Thinking Machines' Inkling-Small Beats Its 975B Teacher at a Quarter of the Cost

Inkling-Small is out, and it is doing something that does not happen very often: the smaller model is beating the one it was distilled from. Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, just released the full weights of Inkling-Small -- a Mixture-of-Experts (MoE) model with 276B total parameters but only 12B active at inference time, trained to match or exceed its 975B-parameter sibling at roughly a quarter of the compute cost.

A model that learned from its teacher's mistakes

The backstory matters here. Inkling-Small began training after its larger counterpart, which gave the team a chance to improve the training process before the smaller model was ever run. That head start compounded in several ways:

• The team made changes to Inkling-Small's pre-training data mix and machine learning recipe.
• On-policy distillation -- a technique where the student model learns by generating its own outputs and then being corrected by the teacher in real time, rather than just imitating a static dataset -- was used with Inkling as the teacher model.
• Starting from an earlier checkpoint, the team continued scaling agentic coding RL for two more weeks.

With these improvements, Inkling-Small surpassed Inkling on reasoning and agentic coding benchmarks. Inkling maintains an advantage on knowledge coverage and factuality. The factuality gap is real: Inkling scores 43.9% on SimpleQA Verified versus Inkling-Small's 20.6%, so if your use case is grounded factual recall, the larger model still wins.

The numbers that matter

On Humanity's Last Exam (HLE) -- a benchmark of extremely hard, expert-level questions across science, math, and humanities -- Inkling-Small scores 31.6%, ahead of Inkling's 29.7%, and the advantage holds at every thinking budget. That last part is key: it is not just a cherry-picked single data point.

The agentic coding story is equally strong:

• On SWEBench-Verified, Inkling-Small exceeds 80% (80.2% to be exact), beating Inkling's 77.6%.
• On Terminal-Bench 2.1, it scores 64.7% , edging past Inkling's 63.8% and ahead of most open-weight models in its class.
• On abstract reasoning (ARC-AGI-2), it scores 40.1% versus Inkling's 36.5% -- a meaningful jump.

The efficiency story is where things get interesting for anyone running at scale. Across agentic tool use, reasoning, and instruction following benchmarks, Inkling-Small is more efficient than Inkling, and its variable thinking effort lets users adapt it to target their use case, balancing cost and performance. Think of thinking effort like a dial from minimal to xhigh -- you pay for reasoning tokens only when you need them.

Natively multimodal, encoder-free

Inkling-Small uses the same natively multimodal encoder-free architecture as Inkling. Audio is represented as dMel spectrograms, while images are divided into 40x40-pixel patches and transformed using a four-layer hMLP. Both are processed jointly with text tokens. There is no separate vision encoder to fine-tune around or worry about -- everything is one unified model.

The team also improved its ability to use Python for visual tasks. The model can combine visual reasoning with operations such as cropping, zooming, and programmatic image inspection, improving usability on documents and charts where relevant information may be small or difficult to inspect directly. On audio, it scores 54.9% on Audio MC, far ahead of specialist omni models like Qwen3-Omni (24.3%) and Gemini 3.5 Flash-Lite (33.6%).

Calibrated, not just confident

One underrated feature is the epistemics training. Calibration involved RL against proper scoring rules on a large corpus of real-world forecasting questions, improving the model's ability to express appropriate confidence and produce calibrated forecasts under uncertainty. On ForecastBench, Inkling-Small (61.3 Brier Index) actually edges out Inkling (60.1) and beats GPT-5.5 (59.3) and Claude Opus 4.8 (56.2). A model that knows what it does not know is genuinely more useful in production pipelines.

Where it falls short

The factuality gap is the clearest limitation. Inkling-Small scores 20.6% on SimpleQA Verified while Inkling scores 43.9% -- a 2x difference. If your application depends on accurate recall of specific facts (legal, medical, financial lookups), the larger model or a fine-tuned version of Inkling-Small on your own data corpus will serve you better. On the safety side, on FORTRESS adversarial -- which measures refusal of genuinely harmful requests -- Inkling-Small scores 71.6% versus Inkling's 78.0% , a modest but real gap worth noting for sensitive deployments.

Getting your hands on it

The full weights of Inkling-Small are available on Hugging Face , and it is live for fine-tuning and inference on Tinker Playground today. Inkling and Inkling-Small are available on Tinker with a limited-time discount and can be chatted with on Tinker Playground using text, image, and audio.

On the Tinker pricing page, Inkling (64K context) is currently $1.87/$4.68 per million input/output tokens at the discounted rate. Inkling-Small pricing is not yet listed separately, but MoE models on Tinker are priced by active parameters, making them significantly more cost-effective than dense models of similar quality -- so expect Inkling-Small to come in meaningfully cheaper than Inkling given its 12B vs 41B active parameter count.

For those who want to self-host, the weights are on Hugging Face under Apache 2.0. The practical use cases that make the most sense:

• Agentic coding pipelines: SWEBench >80% at 12B active parameters is a strong cost-per-solve ratio.
• Document and chart analysis: The Python-augmented visual reasoning is genuinely novel for a model this size.
• Voice and audio agents: Audio MC scores that beat dedicated omni models make it a strong candidate for call center or voice assistant workflows.
• Forecasting and decision support: The calibration training makes it one of the better open-weight options for probabilistic reasoning tasks.
• Fine-tuning base: The smaller active footprint makes it cheaper to run RL and SFT loops on Tinker for domain specialization.

The broader context here is that Thinking Machines raised $2 billion at a $12 billion valuation in its first year , and the strategy is becoming clearer with each release: open weights drive adoption, Tinker drives revenue, and each model in the family is designed to be a fine-tuning starting point rather than a finished product. Inkling-Small is the most compelling argument yet that this approach can produce models that punch above their weight class -- sometimes literally above the model they were trained from.