Mingyu's Library主页
深度学习文档 · AI Daily(08-03 补生成)Deep-Dive · AI Daily (backfilled 08-03)

Thinking Machines Inkling
——开放权重路线深读

Thinking Machines' Inkling
— The Open-Weights Bet, Explained

前 OpenAI CTO Mira Murati 的实验室在 7 月 15 日放出首个从零训练的开放权重模型 Inkling(Apache 2.0),7 月 31 日又用 Inkling Small 证明「三分之一参数、几乎同等智能」。本文讲清:它是什么、架构上有什么讲究、思考档位怎么用、和 Kimi/GLM/DeepSeek 怎么选,以及要避开的坑。

On July 15, the lab of former OpenAI CTO Mira Murati released Inkling, its first open-weights model trained from scratch (Apache 2.0); on July 31, Inkling Small showed "a third of the parameters, nearly the same intelligence." This document covers what it is, what's distinctive in the architecture, how the thinking-effort dial works, how it stacks up against Kimi/GLM/DeepSeek — and the pitfalls to avoid.

📅 调研日期:2026-08-03📅 Researched: 2026-08-03 🌐 主来源:官方发布长文 + Artificial Analysis + 媒体🌐 Primary: official launch post + Artificial Analysis + press 🀄/🅰️ 中英双语,右上角切换🀄/🅰️ Bilingual — toggle top right

1🔭 开篇速览:30 秒版本🔭 At a Glance: The 30-Second Version

是什么:Inkling 是一个 MoE(混合专家)Transformer:975B 总参数、41B 激活,上下文最长 1M token,在 45 万亿 token 的文本+图像+音频+视频上从零预训练,原生多模态推理,Apache 2.0 开源权重。官方定位很坦诚:「不是今天最强的模型」,而是一个为微调定制而生的宽基座——多模态、省 token、可控思考档位,且直接接入自家微调平台 Tinker。

What it is: Inkling is a Mixture-of-Experts transformer: 975B total parameters, 41B active, up to a 1M-token context, pretrained from scratch on 45 trillion tokens of text, images, audio and video, reasoning natively across modalities, with Apache 2.0 open weights. The official positioning is refreshingly candid: "not the strongest overall model available today" — but a broad base built for customization: multimodal, token-efficient, with a controllable effort dial, wired into the lab's Tinker fine-tuning platform.

为什么值得关注:它是开放权重阵营里罕见的「从零训练 + 全家桶思路」样本:模型(Inkling / Inkling-Small)+ 微调平台(Tinker)+ 交互系统(interaction models)一体设计;发布长文还公开了大量训练细节(30M rollouts 的大规模 RL、校准训练、反审查训练),透明度在同类中少见。

Why it matters: it's a rare open-weights specimen of "trained from scratch + full-stack thinking": models (Inkling / Inkling-Small) + a fine-tuning platform (Tinker) + an interaction system, designed together; and the launch post discloses unusual amounts of training detail (30M-rollout large-scale RL, calibration training, censorship-resistance training).

Inkling(旗舰)975B 总 / 41B 激活 · 1M 上下文多模态 · Apache 2.0 Inkling Small276B 总 / 12B 激活 · 256K07-31 发布 · 效率优先 Tinker 微调平台浏览器内微调 / Playgroundcookbook 配方 · tml-renderer64K / 256K 上下文档位 部署生态API:Together / Fireworks /Modal / Databricks / Baseten推理:SGLang · vLLM ·llama.cpp · transformers权重:Hugging Face(含 NVFP4 Blackwell 版) Inkling (flagship)975B total / 41B active · 1M ctxMultimodal · Apache 2.0 Inkling Small276B total / 12B active · 256KShipped 07-31 · efficiency-first Tinker fine-tuningIn-browser fine-tuning / PlaygroundCookbook recipes · tml-renderer64K / 256K context tiers Deployment ecosystemAPIs: Together / Fireworks /Modal / Databricks / BasetenInference: SGLang · vLLM ·llama.cpp · transformersWeights: Hugging Face(incl. NVFP4 for Blackwell)
图 1 · Inkling 家族与生态:模型—微调平台—部署伙伴一体设计。数据:官方发布长文(2026-07-15 / 07-31)。
Fig. 1 · The Inkling family and ecosystem: models, fine-tuning platform, and deployment partners designed as one. Data: official launch posts (2026-07-15 / 07-31).

2🎯 它在赌什么:反「一体适用」🎯 The Bet: Against One-Size-Fits-All

读这个模型,先读它的立场。Thinking Machines 的使命表述是「构建扩展人类意志与判断的 AI」,落到产品上是一个明确的赌注:未来不属于一个万能的封闭模型,而属于「拿一个够宽的开放基座,用自己的数据调成自己的模型」。所以发布的重心不是刷榜,而是三件事:

To read this model, read its stance first. Thinking Machines' mission line is "building AI that extends human will and judgment," which lands as a concrete bet: the future belongs not to one universal closed model, but to "take a broad open base and tune it into your own model with your own data." Hence the launch emphasizes three things over leaderboard-chasing:

宽而平衡的基座A broad, balanced base刻意训练成跨 Agent、推理、代码、指令遵循、事实性、视觉、音频的全能型,而非窄域刷分——因为要适配的下游工作流千差万别。Deliberately trained across agentic, reasoning, coding, instruction-following, factuality, vision and audio rather than narrowly optimized — because downstream workflows vary wildly.
Epistemics(求真品质)Epistemicsofficial 把「校准、指令遵循、反审查」合称模型的 epistemics:用适当的置信度说话(ForecastBench 上与 Grok 4.3、Gemini 3.1 Pro 同档)、该说「不知道」时说不知道;并专门训练了对可能被审查话题的直接回答(Cognition 的 Propaganda and Censorship Eval 验证)。VentureBeat 把「低成本 + 抗审查」总结为其两大卖点。The post groups calibration, instruction following, and censorship resistance as the model's epistemics: speaking with appropriate confidence (on par with Grok 4.3 and Gemini 3.1 Pro on ForecastBench), saying "I don't know" when warranted — and specifically training direct answers on censorable topics (verified via Cognition's Propaganda and Censorship Eval). VentureBeat summarized its pitch as "low cost and resistance to censorship."
微调即产品Fine-tuning as the product发布演示里最有态度的一幕:让 Inkling 在 OpenCode harness 里给自己写微调任务——起草方案、生成评测与合成数据、调 Tinker API 训练、加载新权重完成闭环。模型是原料,Tinker 才是主打商品。The launch's most pointed demo: Inkling, inside the OpenCode harness, fine-tuning itself — drafting the plan, generating evals and synthetic data, calling the Tinker API to train, then loading the new weights. The model is raw material; Tinker is the flagship product.
🧭 一个诚实的定位声明
🧭 An honest positioning statement

官方原话:「Inkling 不是今天可用的最强模型,无论开放或封闭」——它押的是「多模态 + 高 token 效率 + 可微调」这个组合拳。读任何厂商发布都该以此为镜:敢写明短板的发布,数字部分的可信度通常也更高。

In the post's own words: "Inkling is not the strongest overall model available today, open or closed" — the bet is the combination of multimodality, token efficiency, and tunability. Read every vendor launch against this mirror: posts that state their weaknesses plainly tend to have more trustworthy numbers too.

3🏗️ 架构与训练:几处有讲究的偏离🏗️ Architecture & Training: Deliberate Departures

MoE 主体大致沿用 DeepSeek-V3 配方,但有几处为效率和长上下文特意做的偏离,官方长文写得相当具体:

The MoE core largely follows the DeepSeek-V3 recipe, with several departures made deliberately for efficiency and long context — the post is unusually specific:

MoE 层:256 路由专家 + 2 共享专家,每 token 激活 6 个路由专家sigmoid 路由器 + 无辅助损失的负载均衡偏置;选中专家与共享专家联合归一化加权 注意力:滑窗层 : 全局层 = 5 : 1,8 个 KV 头位置编码用相对位置嵌入(非 RoPE)——官方称长序列外推更好;KV 投影与残差出口处加短卷积 多模态:无编码器(encoder-free)设计音频输入 dMel 频谱图;图像切 40×40 像素块过四层 hMLP;轻量嵌入后与文本 token 联合处理 训练:45T token 预训练 + 30M rollouts 大规模异步 RLMuon(大矩阵)+ Adam 混合优化;权重衰减 ∝ 学习率²;NVIDIA GB300 NVL72 训练 MoE layers: 256 routed + 2 shared experts, 6 routed active per tokenSigmoid router with auxiliary-loss-free load balancing; selected + shared experts jointly normalized Attention: sliding-window : global = 5 : 1, 8 KV headsRelative positional embeddings (not RoPE) — better long-sequence extrapolation, per the post; short convs at KV projections and residual outputs Multimodality: encoder-free designAudio as dMel spectrograms; images as 40×40 patches through a 4-layer hMLP; light embedding, processed jointly with text tokens Training: 45T-token pretraining + 30M-rollout async RLMuon (large matrices) + Adam hybrid; weight decay ∝ LR²; trained on NVIDIA GB300 NVL72
图 2 · 架构四层速览。全部细节出自官方发布长文;departures 的动机均为「效率 + 长上下文」。
Fig. 2 · The architecture in four layers. All details from the official launch post; every departure is motivated by efficiency and long context.

训练侧还有两个值得记住的细节:① 后训练冷启动用了包括 Kimi K2.5 在内的开放模型生成的合成数据做初始 SFT(占算力小头),大头是在合成与人工环境上的大规模 RL;② RL 过程中出现了一个未被奖励直接鼓励的涌现现象——思维链随训练越来越「电报化」,丢掉冠词和连接词但保持可读、答案不变,纯粹由效率压力驱动(Cognition 训练 SWE-1.7 时也报告过同样现象)。

Two training details worth remembering: ① post-training was bootstrapped with an initial SFT on synthetic data generated by open models including Kimi K2.5 (a small fraction of compute), with the bulk spent on large-scale RL over synthetic and human-created environments; ② an emergent, unrewarded phenomenon appeared during RL — the chain of thought grew telegraphic over training, dropping articles and connectives while staying comprehensible with unchanged answers, driven purely by efficiency pressure (Cognition reported the same while training SWE-1.7).

4🎚️ 能力与思考档位🎚️ Capabilities & the Effort Dial

基准面上,Inkling 的姿态是「宽而平衡」:HLE(带工具)46.0%、GPQA Diamond 87.2%、SWE-Bench Verified 77.6%、MCP Atlas 76.0%,多数项落在开放权重第一梯队之内、封闭旗舰之下;Design Arena 的盲测人评里排开放模型前列。安全上,FORTRESS 对抗档 78.0% 为其对比的开放模型中最高,且良性误拒率低(Benign 95.9%)。

On benchmarks, Inkling's posture is "broad and balanced": HLE with tools 46.0%, GPQA Diamond 87.2%, SWE-Bench Verified 77.6%, MCP Atlas 76.0% — mostly within the open-weights first tier, below closed flagships; among the top open models in Design Arena's blinded human evals. On safety, its FORTRESS adversarial score of 78.0% is the highest among the open models it compared, with low benign over-refusal (95.9%).

可控思考档位(effort 0.2 – 0.99)
Controllable thinking effort (0.2 – 0.99)
是什么一个推理时旋钮:调高多想、调低省钱,harness 内即可设置。训练方式是在不同样本上通过系统消息+逐 token 成本设定不同档位,让模型「学会」控制思考量。
WhatAn inference-time knob: turn it up to think more, down to save money, settable from within the harness. It was trained by assigning different effort levels via system messages and per-token costs so the model learned to control its own thinking budget.
为什么重要官方展示了完整的「效果 × token 消耗」曲线,而非单点分数:在 Terminal Bench 2.1 上,Inkling 用约三分之一的 token 达到 Nemotron 3 Ultra 的同等成绩。对要跑百万次的生产负载,看整条成本曲线选工作点,比看单点最高分更相关。
Why it mattersThe post shows full performance-vs-token curves rather than single points: on Terminal Bench 2.1, Inkling matches Nemotron 3 Ultra at roughly a third of the tokens. For workloads run millions of times, choosing an operating point on the full cost curve beats comparing single max scores.
呼应「测试时计算可购买」正在成为行业通用语:OpenAI 十证明约 $2000 的 token 账单(见 08-03 文档)、Cursor 的 planner/worker 分账(见 07-31 文档)、Inkling 的档位,是同一逻辑的三种落地。
Echoes"Purchasable test-time compute" is becoming the industry's lingua franca: OpenAI's ~$2,000 token bill for ten proofs (see the 08-03 doc), Cursor's planner/worker cost split (07-31 doc), and Inkling's dial are three implementations of one logic.

5🐜 Inkling Small:三分之一参数的反超🐜 Inkling Small: The Third-Size Overtake

7 月 31 日正式发布的 Inkling Small 把「效率而非规模」推到台前:276B 总参数、12B 激活(旗舰为 975B/41B),用改进后的预训练数据与配方 + 同一套后训练栈,在多个基准上追平甚至反超旗舰。Artificial Analysis 智能指数 40(旗舰 41),并称同尺寸及以下的开放模型无一更高

Officially released July 31, Inkling Small pushes "efficiency over size" to the front: 276B total, 12B active (vs. the flagship's 975B/41B), using improved pretraining data and recipe plus the same post-training stack to match or beat its big sibling on several benchmarks. Artificial Analysis scores it 40 on the Intelligence Index (flagship: 41) and says no open model of equal or smaller size scores higher.

40 vs 41
AA 智能指数:Small 仅比旗舰低 1 分,激活参数不到 1/3
AA Intelligence Index: Small trails the flagship by one point with <1/3 the active parameters
32% / 89%
HLE 与 GPQA Diamond 上反超旗舰(30% / 87%);SWE-Bench Verified 几乎持平(77.4 vs 77.6)
Beats the flagship on HLE and GPQA Diamond (vs. 30% / 87%); near-parity on SWE-Bench Verified (77.4 vs. 77.6)
24K
平均每任务输出 token(DeepSeek V4 Flash 45K、GPT-5.4 mini 78K)——省的是真金白银
Average output tokens per task (DeepSeek V4 Flash: 45K; GPT-5.4 mini: 78K) — real money saved

短板同样明确:Agent 类任务与事实性知识明显落后——Tau 3 Banking 13.6%(旗舰 23.7%)、SimpleQA Verified 仅 20.9%(旗舰 43.9%)。官方给它的定位是成本/延迟敏感的工作负载:编码、当评分器(LLM-as-grader)、生成合成数据。256K 上下文、文本/图像/语音输入、Apache 2.0,可在 Tinker Playground 浏览器内直接微调。

The weaknesses are equally clear: it lags on agentic tasks and factual knowledge — Tau 3 Banking 13.6% (flagship 23.7%), SimpleQA Verified just 20.9% (flagship 43.9%). The official framing: cost/latency-sensitive workloads — coding, LLM-as-grader, synthetic data generation. 256K context, text/image/speech input, Apache 2.0, fine-tunable in the browser via Tinker Playground.

6🧰 怎么用与生态位🧰 How to Use It & the Ecosystem

1
先上手感受:Tinker Playground Get a feel first: Tinker Playground Tinker 控制台内置聊天界面(带 agentic 搜索,限时免费),官方明说选基座模型是「基准数字 + 手感」的综合判断——先玩再决定。 The Tinker console ships a chat interface (with agentic search, free for a limited time); the post itself says picking a base model combines benchmarks with feel — play before committing.
2
托管调用:五家 Day-0 伙伴 Hosted access: five day-0 partners Together AI、Fireworks、Modal、Databricks、Baseten 提供 API;自托管拉 Hugging Face 权重,Blackwell 硬件可用 NVFP4 检查点;推理栈覆盖 SGLang、vLLM、llama.cpp、transformers。 APIs on Together AI, Fireworks, Modal, Databricks, and Baseten; self-hosters pull Hugging Face weights, with an NVFP4 checkpoint for Blackwell; inference support spans SGLang, vLLM, llama.cpp, and transformers.
3
微调落地:cookbook + tml-renderer Fine-tuning in practice: cookbook + tml-renderer Tinker 上提供 64K/256K 两档上下文与限时五折;官方 cookbook 新增三个音频能力配方;tml-renderer 处理工具调用/推理内容/多模态输入的采样与后训练。微调完的检查点可直接部署到上述伙伴平台。 Tinker offers 64K/256K context tiers at a limited-time 50% discount; the cookbook adds three audio recipes; tml-renderer handles sampling and post-training with tool calls, reasoning content, and multimodal inputs. Fine-tuned checkpoints deploy straight to the partners above.

7⚖️ 选型对照与常见坑⚖️ Choosing It — and the Pitfalls

你在选什么选 Inkling / Small 的理由选别家的理由 What you're choosingReasons to pick Inkling / SmallReasons to pick others
开放多模态基座做微调从零训练的原生多模态 + Apache 2.0 + Tinker 一条龙;音频能力在开放阵营名列前茅(VoiceBench 91.4%)纯文本任务上 Kimi K2.6 / GLM 5.2 多项基准更高 An open multimodal base for fine-tuningNatively multimodal from scratch + Apache 2.0 + the Tinker pipeline; top-tier open audio (VoiceBench 91.4%)On text-only tasks, Kimi K2.6 / GLM 5.2 score higher on several benchmarks
高吞吐生产负载effort 档位 + Small 的 24K 平均输出,成本曲线优势明显若需要强事实性(SimpleQA)或复杂 Agent 流程,旗舰封闭模型仍有明显差距优势 High-throughput production workloadsThe effort dial + Small's 24K average output give a clear cost-curve edgeFor strong factuality (SimpleQA) or complex agentic flows, closed flagships keep a sizable lead
合规与可控性Apache 2.0 商用友好;FORTRESS 对抗档开放阵营最高;反审查训练明确需要厂商托管的企业合规栈(审计、SLA)时,封闭 API 更省事 Compliance & controlApache 2.0 is enterprise-friendly; highest open-model FORTRESS adversarial score; explicit censorship-resistance trainingIf you need vendor-managed compliance (audits, SLAs), closed APIs are less work
⚠️ 四个常见坑
⚠️ Four pitfalls

别拿 Small 当知识库:SimpleQA 20.9% 意味着裸问事实极易出错——接检索再用。② Agent 流程先试后上:Tau 3 Banking 13.6% 提示复杂多步工具流程是短板。③ 基准脚注要读:官方 Terminal Bench 成绩剔除了少量「从网络搜索污染」的解;各家 harness 不同,横向对比有水分。④ 时效:Small 于 7 月 31 日刚发布、评分主要来自 Artificial Analysis 单一评测方,更多独立复测还在路上;价格与折扣(Tinker 五折)均为限时信息。

Don't use Small as a knowledge base: 20.9% SimpleQA means bare factual queries fail often — pair it with retrieval. ② Pilot agentic flows before production: 13.6% on Tau 3 Banking flags complex multi-step tool use as the weak spot. ③ Read benchmark footnotes: official Terminal Bench scores zero out solutions contaminated by web search, and harnesses differ across labs. ④ Freshness: Small shipped July 31; scores lean on a single evaluator (Artificial Analysis) with independent replications pending; pricing and the Tinker discount are limited-time.

8📚 学习资源清单📚 Learning Resources

一句话带走

The One-Line Takeaway

Inkling 不是要赢今天的榜,而是要押中明天的分发方式:开放权重 × 原生多模态 × 可购买的思考量 × 微调即产品。Small 用「三分之一参数打平旗舰」证明了配方在进步——但事实性与 Agent 短板提醒你:它是等你雕刻的基座,不是开箱即用的全才。

Inkling isn't built to win today's leaderboard but to bet on tomorrow's distribution: open weights × native multimodality × purchasable thinking × fine-tuning as the product. Small's "third the size, same intelligence" shows the recipe improving — while its factuality and agentic gaps remind you it's a base awaiting your carving, not an out-of-the-box generalist.