Skip to content
全部文章

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。


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

mini-swe-agent v2.4.6. Source: Thinking Machines launch post.

Inkling 2 Mini48.2%
Qwen3.8-27B42.2%
Nemotron 3.5 Super VL24.7%
Muse Glimmer4.4%
Gemma 4 31B0.3%
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

Thinking Machines' Inkling 2 Mini against the four models its launch post compares it with, across coding, agentic search, vision, reasoning and safety, plus the independent evals our catalog holds.

Vendor benchmarksInkling 2 MiniGemma 4 31BMuse Glimmer 30BQwen3.8-27BNemotron 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) / total5.4/26.430.7/30.7 (Highest reported score in this row)30.0/30.0 (Second-highest reported score in this row)27.0/27.012.0/124.5
Third party index
AA Index (v5.0)20.1 (Highest reported score in this row)4.46.320.1 (Highest reported score in this row)13.6 (Second-highest reported score in this row)
Vals Index (v2.2)46.5% (Highest reported score in this row)–––29.8% (Second-highest reported score in this row)
Coding and autoresearch
FrontierCode v2.0stirrup v0.2.115.1% (Highest reported score in this row)0.0%1.5%14.9% (Second-highest reported score in this row)6.1%
DeepSWE v1.1mini-swe-agent v2.4.648.2% (Highest reported score in this row)0.3%4.4%42.2% (Second-highest reported score in this row)24.7%
Terminal Bench v4.0mini-swe-agent v2.4.65.6% (Highest reported score in this row)0.0%0.5% (Second-highest reported score in this row)5.6% (Highest reported score in this row)0.0%
Terminal Bench v2.1Inkling 2 Mini in pi v0.83.071.6% (Second-highest reported score in this row)43.4%51.7%73.0% (Highest reported score in this row)60.7%
SWEBench Pro v2 (Hard)mini-swe-agent v2.4.677.9% (Highest reported score in this row)14.2%44.3%73.8% (Second-highest reported score in this row)44.1%
SWEBench MultilingualInkling 2 Mini with bash only73.7% (Highest reported score in this row)43.0%66.9%70.0% (Second-highest reported score in this row)67.7%
SWEBench Verifiedbash only79.2% (Highest reported score in this row)46.1%76.0%77.8% (Second-highest reported score in this row)71.1%
Vibe Code Bench v1.1openhands (vals)77.9% (Highest reported score in this row)––64.9% (Second-highest reported score in this row)3.5%
PMPP-Hard (AUC, 0-10h)pi v0.83.0455.8 (Highest reported score in this row)––262.5 (Second-highest reported score in this row)–
Agentic search, tool use and knowledge work
BrowseCompInkling 2 Mini with compaction79.0% (Highest reported score in this row)37.0%63.9%61.6%78.9% (Second-highest reported score in this row)
DeepSearchQAInkling 2 Mini with compaction85.3% (Highest reported score in this row)61.7%74.6%79.7% (Second-highest reported score in this row)77.7%
MCP AtlasPublic82.7% (Highest reported score in this row)54.2%75.5%77.3% (Second-highest reported score in this row)67.7%
AA-Briefcase v1.11263.0 (Second-highest reported score in this row)369.0478.01391.0 (Highest reported score in this row)999.0
GDPVal-AA v2.11319.0 (Second-highest reported score in this row)627.0792.01427.0 (Highest reported score in this row)1162.0
JobBench49.6% (Highest reported score in this row)8.2%16.1%33.4%36.5% (Second-highest reported score in this row)
AutomationBench-AA v1.0.641.8% (Second-highest reported score in this row)4.9%6.8%48.2% (Highest reported score in this row)14.8%
EnterpriseOps-Gym42.0% (Second-highest reported score in this row)28.4%34.7%44.2% (Highest reported score in this row)28.1%
Vals Finance Index50.1% (Highest reported score in this row)–––41.0% (Second-highest reported score in this row)
AA-AnalystAgent46.3% (Highest reported score in this row)–20.0% (Second-highest reported score in this row)––
Diligence Benchbest harness48.6% (Highest reported score in this row)–40.9%46.4% (Second-highest reported score in this row)45.0%
Vision interpretation and creation
GDP.pdf16.8% (Highest reported score in this row)6.0%11.8%16.4% (Second-highest reported score in this row)8.0%
PerceptionBench57.5% (Highest reported score in this row)40.7%42.8%52.4% (Second-highest reported score in this row)36.8%
MMMU Prostandard 1074.9% (Second-highest reported score in this row)73.4%74.3%76.3% (Highest reported score in this row)72.3%
OmniDocBench v1.692.8% (Highest reported score in this row)82.5%90.4%91.5% (Second-highest reported score in this row)77.9%
ScreenSpot Prooriginal / with Python76.5/87.7% (Highest reported score in this row)5.9%/–75.4%/– (Second-highest reported score in this row)57.6/78.9%43.5%/–
Chartographyoriginal / with Python22.1/72.5% (Highest reported score in this row)12.0/11.8%17.4/25.1% (Second-highest reported score in this row)12.2/52.1%9.7/29.3%
ZeroBench (main)original / with Python17.0/34.0% (Highest reported score in this row)9.0/12.0%9.0/17.0%12.0%/– (Second-highest reported score in this row)11.0/12.0%
MuirBench77.5% (Second-highest reported score in this row)72.3%77.6% (Highest reported score in this row)65.9%72.7%
BVB68.4% (Second-highest reported score in this row)47.0%50.5%70.7% (Highest reported score in this row)52.1%
Reasoning, knowledge and epistemics
HLE-Diamond15.2% (Second-highest reported score in this row)7.2%9.1%15.3% (Highest reported score in this row)8.8%
CritPt v2full / hard31.4/17.0% (Second-highest reported score in this row)–/4.0%–/7.0%10.6/4.0%36.9/13.5% (Highest reported score in this row)
Apex 202534.4% (Highest reported score in this row)5.2%4.2%18.8%22.9% (Second-highest reported score in this row)
ARC-AGIv1 / v284.5/49.3% (Second-highest reported score in this row)––87.5/42.4% (Highest reported score in this row)22.7/1.5%
LongBench v266.9% (Highest reported score in this row)52.9%33.1%52.3%62.1% (Second-highest reported score in this row)
GraphWalks BFS256K69.7% (Highest reported score in this row)––50.4%63.3% (Second-highest reported score in this row)
MRCR v28-needle 256K-512K46.4% (Highest reported score in this row)––38.0% (Second-highest reported score in this row)5.7%
AA-Omniscience v1.1score-8.6-48.5-32.2-6.4 (Second-highest reported score in this row)-2.9 (Highest reported score in this row)
AA-Omniscience v1.1non-hallucination rate63.4%14.5%16.4%71.4% (Highest reported score in this row)69.5% (Second-highest reported score in this row)
ComplexConstraints v1.1openhands v1.45.033.4% (Second-highest reported score in this row)31.6%35.6% (Highest reported score in this row)30.1%30.1%
Safety and alignment
FORTRESSadversarial public89.5% (Second-highest reported score in this row)67.8%90.3% (Highest reported score in this row)78.8%79.6%
FORTRESSbenign public96.4%97.6% (Second-highest reported score in this row)89.1%97.8% (Highest reported score in this row)92.7%
StrongREJECTnone97.8%98.1%99.0% (Second-highest reported score in this row)99.0% (Second-highest reported score in this row)99.4% (Highest reported score in this row)
Misbehavior (internal)lower is better9.8% (Highest reported score in this row)27.6% (Second-highest reported score in this row)46.4%57.7%52.8%
Covert misbehavior (internal)lower is better8.3% (Highest reported score in this row)24.3% (Second-highest reported score in this row)45.6%33.2%43.4%
Destructive actions (internal)lower is better10.2% (Highest reported score in this row)75.8%68.8%35.2% (Second-highest reported score in this row)65.6%
Independent evalsInkling 2 MiniGemma 4 31BMuse Glimmer 30BQwen3.8-27BNemotron 3.5 Super VL 120B A12B
GPQA DiamondArtificial Analysis–76.3%–––
ARC-AGI-2Epoch AI–––42.4%–
CursorBench–––––
SciCodeArtificial Analysis, Epoch AI–41.1% (Second-highest reported score in this row)–46.6% (Highest reported score in this row)–
SimpleQA (verified)–––––

Sources: Introducing Inkling 2 Mini (Thinking Machines Lab) (vendor-reported, as of 2026-11-21); catalog benchmarks from Artificial Analysis, Epoch AI. A dash means the source reports no value.

Benchmark data from Epoch AI (CC-BY 4.0).

Updated 2026-11-21.

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 positionCompute-equivalent gainLayer span
Middle layers1.51–1.91×Blocks 2 and 3, layers 6–18
Last layers1.24–1.36×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。

AI SDK
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)
Coding agent
anyr claude --model "thinkingmachines/inkling-2-mini"

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。

两分钟发第一笔请求

用自己的 key 免费开始,或充值按 token 计费。Go 每月 $4 额度 — $2/月,或捐一把 provider key 免费开通。