TokenWave AI has released SWE-CoWork, a benchmark that places coding agents in a simulated workplace where 82% of requirements live with teammates, forcing agents to ask the right person to learn what to build. Across 101 tasks drawn from real open-source repositories, the best models pass roughly 63% of individual checks but fully complete only 8 of 101, exposing long-horizon collaboration and information gathering as core weaknesses.
Long-horizon Tasks
长周期任务
Complex agent tasks requiring multi-step planning, resource use, and iterative validation over hours-long sandboxes.
Curated Intelligence
Researchers released FrontierML, a benchmark measuring how well agents convert capital into model performance across 21 end-to-end ML tasks. Each task gives a 12-hour CPU sandbox, a $10 to $200 wallet, and validation samples, and frontier systems already turn that budget into measurable self-improvement. The result highlights advancing AI R&D capability and hints at a machine economy where agents sub-contract work to one another.
The author ran 600 experiments to systematically probe planning in JEPA world models, showing that changing only the planning procedure lets a model solve long-horizon physics tasks with zero additional training. The key finding is that transforming a memory graph to generate subgoals on the fly enables generalization to out-of-distribution goals, though the model still cannot choose its own subgoals without demonstrations.
Google has published the first Android Bench results for Kimi K3.1 Pro and DeepSeek V4.1 Flash, with Kimi reaching a 24.0% pass rate above Claude and Gemini entries while DeepSeek delivers strong cost efficiency at $9.9 per run. The benchmark targets real Android engineering tasks, giving developers a quantified reference for long-horizon agent capability.
Google released Gemini 4 Carbon just two weeks after Argon, positioning it for sustained long-horizon tasks. Early testing shows 0% shortcut rate on Terminal-Bench and 0% prompt-injection success on Gray Swan's benchmark, 2x the secure-code rate of Argon, and 88% vulnerability detection on Sec-Bench Pro. Pricing at $2/$10 per million tokens with 95% off cached input makes multi-day agents economically viable.
Tencent Hunyuan's Hy4 hit quota limits on its first day of launch inside WorkBuddy, prompting a tenfold emergency expansion of its inference cluster and higher free-tier quotas, while Hy3's free access was extended to November 15. Compared with the preview, Hy4 substantially scales pre- and post-training, adds vision for the first time, and is billed as the company's first entry among top open-source models, with weights and a system card promised soon.
Tencent released Hy4, substantially scaling both pre-training and post-training over its preview version and adding vision capabilities for the first time. The model claims major gains in instruction following, long-context understanding, long-horizon tasks and deep research, positioning Hy among the strongest open-source models and emphasizing product co-design for real-world productivity.
Tencent has released the official Hunyuan Hy4 model, substantially scaling pre- and post-training versus the preview and adding vision understanding and computer-use capabilities for the first time, alongside gains in instruction following, long-context comprehension and long-horizon tasks. Tencent claims Hy4 is its first model to rank among the strongest open-source models and plans to release weights with vLLM/SGLang ecosystem support, while acknowledging a sizable gap to top closed models on hard math and long-horizon coding benchmarks.
At CeMAT 2026, Geek+ unveiled its proprietary Gravity embodied AI model to address the trade-off between generalization and efficiency in warehouse robotics. The model enables humanoid robot Gino 1 to handle complex long-horizon tasks and grasp unknown objects, marking a shift from single-unit intelligence to system-level collaborative productivity.
Google DeepMind released Gemini 4 Carbon, claiming state-of-the-art performance in cybersecurity and long-horizon tasks. The model is currently restricted to verified cyber defenders, aiming to enhance proactive digital infrastructure defense through AI.