Thinking Machines Lab · 2026-07-30 · major
Inkling-Small — Thinking Machines' 276B open model matches Inkling at 1/4 the size
Thinking Machines released Inkling-Small under Apache 2.0: a 276B-parameter MoE with 12B active, 1M-token context, and multimodal inputs. Scores 80.2% on SWE-bench Verified and 95.5% on AIME 2026.
A 276B MoE with 12B active parameters that ships full Apache-2.0 weights and matches its 4x-larger sibling.
Quick facts
| Maker | Thinking Machines Lab |
|---|---|
| Parameters | 276B total, 12B active (MoE) |
| Architecture | 42-layer decoder, 256 experts (6+2 routing) |
| Context window | 1M tokens |
| Inputs | Text, image, audio |
| License | Apache 2.0 (open weights) |
| Availability | Hugging Face, Tinker Playground, Tinker API fine-tuning |
Benchmarks
Pricing
| Tinker API output | $1.20 / 1M tokens |
|---|---|
| Tinker Playground · Free chat with text, image, audio | $0 |
| Open weights · Apache 2.0 on Hugging Face | $0 |
What is it?
Inkling-Small is a mixture-of-experts model with 276B total parameters and 12B active per token, released under Apache 2.0 with full weights on Hugging Face. It handles text, image, and audio input, has a 1M-token context window, and reports variable thinking effort at inference time.
How does it work?
The architecture uses a 42-layer decoder with 256 experts, routing each token to six specialised experts plus two shared ones. Encoder-free multimodal handling relies on dMel spectrograms for audio and 40×40 pixel patches for vision. Training used NVIDIA GB300 NVL72 systems, and weights ship in BF16 and NVFP4 formats.
Why does it matter?
Inkling-Small closes most of the gap to the 975B Inkling — 80.2% vs 77.6% on SWE-bench Verified and 64.7% vs 63.8% on Terminal Bench 2.1 — at roughly a quarter of the size. That makes a frontier-class multimodal reasoning model runnable on a single 128 GB Mac Studio via Unsloth's GGUF quantisation, and cheap enough on Tinker ($1.20 per 1M output tokens) for teams to fine-tune on private data.
Who is it for?
developers wanting a top-tier open MoE they can fine-tune or self-host
Frequently asked questions
- How does Inkling-Small compare to the full Inkling model?
- Inkling-Small matches Inkling on most benchmarks despite having a quarter of the total parameters (276B vs 975B). It even edges Inkling on SWE-bench Verified (80.2% vs 77.6%) and Terminal Bench 2.1 (64.7% vs 63.8%), while trailing slightly on factuality. The full model still leads on the Artificial Analysis Intelligence Index (41 vs 40).
- Where can I run Inkling-Small locally?
- Inkling-Small weights are on Hugging Face under `thinkingmachines/Inkling-Small` (BF16) and `thinkingmachines/Inkling-Small-NVFP4` (4-bit). Unsloth has published GGUF conversions that run locally on 128 GB RAM. Hugging Face Inference Endpoints achieves up to 160 tokens/sec on the reference stack.
- How much does the Tinker API cost for Inkling-Small?
- Inkling-Small runs on the Tinker API at $1.20 per 1M output tokens. Fine-tuning on your own data is available through the same API, and chat with text, image, and audio inputs is free in the Tinker Playground.
- Is Inkling-Small really the strongest open model in its size class?
- Thinking Machines and Artificial Analysis both call it the leading US open-weights release for its size — it scores 40 on the AA Intelligence Index and beats similarly-sized peers on SWE-bench, AIME, GPQA Diamond, and MMMU Pro. Larger open models like Kimi K3 (2.8T) still top the raw leaderboard.
Try it
huggingface.co/thinkingmachines/Inkling-Small