[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fzqpbjamrndqf":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":12,"bodyZh":13,"bodyEn":14,"sourceName":15,"sourceUrl":16,"updatedAt":17},"ling-3-1-flash","NEWS","2026-10-02T05:05:00.000Z","蚂蚁百灵发布 Ling-3.1-flash：560B 总参数 \u002F 25B 激活，连续 17 小时写出编译器","Ant's InclusionAI Ships Ling-3.1-flash: 560B Total \u002F 25B Active, Wrote a Compiler in 17 Straight Hours","蚂蚁百灵（InclusionAI）发布 Ling-3.1-flash：约 560B 总参数、25B 激活，混合线性注意力架构，1M 上下文（体验期 256K）；官方演示连续 17 小时从零写出 Lua 到 x86-64 编译器（182 项测试过 178）。目前未开源，官方称转付费后计划开源。","Ant Group's InclusionAI has released Ling-3.1-flash: ~560B total parameters with ~25B active, a hybrid linear-attention architecture, and a 1M-token context (256K during the free trial). Official demos show it writing a Lua-to-x86-64 compiler from scratch over 17 straight hours (178\u002F182 tests passed). Not yet open-sourced — the weights are planned for release when the paid tier opens.",[],[],"## 发布\n\n2026 年 9 月 30 日，蚂蚁集团旗下 **InclusionAI（蚂蚁百灵）** 发布 **Ling-3.1-flash**。相比上一代 Ling-3.0-flash（124B 总参数 \u002F 5.1B 激活），总规模扩大约 **4.5 倍**。\n\n## 架构与规格\n\n- 总参数约 **560B**，每 token 激活约 **25B**（稀疏比约 1:22）\n- 延续**混合线性注意力**架构并提高线性层比例：7 层 KDA + 1 层 Gated MLA\n- 512 个路由专家，每 token 激活 8 个 + 1 个共享专家\n- 上下文上限 **1M token**（当前体验期开放 256K，转付费后开放 1M）\n\n## 官方演示（自报数据）\n\n- 连续约 **17 小时**从零编写 Lua 到 x86-64 ELF 编译器：182 项测试通过 178 项（97.8%）\n- 约 20 小时将 C 图像库移植到 Rust：8.015 倍加速，30 项正确性检查全部通过\n- 基准：GDPVal-AA v2.1 达 1,673 Elo、FrontierSWE 75.16、HealthBench Professional 65.35\n\n## 开源状态：尚未开源\n\n- 官方口径：两周免费体验结束转为付费服务后开放 1M 上下文，**届时计划同步开源模型**\n- 截至 10 月 1 日，Hugging Face 上**没有权重仓库、许可证与模型卡**，所有分数均为官方自报，无第三方独立评测\n- 参考上一代 Ling-3.0-flash 发布到开源约两周间隔，本次节奏预计类似\n\n## 体验渠道\n\n现已通过 **Novita AI**、**Vercel AI Gateway** 与蚂蚁自家 **Ling Studio**（ling.tbox.cn）开放两周免费体验，定位通用智能体、搜索、办公与软件研发，并覆盖医疗、金融、材料科学等专业场景。\n\n> 若按计划开源，560B\u002F25B 激活的稀疏比意味着 decode 显存压力集中在 25B 激活侧——对大显存单卡\u002F多卡本地部署会是一个有意思的新选项。我们将持续关注其权重落地。","## Launch\n\nOn September 30, 2026, Ant Group's **InclusionAI** released **Ling-3.1-flash**. Compared with the previous Ling-3.0-flash (124B total \u002F 5.1B active), the total size grows about **4.5×**.\n\n## Architecture and Specs\n\n- ~**560B** total parameters, ~**25B** active per token (sparsity ratio ~1:22)\n- Continues the **hybrid linear-attention** design with a higher share of linear layers: 7 KDA layers + 1 Gated MLA\n- 512 routed experts, 8 active per token + 1 shared expert\n- Up to **1M-token context** (256K during the trial; 1M unlocks with the paid tier)\n\n## Official Demos (Self-Reported)\n\n- ~**17 straight hours** writing a Lua-to-x86-64 ELF compiler from scratch: 178 of 182 tests passed (97.8%)\n- ~20 hours porting a C image library to Rust: 8.015× speedup, all 30 correctness checks passed\n- Benchmarks: GDPVal-AA v2.1 at 1,673 Elo, FrontierSWE 75.16, HealthBench Professional 65.35\n\n## Open-Source Status: Not Yet Open\n\n- Official line: when the two-week free trial converts to paid, the 1M context opens and the model is **planned to be open-sourced alongside**\n- As of October 1, there is **no weights repository, license, or model card on Hugging Face**; all scores are self-reported with no third-party evaluation\n- The previous Ling-3.0-flash took about two weeks from launch to open weights; expect a similar cadence\n\n## Where to Try\n\nA two-week free trial is live via **Novita AI**, **Vercel AI Gateway**, and Ant's own **Ling Studio** (ling.tbox.cn), positioned for general agents, search, office work, and software development, plus medical, finance, and materials-science scenarios.\n\n> If the weights land as planned, the 560B\u002F25B sparsity means decode-time memory pressure concentrates on the 25B active side — an interesting new option for local multi-GPU or large-VRAM single-card deployments. We'll keep watching.","The BlockBeats","https:\u002F\u002Fwww.theblockbeats.info\u002Fflash\u002F369865","2026-10-02T02:08:31.676Z"]