Inkling-Small matches its larger sibling at a quarter the size
- Source
- x.com
- Date
Today, we are releasing Inkling-Small. Inkling-Small achieves comparable performance to Inkling at a quarter of its size. It features 276B total parameters, 12B active. We are making the full weights available. https://t.co/BtYNcpkDRA Fine-tune it on Tinker today, or chat with it in text, image, and audio on Tinker Playground.
Efficiency is the point. Across agentic tool use (Terminal-Bench 2.1), reasoning (HLE), and instruction following (IFBench), Inkling-Small delivers more performance per FLOP than Inkling. Variable thinking effort lets you pick your point on the cost/performance curve.
Inkling-Small began training after its larger counterpart, so it benefits from everything we learned: an improved pre-training data mix, a refined ML recipe, on-policy distillation with Inkling as the teacher, and two further weeks of agentic coding RL.
It matches or exceeds Inkling on reasoning and agentic tasks. 31.6% on HLE, ahead of Inkling’s 29.7%, and the advantage holds at every thinking budget. On SWEBench-Verified it exceeds 80%.
Context
Inkling-Small shows why fewer parameters need not mean worse performance on every task. Thinking Machines reports that it matches or exceeds the original Inkling on reasoning and agentic coding benchmarks, while the larger model retains an advantage in factuality and knowledge coverage. Small has 276 billion total parameters, with 12 billion active, compared with 975 billion and 41 billion respectively. Release details.
It also received a revised training recipe, including on-policy distillation using Inkling as teacher and further coding training. The comparison therefore reflects both size and training changes. Its appeal is a lower-compute option for particular workloads, with task-specific trade-offs. Training and capability comparisons.
- distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
- pretraining — The first, biggest phase of building a model: training it on enormous amounts of text so it learns language, facts, and reasoning in general.
- SWE-bench — The standard benchmark for AI coding agents: real GitHub issues from real repositories, scored by whether the agent's patch passes the project's own tests.
Checking sign-in…
Loading comments…