Thinking Machines Lab

Inkling

Provider

Thinking Machines Lab

Status

Current

Context

1,000,000 tok

SWE-bench

77.6%

Price

$1 / $4.05 /MTok

Inkling is Thinking Machines Lab’s first model — released on 15 July 2026 as Apache 2.0 open weights, seventeen months after Mira Murati founded the company. It is a 975-billion-parameter mixture-of-experts model (41B active) with a 1-million-token context that reasons natively over text, images and audio — trained on 45 trillion tokens that included video (Thinking Machines). Two weeks later, on 30 July, came Inkling-Small: a 276B model (12B active) that — unusually — beats its larger parent on several benchmarks, including SWE-bench Verified (80.2% vs 77.6%, vendor) and no-tools HLE (VentureBeat).

Two honest framings up front. First, the lab’s own: “Inkling is not the strongest overall model available today, open or closed” — the pitch is customisation via its Tinker fine-tuning service, not the frontier (Simon Willison). Independent scoring agrees: Artificial Analysis rates Inkling 42 on its Intelligence Index — 14th of 107 open-weight models in its class, well above that cohort’s median of 27, far below Kimi K3’s 60. Second, what nothing else in our ranking offers: the Inklings are the only open-weight models that take text, images and audio — the Chinese open cluster is text-only apart from Kimi K3’s vision.

Quick specs

VIEW INKLING →

What Inkling is

Inkling is a mixture-of-experts transformer with an unusually well-specified architecture for an open release: 256 routed experts plus 2 shared experts per layer, with 6 routed experts active per token, a sigmoid-based router with auxiliary-loss-free load balancing, and sliding-window attention interleaved with global attention at a 5:1 ratio (model card). Images enter through a four-layer hMLP encoder and audio as dMel spectrograms — multimodality is native, not bolted on, across 45 trillion training tokens spanning text, images, audio and video.

Its reasoning control is also distinctive: rather than the fixed low/medium/high tiers of most reasoning models, Inkling exposes a continuous thinking-effort dial — the vendor’s benchmark table is run at effort=0.99 — letting you trade cost against depth at any point on the curve.

Inkling-Small shares the architecture at 276B/12B, was trained on NVIDIA GB300 NVL72 systems, and is the more interesting release of the two: at roughly a quarter the parameters it beats its parent on SWE-bench Verified, no-tools HLE (31.6% vs 29.7%) and several agentic-coding rows, while the full-size Inkling keeps its lead on knowledge and factuality (Thinking Machines). For most self-hosters, Small is the one to deploy.

Benchmark performance

Vendor figures (from the model card, at effort=0.99 — a ceiling), then the independent view.

The shape: top-open-cluster coding on Verified (via Small), genuinely strong maths and science scores at maximum effort, a mid-pack independent composite, and no SWE-bench Pro figure in any harness — the same gap we flag for Grok 4.6 and GLM-5.3. No independent score for Inkling-Small had been published at the time of writing, so the Small-beats-parent story rests on vendor numbers.

Pricing and access

The weights are free — both models, Apache 2.0, on Hugging Face — with none of the custom-licence conditions attached to Llama or Kimi K3. There is no first-party inference API: hosted access runs through Together AI, Fireworks, Modal, Databricks and Baseten, where Artificial Analysis lists Inkling at about $1.00 input / $4.05 output per million tokens ($0.34 per index task). Inkling-Small is materially cheaper to serve at 12B active parameters.

The first-party product is Tinker — Thinking Machines’ fine-tuning service, with the Inklings as its raw material and the Tinker Playground for interactive use with text, image and audio input. That is the intended pattern: download or host the base model, fine-tune it to your workload.

How Inkling compares

• vs the open-weight leaders — On independent composites it is not close: Kimi K3 scores 60 and GLM-5.2 51.1 on the index where Inkling posts 42. On SWE-bench Verified, though, Inkling-Small’s vendor 80.2% sits level with the DeepSeek V4 / MiniMax M3 / Kimi K2.6 cluster.
• the multimodal edge — Text + image + audio input is unique among open-weight models on our board; Kimi K3 offers vision only, and the rest of the open cluster is text-only. If your self-hosted workload is multimodal, Inkling is currently the only game in town.
• vs the US open field — Simon Willison’s framing: a new viable contender alongside NVIDIA’s Nemotron and Gemma 4 — the US open-weights bench, which had been thin.
• within the family — Deploy Small unless you specifically need the knowledge/factuality edge of the 975B model; it is stronger on coding, cheaper to serve, and a quarter the download.

Known limitations

Mid-pack on the independent composite — 42 against the open leaders’ 51–60 — and the lab itself says it is not the strongest model available. No SWE-bench Pro figure in any harness, which costs it under our coding-weighted methodology; the flattering Verified figure is vendor-run. The Small-beats-parent claim is vendor-only so far — no independent score for Inkling-Small existed at the time of writing. Thin documentation — Willison’s criticism: a sparse model card with minimal transparency on data sourcing. Text-only output — it reasons over audio and images but does not generate them.

Frequently asked questions

What is Inkling?

Inkling is Thinking Machines Lab’s first model, released 15 July 2026 — a 975-billion-parameter mixture-of-experts model (41B active) with a 1M-token context that natively accepts text, images and audio, released as Apache 2.0 open weights on Hugging Face.

What is Inkling-Small?

A 276B-parameter (12B active) sibling released 30 July 2026 that — on the vendor’s own benchmarks — beats the full-size Inkling on SWE-bench Verified (80.2% vs 77.6%) and no-tools HLE, while the larger model keeps its lead on knowledge and factuality. For most self-hosted deployments it is the one to run.

Is Inkling open source?

The weights are open under Apache 2.0 — the cleanest major open-model licence, free to download, self-host, fine-tune and use commercially, with none of the usage conditions attached to Llama or Kimi K3. Training data and full training code are not published.

How good is Inkling really?

Mid-pack among open models overall, with two genuine distinctions. Artificial Analysis independently scores it 42 on its Intelligence Index — 14th of 107 open-weight models in its class, well above the median of 27 but far below Kimi K3’s 60. The distinctions: it is the only open-weight model taking text, images and audio, and its Small variant posts a vendor 80.2% on SWE-bench Verified, level with the top open coding cluster.

How much does Inkling cost?

The weights are free to self-host. Hosted through partners (Together AI, Fireworks, Modal, Databricks, Baseten), Artificial Analysis lists roughly $1.00 input / $4.05 output per million tokens for Inkling, at $0.34 per completed index task; Inkling-Small is cheaper to serve. There is no first-party API price list.

Why does the smaller model beat the bigger one?

Thinking Machines attributes it to post-training focus — Small was trained later, on NVIDIA GB300 systems, with the lab’s newest recipes. The pattern (a refined small model beating a bigger, earlier sibling on agentic tasks while trailing on knowledge) matches what post-training-heavy releases like GLM-5.3 have shown this year: recipe currently moves these benchmarks more than scale. The caveat: the comparison is vendor-run, with no independent Inkling-Small score yet.

Last verified 19 August 2026. Specs, architecture and benchmark figures are from Thinking Machines’ model card and release posts (vendor-run, at effort=0.99 — treat as a ceiling); the Intelligence Index score, rank and hosted pricing are Artificial Analysis’s independent measurements. The Apache 2.0 licence is stated on the model card. No SWE-bench Pro figure or independent Inkling-Small evaluation existed at the time of writing; we will update when they do.