Inkling Small
Inkling Small is the smaller model in the Inkling family at 276 billion total parameters and 12 billion active. It matches or beats Inkling on many coding, reasoning, and tool-use evaluations, reasons natively over images and audio, and supports a context window of 1M tokens. Your use is subject to Thinkingmachines's Terms & Privacy Policies.
import { streamText } from 'ai'
const result = streamText({ model: 'thinkingmachines/inkling-small', prompt: 'Why is the sky blue?'})About Inkling Small
Inkling Small became available on AI Gateway on July 30, 2026. Inkling Small is a 42-layer decoder-only transformer with a sparse Mixture-of-Experts (MoE) backbone: 276 billion total parameters, 12 billion active per token, and each token routed to 6 of 256 experts plus 2 shared experts. Attention mixes local and global layers, images enter through a hierarchical patch encoder, audio through discrete token encoding, and the context window is 1M tokens. Thinkingmachines released the weights under the Apache 2.0 license.
On agentic coding, Inkling Small scores 80.2% on SWE-bench Verified, 55.9% on SWE-bench Pro (public), 64.7% on Terminal-Bench 2.1, and 48.7% on SciCode. Each of those sits at or above Inkling, which scores 77.6%, 54.3%, 63.8%, and 46.1%. Tool use follows the same pattern: 54.4% on Toolathlon Verified against Inkling's 45.5%, and 79.6% on the public split of MCP Atlas. On reasoning it scores 89.5% on GPQA Diamond, 90.2% on HMMT February 2026, 31.6% on Humanity's Last Exam text-only, and 40.1% on ARC-AGI-2.
The gap is knowledge. Inkling Small scores 20.6% on SimpleQA Verified where Inkling scores 43.9%, and it trails on the AA Omniscience index and on Tau 3 Banking, a multi-turn domain agent evaluation. Audio results sit close behind Inkling at 90.1% on VoiceBench, 77.0% on MMAU, and 54.9% on Audio MC. Treat Inkling Small as a capable reasoner with a smaller store of memorized facts, and give it a retrieval path when questions turn factual.
Vision is a practical strength. Inkling Small scores 74.0% on MMMU Pro and 77.4% on CharXiv reasoning questions, rising to 81.3% when it crops, zooms, and inspects images programmatically. That helps most on documents, forms, and charts where the detail that answers the question is small.
Controllable thinking effort runs from minimal to maximum, so you can trade answer quality against cost and latency per request. Sweep the setting across representative traffic rather than fixing it up front.
Inkling Small is compatible with Zero Data Retention on AI Gateway. Turn it on team-wide from the dashboard, or per request with zeroDataRetention: true, and AI Gateway routes only to providers that delete prompts and responses after each request. Set the model to thinkingmachines/inkling-small in the AI SDK, Chat Completions API, Responses API, Messages API, or other API formats, from TypeScript or Python. To use Inkling Small in a coding agent, run vercel ai-gateway coding-agents setup, then select thinkingmachines/inkling-small in the agent's model configuration.