Research Free — open weights download; Tinker Playground access ★ Editor's pick

Inkling-Small

Thinking Machines Lab's open-weights 276B mixture-of-experts model: 12B active params, native audio and vision, and a 1M-token context window.

Updated 2026-07-31

8.8

AI Score / 10

Visit Inkling-Small

Overview

Inkling-Small is an open-weights language model released by Thinking Machines Lab on July 30, 2026. It's a 276-billion-parameter mixture-of-experts architecture that activates roughly 12 billion parameters per token, handles audio and vision natively rather than through bolted-on encoders, and carries a 1M-token context window. The weights are published on Hugging Face, and the model is also reachable through Thinking Machines' Tinker Playground for anyone who doesn't want to stand up 276B worth of inference hardware.

The interesting part is the name. This is the small model, and Thinking Machines says it beats the larger Inkling on several benchmarks — a claim worth checking against the official results table rather than taking on faith, but one that lines up with a broader industry pattern where sparse MoE models with aggressive post-training outrun bigger, denser siblings on reasoning and agentic coding tasks. If that holds, the sparsity ratio is the headline: 12B active parameters is roughly what a mid-size dense model costs to run per token, at a capability level the lab is positioning near its flagship.

Who it's for: research teams and infrastructure-heavy shops that want frontier-adjacent multimodal capability under weights they control, and anyone whose use case rules out sending audio or documents to a hosted API. Who it isn't for: individuals and small teams who just want a chat assistant. "Open weights" and "free" are not the same thing as "cheap to operate" — a 276B checkpoint needs serious VRAM to serve even at 12B active, so for most people the practical path is Tinker or a third-party host, not local inference. If you're comparing against DeepSeek or Meta Llama on the open-weights shelf, the differentiators here are native audio-plus-vision in one model and the 1M context, not raw parameter count.

Key features

Sparse MoE routing

276B total parameters with about 12B active per token, so serving cost per token tracks a mid-size dense model rather than the full parameter count. The tradeoff is memory: you still need to hold all 276B in VRAM to route across the experts.

Native audio and vision

Audio and image understanding are built into the model rather than handled by a separate encoder pipeline. That matters for tasks where cross-modal context has to survive into reasoning — transcribing a meeting and acting on it, or reading a chart inside a document.

1M-token context

Enough headroom to hold an entire codebase, a long document corpus, or hours of transcribed audio in a single pass. As always with long-context claims, effective recall across the full window is the thing to test rather than assume.

Open weights on Hugging Face

The checkpoint is downloadable, so you can fine-tune, quantize, or run it fully offline for data that can't leave your infrastructure. Check the specific license terms on the model card before commercial deployment — release announcements and license files don't always say the same thing.

Pricing

Free tier: Yes — the weights themselves are free to download and run. Compute is the real cost.

Plan	Price	What's included
Open weights	Free	Full model checkpoint downloadable from Hugging Face. Self-hosted inference costs are yours — a 276B MoE needs substantial GPU memory even at 12B active parameters. Verify license terms on the model card before commercial use.
Tinker Playground	Check website for current pricing	Hosted access through Thinking Machines' Tinker platform for evaluation and fine-tuning without provisioning your own hardware.

Open weights Free

Full model checkpoint downloadable from Hugging Face. Self-hosted inference costs are yours — a 276B MoE needs substantial GPU memory even at 12B active parameters. Verify license terms on the model card before commercial use.

Tinker Playground Check website for current pricing

Hosted access through Thinking Machines' Tinker platform for evaluation and fine-tuning without provisioning your own hardware.

Pros & cons

Pros

✓Open weights on Hugging Face, so it can run fully offline for sensitive audio, documents, or code
✓12B active parameters keeps per-token inference cost near a mid-size dense model despite the 276B total
✓Audio and vision are native to the model rather than a separate encoder stage
✓1M-token context handles whole codebases or long document sets in one pass
✓Hosted Tinker Playground path for teams that want to evaluate before committing hardware

Cons

×"Free weights" is misleading on cost — serving 276B parameters requires enough VRAM to put local deployment out of reach for individuals and small teams
×The claim that it outperforms the larger Inkling comes from the lab's own launch benchmarks; independent third-party evaluations aren't in yet
×Released July 30, 2026, so tooling, quantizations, and community fine-tunes are still thin compared to established open-weights families
×It's a raw model, not a product — no chat app, no agent harness, no support contract unless you build or buy one

How it compares

Tool	Best for	Pricing	Score
Inkling-Small	—	Free — open weights download; Tinker Playground access	8.8/10
Perplexity AI vs Perplexity AI →	Knowledge workers and researchers who want cited, synthesized answers instead of a list of links to click through.	Freemium	9.4/10
NotebookLM vs NotebookLM →	Students, researchers, and professionals who need answers grounded strictly in the specific documents they upload.	Free	9.1/10
Phind vs Phind →	Developers who want technical questions answered with working code examples and cited sources instead of manual doc searches.	Free tier + Pro subscription for advanced models	8.7/10