跳到主内容
@wquguru
精选88Hacker News Best(web_list)模型发布/更新多源精选 ×4

Thinking Machines AI发布Inkling 2 Mini开源模型

Inkling 2 Mini

原文
发到 X
推荐理由

Inkling 2 Mini 展示了循环 Transformer 架构在小参数模型上的推理优势,且完全开源可本地部署,对 Agent 开发有参考价值。

Try on TinkerHugging FaceModel card

在 Hugging Face 上试用 Tinker 模型卡片

Today we are releasing Inkling 2 Mini, a fast, powerful model that’s small enough to run on your laptop.

今天,我们发布 Inkling 2 Mini,这是一款快速、强大的模型,其体积小巧,足以在您的笔记本电脑上运行。

Inkling 2 Mini is a looped transformer, running the middle layers of the model three times for every token. This allows the model to apply more reasoning and reach the right answer with fewer tokens, especially on tasks such as agentic coding that depend on reasoning quality. Inkling 2 Mini leads every open-weights model with less than 100B total parameters on benchmarks such as DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%). It runs faster than similarly performing models as well.

Inkling 2 Mini 是一种循环 Transformer 模型,对于每个 token,它会运行模型中间层三次。这使得模型能够进行更多的推理,并以更少的 token 得出正确答案,特别是在依赖推理质量的任务(如代理编码)中表现更佳。在 DeepSWE v1.1 (48.2%) 和 Vals Index v2.2 (46.5%) 等基准测试中,Inkling 2 Mini 在所有总参数量小于 100B 的开放权重模型中名列前茅。它的运行速度也比性能相似的模型更快。

DeepSWE v1.1

Tasks resolved (%)

解决任务比例 (%)

0%20%40%60%5102050100Tasks resolved (%)Theoretical tokens per second (log scale)Ripple-65B (Preview)Ripple-65B (Preview)Gemma 4 31BMuse GlimmerMuse GlimmerQwen3.8-27BInkling 2 MiniInkling 2 MiniGemma 4 26B A4BGemma 4 26B A4BRecursive FlintRecursive Flint

解决任务比例 (%) 每秒理论 token 数(对数刻度)Ripple-65B (预览) Ripple-65B (预览) Gemma 4 31B Muse Glimmer Muse Glimmer Qwen3.8-27B Inkling 2 Mini Inkling 2 Mini Gemma 4 26B A4B Gemma 4 26B A4B Recursive Flint Recursive Flint

Tasks resolved (%)

解决任务比例 (%)

0%20%40%60%5102050100Tasks resolved (%)Theoretical tokens per second (log scale)Ripple-65B (Preview)Ripple-65B (Preview)Gemma 4 31BMuse GlimmerMuse GlimmerQwen3.8-27BInkling 2 MiniInkling 2 MiniGemma 4 26B A4BGemma 4 26B A4BRecursive FlintRecursive Flint

解决任务比例 (%) 每秒理论 token 数(对数刻度)Ripple-65B (预览) Ripple-65B (预览) Gemma 4 31B Muse Glimmer Muse Glimmer Qwen3.8-27B Inkling 2 Mini Inkling 2 Mini Gemma 4 26B A4B Gemma 4 26B A4B Recursive Flint Recursive Flint

SWE-bench Pro v2 Hard (51 tasks)

SWE-bench Pro v2 Hard (51 个任务)

Tasks resolved (%)

解决任务比例 (%)

0%20%40%60%80%100%5102050100Tasks resolved (%)Theoretical tokens per second (log scale)Ripple-65B (Preview)Ripple-65B (Preview)Gemma 4 31BMuse GlimmerQwen3.8-27BQwen3.6-27BInkling 2 MiniInkling 2 MiniGemma 4 26B A4BGemma 4 26B A4BQwen3.6-35B-A3BQwen3.6-35B-A3B

解决任务比例 (%) 每秒理论 token 数(对数刻度)Ripple-65B (预览) Ripple-65B (预览) Gemma 4 31B Muse Glimmer Qwen3.8-27B Qwen3.6-27B Inkling 2 Mini Inkling 2 Mini Gemma 4 26B A4B Gemma 4 26B A4B Qwen3.6-35B-A3B Qwen3.6-35B-A3B

Tasks resolved (%)

解决任务比例 (%)

0%20%40%60%80%100%5102050100Tasks resolved (%)Theoretical tokens per second (log scale)Ripple-65B (Preview)Ripple-65B (Preview)Gemma 4 31BMuse GlimmerQwen3.8-27BQwen3.6-27BInkling 2 MiniInkling 2 MiniGemma 4 26B A4BGemma 4 26B A4BQwen3.6-35B-A3BQwen3.6-35B-A3B

解决任务比例 (%) 每秒理论 token 数(对数刻度)Ripple-65B (预览) Ripple-65B (预览) Gemma 4 31B Muse Glimmer Qwen3.8-27B Qwen3.6-27B Inkling 2 Mini Inkling 2 Mini Gemma 4 26B A4B Gemma 4 26B A4B Qwen3.6-35B-A3B Qwen3.6-35B-A3B

Reported coding scores for models under 100B total parameters. The green line is the frontier with Inkling 2 Mini, the dashed line the frontier without it, and the shading the ground it adds. Decoding speed is a theoretical ceiling at 614 GB/s with BF16 weights and batch size one, including repeated reads across loops. Evaluation setups vary by model.

总参数量低于 100B 的模型的报告编码分数。绿线代表包含 Inkling 2 Mini 的前沿水平,虚线代表不包含它的前沿水平,阴影部分表示它所提升的范围。解码速度是在 BF16 权重和批量大小为 1 的情况下,包括跨循环的重复读取的理论上限,为 614 GB/s。评估设置因模型而异。

Inkling 2 Mini is a Mixture-of-Experts model with 26.4B total parameters, 3.4B active, and 5.4B weight activations per token to account for looping. Looping allows the model to reason longer without writing its reasoning down, which matters for monitoring: our ability to oversee and understand what the model does by reading its chain of thought (CoT).

Inkling 2 Mini 是一个混合专家(MoE)模型,总参数量为 26.4B,活跃参数为 3.4B,每个 token 的权重激活量为 5.4B(以考虑循环因素)。循环允许模型在不将推理过程写下来的情况下进行更长时间的推理,这对监控至关重要:即通过阅读其思维链(CoT),我们能够监督和了解模型的行为。

We measured the impact of looping on both performance and monitorability with Inkling 2 Mini and a research-specific variable-loop variant we are releasing alongside it. The variable-loop model’s no-CoT time horizon improved from 16 seconds at one pass to 37 seconds with three, and its performance on a benchmark of serial reasoning kept improving out to ten loops—twice as many as it was trained with. The model’s chain of thought grew shorter with more loops but stayed legible throughout. The model didn’t trend higher on tests of agentic sabotage as loops were added.

我们使用 Inkling 2 Mini 以及与其同步发布的研究专用可变循环变体,测量了循环对性能和可监控性的影响。该可变循环模型在无需思维链(CoT)的情况下,其时间范围从一次通过的 16 秒提升至三次通过时的 37 秒;其在串行推理基准测试上的表现随着循环次数增加而持续提升,直至十次循环——这比其训练时所采用的循环次数多了一倍。随着循环次数增加,模型的思维链变得更短,但始终保持清晰可读。随着循环的加入,模型在代理破坏性测试中的得分并未呈现上升趋势。

We’re excited to offer some of the first models with looped architecture available to researchers. The variable-loop model was used for our monitorability experiments as it allows for cleaner comparisons between different looping depths.

我们很高兴提供首批支持循环架构、可供研究人员使用的模型之一。可变循环模型被用于我们的可监控性实验,因为它允许在不同循环深度之间进行更清晰的比较。

Most of Inkling 2 Mini’s weights were trained and are deployed natively in NVFP4, with the last four layers in MXFP8. The total memory footprint of Inkling 2 Mini is ~20GB, allowing it to run on local devices such as consumer laptops and desktop AI computers. Our researchers took Inkling 2 Mini on a flight, where it ran offline on a MacBook Pro and built a playable flight simulator in a few minutes.

Inkling 2 Mini 的大部分权重采用 NVFP4 格式进行训练并以原生方式部署,最后四层采用 MXFP8 格式。Inkling 2 Mini 的总内存占用约为 20GB,使其能够在消费级笔记本电脑和桌面 AI 计算机等本地设备上运行。我们的研究人员带着 Inkling 2 Mini 乘坐飞机,它在 MacBook Pro 上离线运行,并在几分钟内构建了一个可玩的飞行模拟器。

DemoArtifact

演示工件

MacBook Pro (M5 Pro, 48GB) Airplane mode

MacBook Pro (M5 Pro, 48GB) 飞行模式

Open video: PlaneBench flight simulator demo, screen with phone video inset

打开视频:PlaneBench 飞行模拟器演示,屏幕附带手机视频画中画

0:0015×

Loading the flight simulator Inkling 2 Mini built…To try the flight simulator, open this page on desktop.

正在加载 Inkling 2 Mini 构建的飞行模拟器……要体验该飞行模拟器,请在桌面上打开此页面。

Building and flying a flight simulator with Inkling 2 Mini. The timer shows real time; most of the recording plays at 15× speed.

使用 Inkling 2 Mini 构建并驾驶飞行模拟器。计时器显示的是实时时间;大部分录制内容以 15 倍速播放。

Inkling 2 Mini is released with open weights and is available on Tinker today for inference and customization.

Inkling 2 Mini 已开源权重,今天可在 Tinker 上获取,用于推理和定制。

Capabilities

能力

Inkling 2 Mini’s strengths mirror those of Inkling 2 Small: it does best on long-horizon agentic work and is the best among open-weights models of similar size on most tasks we measured.

Inkling 2 Mini 的优势与 Inkling 2 Small 相似:它在长周期代理任务中表现最佳,在我们测量的大多数任务中,它是同类大小开源权重模型中的佼佼者。

Inkling 2 Mini scores ahead of every comparable model on agentic coding benchmarks such as DeepSWE v1.1 and SWE-bench Pro v2, and is behind only Qwen3.8-27B on others. On PMPP-Hard, a kernel-optimization benchmark, it approaches the score of the much larger Inkling 2 Small.

在 DeepSWE v1.1 和 SWE-bench Pro v2 等代理编码基准测试中,Inkling 2 Mini 的得分高于所有可比模型,在其他测试中仅落后于 Qwen3.8-27B。在 PMPP-Hard(一种内核优化基准测试)中,它接近规模大得多的 Inkling 2 Small 的得分。

While looping improves the model’s reasoning, it doesn’t enable its weights to store more facts—knowledge is mostly a function of total parameter count. Inkling 2 Mini’s score on AA-Omniscience, a measure of factuality, is in line with other models of the same size. Studying the variable-loop model, we saw that performances on long-context benchmarks improve with more passes, such as GraphWalks BFS and MRCR v2; but knowledge-intensive benchmarks doesn’t improve with more passes, such as AA-Omniscience, SimpleQA Verified, SimpleVQA, and WorldVQA.

虽然循环推理提升了模型的推理能力,但它并不能让模型的权重存储更多事实——知识主要取决于总参数量。Inkling 2 Mini 在 AA-Omniscience(一项衡量事实准确性的指标)上的得分与同规模的其他模型持平。在对可变循环模型的研究中,我们发现其在长上下文基准测试中的表现会随着迭代次数增加而提升,例如 GraphWalks BFS 和 MRCR v2;但在知识密集型基准测试中,如 AA-Omniscience、SimpleQA Verified、SimpleVQA 和 WorldVQA,性能并不会因增加迭代次数而改善。

Inkling 2 Mini overcomes the knowledge gap by efficiently using search and tools. It leads open-weights models of similar size on BrowseComp, DeepSearchQA, and MCP Atlas. It is also strong on knowledge work benchmarks that incorporate search and tool use.

Inkling 2 Mini 通过高效使用搜索和工具来弥补知识差距。它在 BrowseComp、DeepSearchQA 和 MCP Atlas 上领先于其他同规模的开源权重模型。它在结合搜索和工具使用的知识工作基准测试中也表现出色。

更进一步:量化金融体系

看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力

进入量化体系 →