# Inkling 2 Mini：可在笔记本上运行的循环Transformer

> Thinking Machines 的新开源权重模型对每个 token 将中间层运行三次。它在 100B 以下开源权重模型的智能体编程上领先，仅需约 20GB，并已通过 Tinker 与 OpenRouter BYOK 上线 AnyRouter。

- Date: November 21, 2026
- Category: product
- HTML: https://anyrouter.dev/blog/zh/inkling-2-mini

## 一个每个 token 思考更多次的小模型

Inkling 2 Mini 是一个混合专家模型，总参数 26.4B，激活参数 3.4B。它的特别之处在于「循环」：网络中间层对每个 token 运行三次，让模型在不增加参数的情况下完成更多串行计算。Thinking Machines 表示，它在 100B 以下开源权重模型中领先于智能体编程任务，DeepSWE v1.1 得分 48.2%，Vals Index v2.2 得分 46.5%。

它以 NVFP4 原生运行，整体占用约 20GB，足以在消费级笔记本上运行。实验室曾在飞行模式下用 M5 Pro MacBook Pro 离线运行它，几分钟内就构建出一个可玩的飞行模拟器。

**DeepSWE v1.1 — tasks resolved**

- Inkling 2 Mini: 48.2% %
- Qwen3.8-27B: 42.2% %
- Nemotron 3.5 Super VL: 24.7% %
- Muse Glimmer: 4.4% %
- Gemma 4 31B: 0.3% %

_Figure 1: DeepSWE v1.1 scores from the Thinking Machines comparison table._

## 可交互的基准测试表

发布公告将 Inkling 2 Mini 与 Gemma 4 31B、Muse Glimmer、Qwen3.8-27B 和 Nemotron 3.5 Super VL 在编程、智能体搜索、视觉、推理与安全等方面做了对比。我们把公布的每一项数字都整理进一个 compare collection。你可以在表格、Pareto 和饱和度假签之间切换，按模型排序，并分享任意视图的直链。

**Inkling 2 Mini vs open-weight peers**

| Vendor benchmarks | Inkling 2 Mini | Gemma 4 31B | Muse Glimmer 30B | Qwen3.8-27B | Nemotron 3.5 Super VL 120B A12B |
| --- | --- | --- | --- | --- | --- |
| Model info |  |
| Inkling 2 Mini has 3.4B unique active parameters. The vendor's 'activated' column below counts weight activations per token (5.4B), which includes the three loop passes. |  |
| Params (B) (activated per token (incl. loops) / total) | 5.4/26.4 | **30.7/30.7** | 30.0/30.0 | 27.0/27.0 | 12.0/124.5 |
| Third party index |  |
| AA Index (v5.0) | **20.1** | 4.4 | 6.3 | **20.1** | 13.6 |
| Vals Index (v2.2) | **46.5%** | – | – | – | 29.8% |
| Coding and autoresearch |  |
| FrontierCode v2.0 (stirrup v0.2.1) | **15.1%** | 0.0% | 1.5% | 14.9% | 6.1% |
| DeepSWE v1.1 (mini-swe-agent v2.4.6) | **48.2%** | 0.3% | 4.4% | 42.2% | 24.7% |
| Terminal Bench v4.0 (mini-swe-agent v2.4.6) | **5.6%** | 0.0% | 0.5% | **5.6%** | 0.0% |
| Terminal Bench v2.1 (Inkling 2 Mini in pi v0.83.0) | 71.6% | 43.4% | 51.7% | **73.0%** | 60.7% |
| SWEBench Pro v2 (Hard) (mini-swe-agent v2.4.6) | **77.9%** | 14.2% | 44.3% | 73.8% | 44.1% |
| SWEBench Multilingual (Inkling 2 Mini with bash only) | **73.7%** | 43.0% | 66.9% | 70.0% | 67.7% |
| SWEBench Verified (bash only) | **79.2%** | 46.1% | 76.0% | 77.8% | 71.1% |
| Vibe Code Bench v1.1 (openhands (vals)) | **77.9%** | – | – | 64.9% | 3.5% |
| PMPP-Hard (AUC, 0-10h) (pi v0.83.0) | **455.8** | – | – | 262.5 | – |
| Agentic search, tool use and knowledge work |  |
| BrowseComp (Inkling 2 Mini with compaction) | **79.0%** | 37.0% | 63.9% | 61.6% | 78.9% |
| DeepSearchQA (Inkling 2 Mini with compaction) | **85.3%** | 61.7% | 74.6% | 79.7% | 77.7% |
| MCP Atlas (Public) | **82.7%** | 54.2% | 75.5% | 77.3% | 67.7% |
| AA-Briefcase v1.1 | 1263.0 | 369.0 | 478.0 | **1391.0** | 999.0 |
| GDPVal-AA v2.1 | 1319.0 | 627.0 | 792.0 | **1427.0** | 1162.0 |
| JobBench | **49.6%** | 8.2% | 16.1% | 33.4% | 36.5% |
| AutomationBench-AA v1.0.6 | 41.8% | 4.9% | 6.8% | **48.2%** | 14.8% |
| EnterpriseOps-Gym | 42.0% | 28.4% | 34.7% | **44.2%** | 28.1% |
| Vals Finance Index | **50.1%** | – | – | – | 41.0% |
| AA-AnalystAgent | **46.3%** | – | 20.0% | – | – |
| Diligence Bench (best harness) | **48.6%** | – | 40.9% | 46.4% | 45.0% |
| Vision interpretation and creation |  |
| GDP.pdf | **16.8%** | 6.0% | 11.8% | 16.4% | 8.0% |
| PerceptionBench | **57.5%** | 40.7% | 42.8% | 52.4% | 36.8% |
| MMMU Pro (standard 10) | 74.9% | 73.4% | 74.3% | **76.3%** | 72.3% |
| OmniDocBench v1.6 | **92.8%** | 82.5% | 90.4% | 91.5% | 77.9% |
| ScreenSpot Pro (original / with Python) | **76.5/87.7%** | 5.9%/– | 75.4%/– | 57.6/78.9% | 43.5%/– |
| Chartography (original / with Python) | **22.1/72.5%** | 12.0/11.8% | 17.4/25.1% | 12.2/52.1% | 9.7/29.3% |
| ZeroBench (main) (original / with Python) | **17.0/34.0%** | 9.0/12.0% | 9.0/17.0% | 12.0%/– | 11.0/12.0% |
| MuirBench | 77.5% | 72.3% | **77.6%** | 65.9% | 72.7% |
| BVB | 68.4% | 47.0% | 50.5% | **70.7%** | 52.1% |
| Reasoning, knowledge and epistemics |  |
| HLE-Diamond | 15.2% | 7.2% | 9.1% | **15.3%** | 8.8% |
| CritPt v2 (full / hard) | 31.4/17.0% | –/4.0% | –/7.0% | 10.6/4.0% | **36.9/13.5%** |
| Apex 2025 | **34.4%** | 5.2% | 4.2% | 18.8% | 22.9% |
| ARC-AGI (v1 / v2) | 84.5/49.3% | – | – | **87.5/42.4%** | 22.7/1.5% |
| LongBench v2 | **66.9%** | 52.9% | 33.1% | 52.3% | 62.1% |
| GraphWalks BFS (256K) | **69.7%** | – | – | 50.4% | 63.3% |
| MRCR v2 (8-needle 256K-512K) | **46.4%** | – | – | 38.0% | 5.7% |
| AA-Omniscience v1.1 (score) | -8.6 | -48.5 | -32.2 | -6.4 | **-2.9** |
| AA-Omniscience v1.1 (non-hallucination rate) | 63.4% | 14.5% | 16.4% | **71.4%** | 69.5% |
| ComplexConstraints v1.1 (openhands v1.45.0) | 33.4% | 31.6% | **35.6%** | 30.1% | 30.1% |
| Safety and alignment |  |
| FORTRESS (adversarial public) | 89.5% | 67.8% | **90.3%** | 78.8% | 79.6% |
| FORTRESS (benign public) | 96.4% | 97.6% | 89.1% | **97.8%** | 92.7% |
| StrongREJECT (none) | 97.8% | 98.1% | 99.0% | 99.0% | **99.4%** |
| Misbehavior (internal) (lower is better) | **9.8%** | 27.6% | 46.4% | 57.7% | 52.8% |
| Covert misbehavior (internal) (lower is better) | **8.3%** | 24.3% | 45.6% | 33.2% | 43.4% |
| Destructive actions (internal) (lower is better) | **10.2%** | 75.8% | 68.8% | 35.2% | 65.6% |

| Independent evals | Inkling 2 Mini | Gemma 4 31B | Muse Glimmer 30B | Qwen3.8-27B | Nemotron 3.5 Super VL 120B A12B |
| --- | --- | --- | --- | --- | --- |
| GPQA Diamond (Artificial Analysis) | – | 76.3% | – | – | – |
| ARC-AGI-2 (Epoch AI) | – | – | – | 42.4% | – |
| CursorBench | – | – | – | – | – |
| SciCode (Artificial Analysis, Epoch AI) | – | 41.1% | – | **46.6%** | – |
| SimpleQA (verified) | – | – | – | – | – |

_Sources: [Introducing Inkling 2 Mini (Thinking Machines Lab)](https://thinkingmachines.ai/news/inkling-2-mini/) (as of 2026-11-21); [Epoch AI](https://epoch.ai/benchmarks) (CC-BY 4.0). Interactive version: https://anyrouter.dev/compare/inkling-2-mini_

## 循环究竟改变了什么

模型共 24 层，分为四个六层区块。首尾区块各运行一次，中间两个区块运行三次。实验室在 scaling 阶梯上比较了循环位置，发现把循环放在中间比放在末尾带来更大的等效算力收益。

| Loop position | Compute-equivalent gain | Layer span |
| --- | --- | --- |
| Middle layers | 1.51–1.91× | Blocks 2 and 3, layers 6–18 |
| Last layers | 1.24–1.36× | Last 12 layers |

_Figure 2: Compute-equivalent gain by loop position at training step 2,000._

注意力残差设计让每一轮都能读取更早区块的输出，循环消息由门控控制，模型可以学习何时使用它。约 20 层以 NVFP4 部署，最后四层使用 MXFP8，与训练时精度一致，因此没有量化转换损失。

## 在 AnyRouter 上使用

Inkling 2 Mini 以 `thinkingmachines/inkling-2-mini` 提供。两条路线都是 BYOK：在 Dashboard → BYOK 添加 Tinker 或 OpenRouter 密钥，由提供商直接向你收费，AnyRouter 对 BYOK 路线不额外收费。Tinker 列出的 256K 训练 / 1M 采样版本价格为每 100 万输入 token $0.15、每 100 万输出 token $0.50。

```ts
import { createAnyRouter } from "@anyr/ai-sdk-provider"
import { generateText } from "ai"

const anyrouter = createAnyRouter()
const { text } = await generateText({
  model: anyrouter("thinkingmachines/inkling-2-mini"),
  prompt: "Explain the difference between looping and chain of thought.",
})
console.log(text)
```

_AI SDK._

```bash
anyr claude --model "thinkingmachines/inkling-2-mini"
```

_Coding agent._

## 它仍有不足的地方

实验室坦率说明了局限。它并非多轮对话或含大量约束指令场景下的最强小模型，MultiChallenge 得分为 50.0%，ComplexConstraints v1.1 为 33.4%。循环增加推理深度但不增加存储的知识，因此事实性与同规模模型相当；在 CritPt v2 和 HLE-Diamond 等重推理基准上仍落后于更大的模型。在知识类工作上，它更多依赖搜索和工具。
