Mostik team with Fields Medalist claims hybrid 4B Qwen + 753B GLM boosts ARC-AGI 3 via efficient inter-model communication
A four-month-old team called Mostik, which includes Fields Medalist Stanislav Smirnov, claims that combining a 4B Qwen model on-device with a 753B GLM-5.2 in the cloud achieves high ARC-AGI 3 scores by improving inter-model communication efficiency. The team says their bridge method reduces the performance gap between the small and large models by about 50%, boosts the small model's accuracy by 25%, and cuts the large model's inference cost to roughly one-twentieth. Technical details are withheld because the competition is still ongoing.
Coverage timeline
量子位量子位
菲尔兹奖得主入局大模型！4B手机Qwen+云端GLM刷爆ARC-AGI 3 衡宇 2026-09-07 16:36:30 来源： 量子位 “在两个模型之间找到数学上的共同基础其实非常困难” 衡宇 发自 凹非寺 量子位 | 公众号 QbitAI 就离谱？ GPT-6 Astra刷到快满分的ARC-AGI 3， 一个手机能跑的4B小模型也能刷爆 。 更准确地说，是4B Qwen+753B GLM 5.2的组合。 其中，GLM-5.2待在云端，手机端只需运行4B的Qwen-3.5。大模型全程不输出一个字，小模型负责把最终答案写出来。 具体技术细节尚未公开，背后团队Mostik拒绝在此刻披露，“因为比赛还没结束”。 但他们公开了核心发现： AI模型之间一直在用一种极其低效的方式交流。 Mostik认为，他们搭了一座桥，让模型之间的信息传递效率提升了几个数量级。 按团队披露的实验结果，这套方案能让4B模型补上与753B模型之间约50%的性能差距，自身准确率提高25%。在大小模型差距更明显的难题子集上，提升达到2倍。 而大模型侧的推理成本，被压到了原本约二十分之一。 Mostik成立仅4个月，团队共15人，其中12名博士。 团队里还有一位菲尔兹奖得主——2010年证明了统计物理中“渗流模型”和“平面Ising模型”的共形不变性的俄罗斯数学家 Stanislav Smirnov （斯坦尼斯拉夫·斯米尔诺夫）。 两兆字节与十七比特 大模型生成一个token的过程中，内部会构建超过一百个隐藏向量，总共包含大约一百万个数值，约两兆字节的内部状态信息。 然后，它从大约15万词的词表中挑出一个token输出。 17个比特离开模型，两兆字节被丢弃，不会再被使用。 目前所有让多个模型协作的系统，包括编程子智能体、模型委员会、路由调度，全部建立在这17个比特之上。 模型之间唯一的沟通渠道就是文本。 那么被丢掉的两兆字节里到底装着什么？ 近两年的可解释性研究给出了明确答案，是深层计算的核心内容。 一篇ICLR 2026发表的论文显示，Qwen-3在写出“accountant”这个词之前好几个token，内部就已经携带了这个词的表征，正是这个表征让模型提前选对了冠词“an”而非“a”。 而且 模型越大，这种提前规划的能力越强，内部状态中携带的”未说出口”的信息越多。 Anthropic的可解释性团队
Related stories
Z.ai debuts GLM-5.3 with post-training scaling gains, open weights in two weeks
Z.ai (Zhipu AI) announced GLM-5.3, an open-source model built on the same base as GLM-5.2 with substantially extended post-training, yielding a 50% gain on internal coding benchmarks and state-of-the-art results among open models on Terminal Bench 3.0 and Agents' Last Exam (CLI). The model matches Mythos 5 in white-box code review and vulnerability discovery, and scored 60 on the Artificial Analysis Intelligence Index, tying Kimi K3. Model weights will be released in two weeks after security hardening; API pricing is set at $1.40 per million input tokens and $4.40 per million output tokens.
9 reports · 6 sources1154 points · 573 comments
Z.ai confirms Ox Alpha is GLM-5.3 Flash, releases open weights
Z.ai confirmed that the anonymous model 'Ox Alpha', which went viral on OpenRouter, is a new iteration of its GLM series, specifically GLM-5.3 Flash, and released its weights on Hugging Face under a new license. The model, a 320B-parameter natively multimodal model running on domestic chips, topped OpenRouter's leaderboard and ended DeepSeek's 56-day streak on OpenCode. The license requires companies with over $10B in revenue to pass a security review.
9 reports · 4 sources801 points · 281 comments
Zhipu AI open-sources GLM-5.3-Flash, a 320B multimodal model priced at 1/40 of Claude Opus 4.8
Zhipu AI released and open-sourced GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series, with 320B total parameters (18B active). It scores 57 on the Artificial Analysis Intelligence Index, matching Claude Opus 4.8, and is priced at one-tenth of GLM-5.3, or one-fortieth of Opus 4.8. The model uses a hybrid sparse and linear attention architecture, runs on domestic Chinese chips, and was tested anonymously as Ox-Alpha on OpenRouter and OpenCode.
8 reports · 5 sources1126 points · 573 comments
Alibaba open-sources Qwen3.8-27B, a 27B multimodal dense model
Alibaba's Qwen team released the open-weights Qwen3.8-27B, a 27B-parameter native multimodal dense model under the Apache 2.0 license. It supports 262K native context (extendable to 1M tokens) and outperforms Qwen3.7-Plus in coding and office tasks, with independent benchmarks showing it matching GPT-5.6 Luna on the Artificial Analysis Intelligence Index. The model surpassed 1 million downloads within days of release.
11 reports · 4 sources1421 points · 788 comments
News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe