5.2 improves robot control on RoboDojo benchmark
Fable 5.2 shows substantially improved skills in controlling physical robots, even on tasks it has never explicitly been trained to do. On RoboDojo, a simulated two-arm robot benchmark, it outperformed other models and most AI systems built to control robots.
Key points
Improved robot control on tasks without explicit training.
Outperformed other models and most AI systems on the RoboDojo two-arm robot benchmark.
Sources
Fable 5.2 improves robot control on RoboDojo benchmark
Fable 5.2 shows substantially improved skills in controlling physical robots, even on tasks it has never explicitly been trained to do. On RoboDojo, a benchmark with a simulated two-arm robot, it outperformed other models and most AI systems built to control robots.
Key points
Improved robot control skills on unseen tasks
Outperformed other models on RoboDojo two-arm robot benchmark
Sources
HiRE training-free reward recipe calibrates similarity-based rewards
HiRE is a training-free reward recipe that calibrates similarity-based rewards such as DINOv2 into control-aware rewards. It uses only hindsight from the robot’s own rollouts.
Key points
Introduces HiRE, a training-free reward recipe.
Calibrates similarity-based (e.g., DINOv2) rewards into control-aware ones using only hindsight from the robot’s own rollouts.
