New Model, Big Claim

A Small Model Just Rattled the Leaderboard

Inkling Small launched July 30 with a fraction of its sibling's size and benchmark scores that turned heads fast.

The Setup

276 Billion Parameters, Only 12 Billion Active

Thinking Machines Lab's second model runs open weight under Apache 2.0, built for cheaper self hosting than the flagship.

The Numbers

It Beat Its Own Bigger Sibling on Key Tests

SWE-bench Verified hit 80.2% vs 77.6%. GPQA Diamond hit 89% vs 87%. Humanity's Last Exam: 32% vs 30%.

The Twist

DeepSeek Quietly Updated Its Model a Day Later

DeepSeek V4 Flash 0731 jumped to 82.7 on Terminal Bench 2.1, making Inkling Small's launch comparisons instantly outdated.

The Takeaway

The Real Story Was How Fast Comparisons Aged

Inkling Small still shines on coding and reasoning, but DeepSeek's update erased its agentic edge within 24 hours.

Go Deeper

See the Full Benchmark Breakdown

Read the complete inkling small benchmarks comparison, including every score, source and what it means for developers.