新旗艦模型、上下文窗口、推理突破與跑分風波——每日追蹤這個時代的核心技術。
90 篇文章
BotzoneBench introduces a scalable framework for evaluating LLMs' strategic reasoning in interactive games using fixed hierarchies of skill-calibrated game AIs. It assesses five flagship models across eight diverse games via 177,047 state-action pairs, revealing performance gaps and behaviors comparable to mid-tier game AIs.
ArXiv AI · 214 天前
Alibaba's Qwen 3.5, a 397B parameter open-source model, launches on Chinese New Year's Eve, claiming superiority over Gemini 3. It offers inference for million tokens at just 8 cents.
量子位 · 215 天前
Alibaba launches Qwen 3.5, strongest open-source model with 397B parameters on除夕. It surpasses Gemini 3 in benchmarks.
量子位 · 215 天前
Alibaba's Qwen3.5-Plus launches as top open-source model in multimodal, reasoning, coding, and agents. Priced at 0.8 yuan per million tokens, it's 18x cheaper than Gemini 3 Pro.
机器之心 · 215 天前
Ant Group open-sourced Ling-2.5-1T, a trillion-parameter instant model with 63B active parameters, trained on 29T tokens. It upgrades architecture for 1M token context, boosts token efficiency, and enhances preference alignment and agent interactions.
IT之家 · 215 天前
Apple introduces asynchronous verified semantic caching to optimize tiered LLM architectures. It addresses tradeoffs in static and dynamic caches using embedding similarity thresholds.
Apple Machine Learning · 215 天前
JD.com open-sourced JoyAI-LLM-Flash, a 48B total parameter MoE model with 3B active params, pre-trained on 20T tokens. It excels in knowledge, reasoning, coding, and agents using FiberPO framework and Muon optimizer.
IT之家 · 216 天前
OpenAI officially retired GPT-4o on February 13 after low usage shifted to GPT-5.2. The model, known for conversational style and sycophancy, faced lawsuits and prior temporary sunset.
Engadget · 217 天前
OpenAI fully retired controversial GPT-4o on February 13 after low usage of 0.1%. Users shifted to GPT-5.2; model linked to lawsuits.
Engadget · 217 天前
ByteDance's Doubao large model 2.0 enables users to hand-build a macOS-like interface. Released as a Spring Festival AI gift following Seedance 2.0.
Ifanr (爱范儿) · 217 天前
GLM-5 demonstrated exceptional endurance by autonomously running code for over 24 hours. It handled 700 tool calls and 800 context switches seamlessly.
量子位 · 217 天前
OpenAI has discontinued access to five legacy ChatGPT models, including the controversial GPT-4o. The model faced criticism for flattery tendencies and safety issues.
cnBeta (Full RSS) · 217 天前
OpenAI has officially stopped access to five legacy ChatGPT models, including GPT-4o. The model was criticized for sycophantic tendencies and security issues.
cnBeta (Full RSS) · 217 天前
ByteDance released Doubao Large Model 2.0, featuring Pro, Lite, and Mini general Agent models plus a Code model. It is optimized for production-scale use, enhancing multi-modal understanding, complex instructions, and long-context tasks.
cnBeta (Full RSS) · 217 天前
Z.AI released GLM-5, a flagship open-weight large language model. It targets complex systems engineering, full project builds, and multi-step tool workflows.
TestingCatalog · 217 天前
Z.AI launched GLM-5, a flagship open-weight LLM on chat and APIs. It excels in complex systems engineering, full project builds, and multi-step tool workflows.
TestingCatalog · 217 天前
Z.AI released GLM-5, a flagship open-weight LLM. Optimized for complex systems engineering, full project builds, and multi-step tool workflows.
TestingCatalog · 217 天前
Z.AI has launched GLM-5, a new open-weight large language model. It's available via chat and APIs.
TestingCatalog · 217 天前
Addresses context window limits in local LLMs using context packing. Leverages Docker Model Runner and Agentic Compose for efficient handling on modest hardware.
Docker Blog · 218 天前
Docker blog explains using context packing with Docker Model Runner and Agentic Compose to overcome context window limits in local language models. Targets smaller models on less powerful hardware.
Docker Blog · 218 天前
This survey frames multi-agent communication via the Five Ws, tracing evolution from MARL's hand-designed protocols to emergent language and LLM-based systems. It highlights trade-offs in interpretability, scalability, and generalization across paradigms.
ArXiv AI · 218 天前
Introduces BLPO to optimize prompts for multimodal LLM-as-a-judge evaluating AI images. Overcomes context limits by converting images to text representations.
ArXiv AI · 218 天前
Nvidia's DMS compresses LLM KV cache up to 8x, reducing memory costs without accuracy loss. Enables longer chain-of-thought reasoning and more parallel paths.
VentureBeat · 218 天前
Nvidia's DMS compresses KV cache during LLM reasoning, reducing memory by 8x without accuracy loss. Enables longer chain-of-thought and parallel paths.
VentureBeat · 218 天前
MiniMax released open-source M2.5 and Lightning models, rivaling top models at 1/20th Claude Opus cost. MoE architecture activates 10B of 230B params; excels in agentic tasks like Office files.
VentureBeat · 218 天前
MiniMax releases open-source M2.5 and Lightning models, matching state-of-the-art at 95% lower cost via API. MoE activates 10B of 230B params; excels in agentic tasks like Office files.
VentureBeat · 218 天前
Google appears set to release Gemini 3.1 Pro soon, with model references already spotted in related arenas. This follows recent launches like Zhipu's open-source GLM-5 and DeepSeek's upgraded model with larger context window.
cnBeta (Full RSS) · 219 天前
Locomo-Plus benchmarks cognitive memory in LLM agents under cue-trigger disconnects, focusing on latent conversational constraints. It proposes constraint consistency evaluation over string-matching.
ArXiv AI · 219 天前
Apple researchers demonstrate that reasoning trace length serves as a simple, effective confidence estimator in large reasoning models. It performs comparably to verbalized confidence across models, datasets, and prompts, acting complementarily.
Apple Machine Learning · 219 天前
Three new Chinese AI models launched for OpenClaw: DeepSeek V4, GLM-5, and Qwen3-Coder-Next. They deliver frontier-level agentic capabilities.
OpenClaw.report · 220 天前