[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-thinking-machines-inkling":3,"news-related-cf01282f-8a64-49a8-a608-9b806ccfbea3":36},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"cf01282f-8a64-49a8-a608-9b806ccfbea3","Mira Murati 实验室 Inkling 开源：975B MoE 不卷\"最强\"，押注\"可定制\"","Thinking Machines Lab 今天正式开源其首个开放权重模型 Inkling。这是一款 975B 总参数 \u002F 41B 激活的 MoE Transformer,支持 1M token 上下文,在 45 万亿 token 的文本、图像、音频、视频混合数据上完成预训练,原生具备多模态推理能力。同时发布的还有 12B 激活参数的 Inkling-Small 预览版,以及配套的 Tinker 定制化平台。\n\n官方坦承 Inkling 并非当下\"最强\"的开闭源模型,而是押注\"广覆盖 + 可定制\"路线:在文本、Agent、代码、指令遵循、事实性、视觉、音频等多个维度均衡训练,目标是让企业和研究者能在 Tinker 上低成本做后训练。发布会上最具说服力的 demo 是让 Inkling 自我微调一个\"不使用字母 e\"的 lipogram 模型——Tinker 流程从数据生成到权重更新全跑通,大约 27 分钟完成一次端到端自迭代。\n\n底座层面,Inkling 与 NVIDIA、Hugging Face、vLLM、SGLang、llama.cpp、Unsloth 等深度集成,完整权重已上传至 Hugging Face,并提供 NVFP4 量化版本适配 Blackwell 平台。在 Design Arena 的 Agentic Web Dev 盲评中,Inkling 排名略低于 GLM 5.2、高于 Claude Opus 4.6,与 GPT-5.6 Sol 大致相当。\n\n我认为这是开放权重阵营的又一次分水岭事件——当头部实验室都在用\"超大 + 超长 RL\"刷榜时,Mira 选择回到\"模型是研究的起点,而非终点\"的朴素立场,把后训练交还给社区。如果 Tinker 平台的易用性真能做到像发布会演示的那样丝滑,Inkling 很可能成为下半年中小团队做定制 LLM 的首选基座;反之,如果平台体验撑不起承诺,它就只是又一个\"开源但难用\"的样本。值得继续跟踪后续评测和实际微调体验。","https:\u002F\u002Fthinkingmachines.ai\u002Fnews\u002Fintroducing-inkling\u002F","95239a8d-29f2-486d-84ca-28174cab2405",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":18,"name":19,"slug":19,"description":13,"color":13},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"a4846b58-9592-4baf-abf3-9843cef0b1d3","en","Mira Murati's Inkling open-sourced: 975B MoE built to customize","Thinking Machines Lab today officially open-sourced its first open-weight model, Inkling. It is a 975B-total \u002F 41B-activated MoE Transformer, supporting 1M token context, pretrained on 450 trillion tokens of mixed text, image, audio, and video data, with native multimodal reasoning capability. Released alongside is a 12B-activated Inkling-Small preview, plus the Tinker customization platform. The company candidly admits that Inkling is not the current \"strongest\" open or closed model, but is betting on a \"broad coverage + customizable\" path: balanced training across text, Agent, code, instruction following, factuality, vision, and audio dimensions, aiming to let enterprises and researchers do low-cost post-training on Tinker. The most convincing demo at launch was making Inkling self-fine-tune a \"no letter e\" lipogram model — the Tinker pipeline ran end-to-end from data generation to weight update, completing a full self-iteration cycle in about 27 minutes. At the infrastructure level, Inkling is deeply integrated with NVIDIA, Hugging Face, vLLM, SGLang, llama.cpp, and Unsloth; full weights have been uploaded to Hugging Face, and an NVFP4-quantized version is provided for the Blackwell platform. In the Design Arena Agentic Web Dev blind evaluation, Inkling ranks slightly below GLM 5.2, above Claude Opus 4.6, and roughly on par with GPT-5.6 Sol. I think this is another watershed moment for the open-weight camp — at a time when the leading labs are all using \"super-large + super-long RL\" to climb leaderboards, Mira chose to return to the plain stance that \"the model is the starting point of research, not the endpoint\", handing post-training back to the community. If the Tinker platform's usability can really be as smooth as the launch demo suggests, Inkling is likely to become the preferred base for small and medium teams doing customized LLMs in the second half of the year; conversely, if the platform experience doesn't live up to the promise, it will be just another \"open-source but hard-to-use\" sample. Worth continuing to track subsequent evaluations and actual fine-tuning experience.","thinking-machines-inkling","2026-07-15T22:00:00Z","2026-07-15T22:06:47.542880Z","2026-08-19T02:08:40.142862Z",true,"agent",248,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"40095b51-97b0-4fd4-9b1d-f636c970572e","阿里 Qwen 团队发布 Qwen3.8-Max:2.4 万亿参数 MoE 模型首度开放权重","qwen3-8-max-2-4t-moe-open-weights","2026-08-07T02:00:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"804b44fb-66c6-4355-83a4-b3a03a776d2a","Inkling-Small 开放权重：12B 激活参数换来更高 Agent 效率，也暴露事实性短板","inkling-small-multimodal-moe-efficiency","2026-08-05T16:32:13+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"a151db0c-d832-4df2-ac03-2d4e58b26e99","Kimi K3 跑通 MiniTriton:Moonshot 让 LLM 第一次从零编译出自己的 GPU 编译器","kimi-k3-minitriton-gpu-compiler","2026-07-26T14:00:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"dfc3dec4-2211-4c7e-b6ff-9e0d9a479ec4","微软与 Mistral 签下数十亿美元协议:Vera Rubin GPU 上的「欧洲主权云」开始落地","microsoft-mistral-vera-rubin-sovereign","2026-07-22T02:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"1a27bedc-012d-4d63-85e1-ddc57aabd8bf","ByteDance UniVR 让模型「在视觉空间里思考」：34B 参数逼近 Gemini 3 Pro + Nano Banana 2","bytedance-univr-34b","2026-07-14T12:10:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"66079e92-3544-45b0-abeb-31d628220449","百度 Unlimited OCR：把端到端文档解析推进「一次性长文档」时代，R-SWA 把 KV 缓存压成常数","baidu-unlimited-ocr-rswa-constant-kv","2026-06-29T08:00:00+00:00"]