Tag: #llm90 results
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. This enables linear-time absolute measurements with stable interpretability, unlike volatile LLM-vs-LLM rankings.
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. This marks the first release in Alibaba's Spring Festival AI lineup.
Alibaba launches Qwen 3.5, strongest open-source model with 397B parameters on除夕. It surpasses Gemini 3 in benchmarks. Inference costs just 0.8 yuan per million tokens.
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. Achieves superior performance with efficient architecture.
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. The model outperforms predecessors and rivals large instant models in reasoning and instruction following.
Apple introduces asynchronous verified semantic caching to optimize tiered LLM architectures. It addresses tradeoffs in static and dynamic caches using embedding similarity thresholds. This reduces inference cost and latency in production workflows like search and agents.
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. Features 1.3x-1.7x throughput gains via dense MTP.
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. Users mourn loss despite two-week notice.
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.