Inkling
Inkling is an open-weights multimodal Mixture-of-Experts model that reasons over text, images, and audio. It supports controllable thinking effort and a context window of 256K tokens. Call Inkling on AI Gateway with thinkingmachines/inkling. Your use is subject to Thinkingmachines's Terms & Privacy Policies.
import { streamText } from 'ai'
const result = streamText({ model: 'thinkingmachines/inkling', prompt: 'Why is the sky blue?'})Playground
Try out Inkling by Thinkingmachines. Usage is billed to your team at API rates. Free users (those who haven't made a payment) get $5 of credits every 30 days.
Inkling
Inkling
Providers
Route requests across multiple providers. Copy a provider slug to set your preference. Visit the docs for more info. Using a provider means you agree to their terms, listed under Legal.
P50 throughput on live AI Gateway traffic, in tokens per second (TPS). Visit the docs for more info.
P50 time to first token (TTFT) on live AI Gateway traffic, in milliseconds. View the docs for more info.
Direct request success rate on AI Gateway and per-provider. Visit the docs for more info.
More models by Thinkingmachines
| Model |
|---|
| 1M | 0.2s | 321tps | $0.50/M | $1.20/M | Read:$0.1/M Write:— | — | +2 |    | 07/30/2026 |
About Inkling
Inkling became available on AI Gateway on July 15, 2026. Inkling is a decoder-only transformer with a sparse Mixture-of-Experts (MoE) backbone: 66 layers, 975 billion total parameters, and 41 billion active per token, with each token routed to 6 of 256 experts plus 2 shared experts. Attention mixes local and global layers, and the context window is 256K tokens. Thinkingmachines released the weights under the Apache 2.0 license.
Multimodality is native rather than bolted on. Images enter through a hierarchical patch encoder and audio through discrete token encoding, and both are processed jointly with text by the same decoder. Inkling accepts pixel-based images with each dimension between 40px and 4096px, and WAV audio sampled at 16kHz. Inkling transcribes speech, follows spoken instructions, and answers questions about recordings, scoring 91.4% on VoiceBench, 77.2% on MMAU, and 56.6% on Audio MC. On vision, Inkling scores 73.5% on MMMU Pro and 78.1% on CharXiv reasoning questions, rising to 82.0% when it uses a Python tool to zoom into and crop the image.
On agentic and reasoning evaluations at maximum effort, Inkling scores 77.6% on SWE-bench Verified, 54.3% on SWE-bench Pro (public), 63.8% on Terminal-Bench 2.1, 76.0% on MCP Atlas, and 45.5% on Toolathlon Verified. Reasoning results include 87.2% on GPQA Diamond, 97.1% on AIME 2026, and 29.7% on Humanity's Last Exam text-only, which rises to 46.0% with tools. Inkling scores 79.8% on IFBench for instruction following.
Controllable thinking effort is the setting you tune most. Raising effort spends more thinking tokens for higher scores, and lowering it returns answers sooner for less. Inkling reaches a given score at fewer thinking tokens than the open-weights models Thinkingmachines compared it against, so sweep the effort setting across a representative slice of your traffic before you fix a value.
Inkling also aims for calibrated confidence. It hedges or says it doesn't know rather than guessing, which helps in forecasting and in any workflow where a confident wrong answer costs more than an uncertain one.
Set the model to thinkingmachines/inkling in the AI SDK, Chat Completions API, Responses API, Messages API, or other API formats, from TypeScript or Python. AI Gateway serves Inkling through Baseten, Together AI, with retries and failover, and mirrors provider pricing with no markup and no platform fee on inference, including on Bring Your Own Key (BYOK) requests.
What To Consider When Choosing a Provider
- Configuration: Thinkingmachines states plainly that Inkling is not the strongest overall model available, open or closed. Its case is breadth: one model that accepts text, images, and audio, follows instructions closely, and exposes an effort dial. Factual recall is the weakest area, with 43.9% on SimpleQA Verified, so pair Inkling with retrieval or search when answers depend on specific facts.
- Zero Data Retention: AI Gateway supports Zero Data Retention for this model via direct gateway requests (BYOK is not included). To configure this, check the documentation.
- Authentication: AI Gateway authenticates requests using an API key or OIDC token. You do not need to manage provider credentials directly.
When to Use Inkling
Best for
- Multimodal Agent Backends: Text, image, and audio input handled by a single model rather than three
- Speech and Audio Reasoning: Transcription, spoken instructions, and questions over longer recordings
- Chart and Document Vision: Charts, diagrams, and visual math, with a Python tool for zooming and cropping
- Tool-Heavy Agent Workflows: Broad tool use across harnesses, with 76.0% on MCP Atlas
- Mixed Workload Consolidation: One generalist covering reasoning, coding, chat, and multimodal input
- Effort-Tuned Cost Control: Thinking effort set per request to balance answer quality against latency
Consider alternatives when
- Peak Coding Scores: Dedicated coding models post higher SWE-bench and Terminal-Bench 2.1 results
- Factual Recall Workloads: 43.9% on SimpleQA Verified means knowledge-heavy answers need a retrieval layer
- Lower Cost Per Task: Inkling Small matches or beats Inkling on many evaluations at a quarter of the size
- Text-Only Pipelines: A text-focused model fits better when image and audio input never apply
Conclusion
Inkling is a broad, balanced model rather than a leader on any single benchmark family. Use Inkling when one model needs to read images, listen to audio, call tools, and write code, and when you want an effort dial to control what each request spends. Try it in the playground at https://ai-sdk.dev/playground/baseten:thinkingmachines/inkling.