AI HOT (Curated Pool)Jul 31, 2026, 01:47 UTC+8

Inkling-Small released: 276B total params, 12B active, matches original Inkling performance

Inkling-Small 发布，276B 参数性能持平原版

Thinking Machines open-sourced Inkling-Small: 276B total parameters, 12B active, one-quarter the size of the original Inkling yet matching its performance. Full weights are available. You can fine-tune it on Tinker or chat with it via text, image, and audio in Tinker Playground. The post doesn't disclose specific benchmark scores or comparison details, so I'd hold off on the 'matching performance' claim until independent evals appear.

Why it matters: Thinking Machines released Inkling-Small: 276B total params, 12B active at inference, full weights open for fine-tuning. The compression ratio and open release are strong, but the post doesn't disclose specific benchmark numbers, so the score stays below 80.

Read the original ↗Export Markdown

Take

Thinking Machines shrunk their Inkling model to one-quarter size and open-sourced the full weights. It's a 276B MoE with 12B active parameters per inference—light enough to run on modest hardware if the architecture is efficient.

I'd hold off on the 'matching original performance' claim. The announcement gives no benchmarks, no comparison tables, no task-specific scores. The original Inkling didn't have a ton of public evals either, so there's nothing to cross-check against. This could be solid, but right now it's a claim without receipts.

The useful bit: weights are live on Tinker for fine-tuning, and the Playground handles text, image, and audio inputs. If you're already in the Thinking Machines ecosystem, you can kick the tires today. If you're comparing general-purpose models, wait for independent benchmarks.

Source: AI HOT (Curated Pool) · x.com
Published: Jul 31, 2026, 01:47 UTC+8 · Jul 31, 2026