Token efficiency
WinnerAA tokens per Intelligence Index task
Lower is better. Launch coverage cites Inkling at ~25k tokens per AA Intelligence task, leaner than several higher-scoring open peers.
SPIRITT
InklingInkling is the open model you shape, not just prompt
Thinking Machines Lab's first open-weights model: a 975B MoE (41B active) generalist that reasons across text, images, and audio with controllable thinking effort. On SPIRITT, you run it inside a fully equipped cloud environment with tools, browser, and durable memory.
Launch day on X
People are reacting to a US open-weights multimodal release built to be customized, not just prompted.
Official launch card
From Thinking Machines Lab: Inkling is a broad, balanced open-weights foundation model trained from scratch for customization. Multimodal inputs, controllable thinking effort, and native fine-tuning on Tinker.
What it was built for
A Mixture-of-Experts transformer with 975B total parameters and 41B active, pretrained on 45 trillion tokens of text, images, audio, and video. Native context up to 1M tokens; Tinker currently serves 64K and 256K options.
Thinking Machines is explicit that Inkling is not the absolute strongest overall model. The pitch is a customizable generalist: agentic coding and tool use, efficient reasoning, multimodality, calibrated confidence, and open weights you can fine-tune.
975B MoE / 41B activeOpen weightsText + image + audio1M contextControllable effortFine-tune on Tinker
AA Intelligence Index
WinnerIndependent Artificial Analysis. Inkling debuts at 41, the highest US open-weights score reported in launch coverage, ahead of Nemotron 3 Ultra (38) and Gemma 4 31B (29). Chinese open labs still lead the absolute top.
Where Inkling is strong for agent work
Primary scores from the Thinking Machines launch tables (effort 0.99) plus independent AA framing on intelligence and token efficiency.
Agentic coding
Near frontierSWE-Bench Verified
Thinking Machines launch table, effort 0.99. Competitive among open-weights peers.
Terminal agents
Terminal Bench 2.1 (best harness)
Launch table. Inkling is competitive with open peers but not the absolute leader.
Tool use
Near frontierMCP Atlas
Scaled multi-tool / MCP use from the launch table.
Reasoning + tools
Humanity's Last Exam (with tools)
AA / launch-reported HLE with tools. Strong but not the top open score.
Instruction following
Near frontierIFBench
Launch table. Inkling is near the open-weights leaders on hard instruction following.
Vision
MMMU Pro (Standard 10)
Native multimodal vision from the launch table.
Audio
WinnerVoiceBench
Among open omni models in the launch multimodal table, Inkling leads VoiceBench.
Safety
WinnerFORTRESS (Adversarial)
Launch table: strongest built-in safeguards among the open-weights models they compared.
Sources: Thinking Machines Inkling launch post and model card (Jul 15, 2026); Artificial Analysis Intelligence Index and token-efficiency figures cited in independent launch coverage. Primary capability rows are vendor-reported at effort 0.99 unless noted as AA. Inkling is positioned as a customizable open generalist, not the absolute top of every frontier board. On SPIRITT, you run the model inside a workspace with tools, browser, and durable context already set up.
How It Works
From zero to Inkling running real multimodal agent work
01
Open a workspace
Start free, including trial. You land in a fully equipped cloud computer: browser, files, terminal, integrations, and memory. No local setup. No thin chat box pretending to be an agent.
02
Pick Inkling
Open the model picker and choose Inkling. The open-weights multimodal model lands inside a workspace that already has browser control, files, terminal, and durable context, ready for coding, tool use, and fine-tuned workflows.
03
Build or automate
Tell it what to ship or what to run. Inkling can code, call tools, drive the browser, coordinate multi-step work, and keep going while you step away. The point is not another chat window. It is an agentic environment where Inkling actually does the job.
Try Inkling in an agentic environment
A real cloud environment, model picker, tools, and memory. Included in all plans, including trial.
Questions
Frequently asked questions
01What is Inkling?+
Inkling is Thinking Machines Lab's first open-weights foundation model: a Mixture-of-Experts transformer with 975B total parameters and 41B active. It accepts text, image, and audio inputs, supports long context, and is designed as a customizable generalist for agentic and multimodal work.
02How is Inkling different from closed frontier models?+
Thinking Machines does not claim it is the strongest overall model. The bet is open weights plus easy customization: controllable thinking effort, strong multimodality, and fine-tuning on Tinker so teams can adapt the model instead of only calling a fixed API.
03What is Inkling best for?+
Agentic coding and tool use, multimodal workflows that mix screenshots or audio with text, long refinement loops, and domain fine-tuning. Independent AA framing also highlights competitive intelligence with relatively lean token use.
04What are the main limitations?+
It is not the absolute top of every pure-coding or closed-frontier board. Hosted Tinker context options (64K / 256K) are below the model's native 1M window. Running full weights yourself is heavy hardware, so most teams will use hosted inference or fine-tuning platforms.
05What is a good way to test it?+
Open a SPIRITT workspace, pick Inkling, and give it a multi-step job: build a small app, use browser tools, refine against feedback, or run a vision/audio-aware workflow. That tests the customizable generalist claim better than a single chat prompt.
06Where is Inkling best to test for full agentic capabilities?+
SPIRITT Workspaces. Pick Inkling in the model picker and run real work in a fully equipped cloud environment: tools, browser, files, terminal, memory, and durable sessions. Available on every plan, including trial. That is where the model can act as an agent, not just chat.
Buy from builders who use what they sellBuilt using