Skip to content

Inkling 2 Small

Thinking Machines' second-generation open-weights model joins the catalog, 594B MoE, 17B active, vision input, up to 1M context.

Open Inkling 2 Small
Cover art for the Inkling 2 Small announcement

thinkingmachines/inkling-2-small is Thinking Machines Lab's second-generation foundation model. It is a 594B-parameter mixture-of-experts model with 17B active parameters, a 1M-token context window, switchable reasoning effort and tool use, and image input (no audio) on top of text. On the Artificial Analysis Intelligence Index it scores 31, a large jump over previous generations, led by agentic coding, vision and knowledge work. Weights are not open yet: they sit under the Tinker Terms of Use with a full Apache 2.0 release pledged after safety testing.

It is BYOK-only on Tinker's OpenAI-compatible API. Create a Tinker key and add it under Dashboard → BYOK. Tinker lists it at $0.30 per 1M input tokens and $1.00 per 1M output tokens and bills you directly; AnyRouter charges nothing for a BYOK route.

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-small"),
  prompt: "Draw a diagram of how a moat bridges a castle.",
})
console.log(text)
Run it in the terminal with your coding agent
anyr claude --model "thinkingmachines/inkling-2-small"

Call it with your own key. AnyRouter adds no per-token charge.

Inkling 2 Small vs open-weight peers

Thinking Machines' Inkling 2 Small against the four models its launch post compares it with, on agentic coding and autoresearch, plus the independent evals our catalog holds.

Coding and AutoresearchInkling 2 SmallGLM-5.3-FlashDeepSeek-V4.1-Flash-1017Muse Spark OpenKimi K3
FrontierCode v2.0stirrup v0.2.125.4% (Second-highest reported score in this row)17.6%28.5% (Highest reported score in this row)7.6%17.3%
DeepSWE v1.1mini-swe-agent v2.4.672.1% (Second-highest reported score in this row)63.4%77.0% (Highest reported score in this row)60.0%68.5%
SWE-Atlas v1.1mini-swe-agent v2.4.661.9% (Second-highest reported score in this row)–62.1% (Highest reported score in this row)42.3%61.0%
SWEBench Pro v2 (Hard/Full)mini-swe-agent v2.4.686.3/97.7%82.8%/–89.7%/– (Highest reported score in this row)82.4/95.0%88.2/97.7% (Second-highest reported score in this row)
Vibe Code Bench v1.1openhands (vals)81.6%30.8%87.7% (Highest reported score in this row)76.2%85.0% (Second-highest reported score in this row)
PostTrainBench v1.2Inkling 2 Small in Claude Code31.9% (Second-highest reported score in this row)36.8% (Highest reported score in this row)30.5%–31.5%
Terminal Bench Science v0.1mini-swe-agent v2.4.65.7%4.8%12.9% (Highest reported score in this row)1.9%8.1% (Second-highest reported score in this row)
Preliminary kernel autoresearch results

DeepSeek-V3.2 sparse attention, forward + backward. Best of four sessions on one B300 GPU.

Claude Opus 5.5 · 18.80×GPT-6 Astra · 10.26×Inkling 2 Small · 6.02×DeepSeek V4.1 Flash · 5.66×GLM 5.3 · 5.03×Mistral Large 4 · 3.43×0510152002.557.510Hours into the sessionSpeedup over originalClaude Opus 5.5: 18.80×GPT-6 Astra: 10.26×Inkling 2 Small: 6.02×DeepSeek V4.1 Flash: 5.66×GLM 5.3: 5.03×Mistral Large 4: 3.43×

Final speedups verified on hidden tests. The source shows each session's full step curve; it publishes only these endpoints as numbers. Source: Introducing Inkling 2 Small (Thinking Machines Lab), as of 2026-11-18.

Sources: Introducing Inkling 2 Small (Thinking Machines Lab) (vendor-reported, as of 2026-11-18); 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 2026-11-18.

Playbooks: AI SDK, OpenAI SDK, coding agents, playground.

Terminal replay scoring panel of an agentic task run
Agentic tool-use replay from the launch demonstration. Source: Thinking Machines.
Three glasses of red wine on a restaurant table in order
Wine-tasting Three.js scene generated by Inkling 2 Small. Source: Thinking Machines.

Looking for the full catalog? Browse all models.