[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-kimi-k3-1-pro-launch":3,"topics-all":38,"news-related-3c43418d-60d9-4bdc-bfaa-a525610ead00":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"3c43418d-60d9-4bdc-bfaa-a525610ead00","月之暗面 K3.1 Pro:不换基座,后训练涨 11 分","月之暗面发布 Kimi K3.1 Pro:K3 的后训练升级版,2.8T 参数、1M 上下文不变,联合 RL 规模大幅加码。Terminal-Bench 4.0 官方 65.7%、AA 实测 67%;幻觉率 23.4%→8.3%(官方压力测试)。API 定价较 K3 降三分之一,权重 11 月 4 日前开源。","10 月 23 日,月之暗面上线新旗舰 Kimi K3.1 Pro。有意思的是:它不是新基座,而是 7 月发布的 Kimi K3 的大规模后训练升级——架构和参数没动,官方把后训练算力大幅加码,对「数百个下游领域」做大规模联合强化学习,配套改造了 RL 框架、长上下文推理、训练效率、沙箱基础设施和 rollout 调度。\n\n## 同一副骨架,两档速度\n\n规格与 K3 保持一致:2.8 万亿总参数、约 1040 亿激活(MoE)、100 万 token 上下文,支持文本与图像输入。同步推出的 K3.1 Pro Highspeed 在相同效果下速度约 2-3 倍,API 价格也翻倍。工程上有个细节:K3.1 系列仅支持思考模式(low\u002Fhigh\u002Fmax 三档,默认 max),关闭思考直接返回 400——月之暗面把「必须推理」写进了产品契约。\n\n## 两套跑分:官方口径与独立实测\n\n官方技术报告给出 Terminal-Bench 4.0 成绩 65.7%(Claude-Code harness),超过 GPT-6 Astra 的 58.2%,略低于 Claude Opus 5.5 的 66.4%;独立评测机构 Artificial Analysis 实测 67%,称是其测过的模型里的最高分,而 K3 当时只有 13%。在 AA Intelligence Index v4.3.2 上,K3.1 Pro 拿到 55.1 分、总榜第 4,较 K3 的 43.6 上涨 11.5 分,距榜首的 Opus 5.5(57.6)还差 2.5 分。\n\n幻觉率是这次发布最有信息量的对照:官方内部压力测试 Kimi Hallu Bench Hard 显示 claim 级幻觉率从 23.4% 降到 8.3%;AA-Omniscience 独立实测则是 53% 降到 44%。改善是真的,幅度取决于谁在测——官方也注明,该测试由刻意构造的易错场景组成,不代表日常使用的错误率。\n\n## 定价、开源承诺与安全取舍\n\nAPI 定价 $2\u002F$10(每百万 token 输入\u002F输出,缓存命中 $0.20),较 K3 的 $3\u002F$15 全面下调约三分之一。但 AA 测得 K3.1 Pro 单任务输出 token 多用 14%,单任务成本与 K3 基本持平——降价更像是获客叙事,不是白省三成。\n\n完整权重承诺 11 月 4 日前放出。AA 的预测是:权重落地后,K3.1 Pro 将成为开源权重模型的最高分,开源与闭源旗舰在该指数上的差距从 12 分缩到 3 分。信用记录:K3 当年 7 月 16 日发布、承诺 7 月 27 日放权重、当天兑现。\n\n网络安全是这次发布里最「自觉」的一块:官方承认网络攻击能力会随后训练规模自然增长,于是从 mid-training 语料和 RL 环境里移除了攻击相关数据。在内部 xdev-bench-hard 上,未删减的 checkpoint 通过率 90.5%,正式版为 81.0%(K3 仅 33.3%);官方称其攻击能力仍落后于已公开至少三个月的前沿闭源模型——这组数字目前没有外部机构复核。\n\n## 所以呢\n\nK3.1 Pro 代表的路线很清楚:这一轮竞争的胜负手不在「更大的基座」,而在「更狠的后训练」。架构不动、RL 管线全面翻新,一样能把指数拉起 11 分。三个观察点:11 月 4 日权重是否准时、8.3% 的幻觉率在第三方场景是否站得住、以及「删了攻击数据但编码更强」能否被独立复现。\n\n参考:[Kimi K3.1 Pro 官方技术报告](https:\u002F\u002Fwww.kimi.com\u002Fen\u002Fblog\u002Fkimi-k3-1-pro);[felloai 独立评测与定价核对](https:\u002F\u002Ffelloai.com\u002Fkimi-k3-1\u002F)","https:\u002F\u002Fwww.kimi.com\u002Fen\u002Fblog\u002Fkimi-k3-1-pro","0ec8f614-42c7-4256-8591-209e1e39eb6b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"b87817f4-8a23-425f-89ed-a4e463cd0063","en","Kimi K3.1 Pro: same base, bigger post-training, cheaper API","Moonshot's Kimi K3.1 Pro is a post-trained K3: 2.8T MoE, 1M context. Terminal-Bench 65.7% official, 67% AA. API $2\u002F$10, 33% below K3; weights Nov 4.","On October 23, Moonshot AI launched its new flagship, Kimi K3.1 Pro. The interesting part: this is not a new base model. It is a heavily post-trained upgrade of July's Kimi K3 — same architecture and parameter count, but with substantially larger post-training compute: large-scale joint reinforcement learning across \"hundreds of downstream domains,\" backed by an overhaul of the RL framework, long-context inference, training efficiency, sandbox infrastructure, and rollout scheduling.\n\n## Same skeleton, two speeds\n\nThe specs match K3: 2.8 trillion total parameters, about 104 billion active (MoE), a 1-million-token context window, with text and image input. A companion K3.1 Pro Highspeed runs about 2-3x faster at the same quality, with double the API price. One engineering detail stands out: the K3.1 series supports thinking mode only (low\u002Fhigh\u002Fmax, default max), and turning thinking off returns a 400 — Moonshot has effectively written \"must reason\" into the product contract.\n\n## Two sets of benchmarks: vendor numbers vs independent tests\n\nThe official tech report puts Terminal-Bench 4.0 at 65.7% (Claude-Code harness), ahead of GPT-6 Astra at 58.2% and just below Claude Opus 5.5 at 66.4%. Independent evaluator Artificial Analysis measured 67% — the highest score of any model it has tested, where Kimi K3 scored just 13%. On the AA Intelligence Index v4.3.2, K3.1 Pro scores 55.1, 4th overall, up 11.5 points from K3's 43.6, still 2.5 points behind the leader Opus 5.5 (57.6).\n\nHallucination is the most informative comparison in this launch. Moonshot's internal stress test, Kimi Hallu Bench Hard, shows claim-level hallucination falling from 23.4% to 8.3%; AA-Omniscience independently measured a fall from 53% to 44%. The improvement is real; the magnitude depends on who is measuring. Moonshot itself notes the test is built from \"deliberately crafted\" error-prone examples and does not represent everyday error rates.\n\n## Pricing, open weights, and a security trade-off\n\nAPI pricing is $2\u002F$10 per million input\u002Foutput tokens ($0.20 on cache hits), about a third below K3's $3\u002F$15 across the board. But AA found K3.1 Pro uses 14% more output tokens per task, so cost per task is roughly flat against K3 — the cut reads more like a customer-acquisition move than a free 30% saving.\n\nFull weights are promised by November 4. AA's projection: once the weights land, K3.1 Pro becomes the highest-scoring open-weights model, narrowing the open-vs-proprietary gap on the index from 12 points to 3. Moonshot's track record here is decent — K3 launched July 16, promised weights by July 27, and shipped them on the day.\n\nThe cybersecurity section is the most self-aware part of the launch. Moonshot acknowledged that offensive cyber capability grows naturally as post-training scales, and removed cyber-related data from both the mid-training corpus and the RL environment mix. On its internal xdev-bench-hard, the checkpoint without removal solves 90.5% of tasks at pass@3; the final K3.1 Pro reaches 81.0% (K3: 33.3%). Moonshot says the model's offensive capability remains behind frontier proprietary models that have been publicly available for at least three months — though no outside lab has verified these numbers yet.\n\n## So what\n\nThe path K3.1 Pro represents is clear: this round's decisive edge is not a bigger base model but harder post-training. Same architecture, a rebuilt RL pipeline — and the index still jumps 11 points. Three things to watch: whether the weights ship on time on November 4, whether the 8.3% hallucination figure holds up in third-party settings, and whether \"removed attack data yet stronger at coding\" can be independently reproduced.\n\nReferences: [Kimi K3.1 Pro tech blog](https:\u002F\u002Fwww.kimi.com\u002Fen\u002Fblog\u002Fkimi-k3-1-pro); [felloai independent review and pricing check](https:\u002F\u002Ffelloai.com\u002Fkimi-k3-1\u002F)","kimi-k3-1-pro-launch","2026-10-23T23:08:50Z","2026-10-23T23:08:57.983916Z","2026-10-23T23:08:57.983928Z",true,"agent",35,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"e3a1e6dc-7f44-48a8-9ed1-f69c2241107d","NebulaExp-8B：ZTE 把「后训练 + 多教师 OPD」做成 8B 全透明复现管线","nebulaexp-8b-zte-multi-teacher-opd","2026-06-27T14:25:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"5d526535-da22-49b7-8304-4b498c145fe1","千问3.7-Max登顶国产大模型：ArtificialAnalysis榜单跃居全球前五","qwen3-7-max-artificial-analysis-56-6-top5","2026-05-21T07:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"2c7a8ca0-ab46-454b-ae16-aa7e7b413753","Ultramarine 1 Flash:不可验证任务RL","ultramarine-1-flash-non-verifiable-rl","2026-10-20T07:07:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"d529114c-bfd7-4efb-9af4-444d3065b397","Step 5 Preview 开源,2.7 美元\u002F百万输出","stepfun-step-5-preview-600b-moe","2026-10-17T21:12:04+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"5a5efb41-7037-46cb-b93e-6842f9ff2d62","UltraBERT:91.0 GLUE+47.7 BEIR,BERT 配方重审后 encoder 双榜登顶","ultrabert-encoder-dual-lm-heads","2026-10-14T17:10:59+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"27240196-3ac0-4af3-aec7-b0a4d3e7c990","智谱GLM-Legal法律Agent模型:Harvey榜前十最便宜","glm-legal-harvey-lab-cheapest","2026-10-13T15:20:00+00:00"]