Thinking Machines Lab releases Inkling-Small, matching Inkling at a quarter of the size

Thinking Machines Lab introduced Inkling-Small, a smaller Mixture-of-Experts model it says delivers “comparable performance to Inkling at a quarter of its size.” The official post is dated July 30, 2026 (earlier reports gave July 31).

Details

• Size: 276B total and 12B active parameters, versus 975B and 41B for Inkling. Context is up to 1M tokens with variable thinking effort.
• Multimodal design: Natively multimodal and encoder-free. Audio is represented as dMel spectrograms, and images are split into 40x40-pixel patches processed by a four-layer hMLP.
• Training: An adjusted pretraining data mix and recipe. Post-training started from an earlier Inkling-Small (preview) checkpoint trained by on-policy distillation with Inkling as teacher, followed by two weeks of scaled agentic coding RL.
• Benchmarks vs Inkling (lab-reported): SWEBench Verified 80.2% vs 77.6%, Terminal Bench 2.1 64.7% vs 63.8%, GPQA Diamond 89.5% vs 87.2%, HLE text-only 31.6% vs 29.7%. Inkling is ahead on AIME 2026 (97.1% vs 95.5%) and SimpleQA Verified (43.9% vs 20.6%), so the post says Inkling keeps its edge on knowledge and factuality.
• Availability: Full weights on Hugging Face (the model page is tagged Apache-2.0; the post text does not name the license), fine-tuning on Tinker, and chat with text, image and audio in the Tinker Playground. Both models carry a limited-time Tinker discount.
• Evaluation notes: Most evals ran at effort 0.99; Terminal Bench 2.1 used an internal harness and zeroed solutions contaminated by web search.

What happened next

The release makes a frontier-style open model far cheaper to serve and fine-tune, and Tinker’s one-year post (October 15) lists Inkling among supported models. The trade-off the lab itself reports is weaker factual recall. All benchmark numbers are first-party and unverified independently.