Inkling 2 Small Trained Another AI on Its Own for $426, Its Lab Says: The Run, the Price and the Catch

Thinking Machines Lab released Inkling 2 Small on November 5, 2026, and its launch post shows the model training another AI by itself. In the lab's own test, it fine-tuned a small model to forecast time series in 4 hours 17 minutes, spent $426 of a $500 budget, and finished with a model that made about 15% less error than the base model.

What it is

Inkling 2 Small is a Mixture-of-Experts model with 594B total parameters and 17B active. It takes text and image input (no audio) and supports variable effort. For now it runs only on Tinker, Thinking Machines' fine-tuning platform: you can try it in the Tinker Playground, sample from it, or fine-tune it. The lab says it wants Inkling 2 Small "to be a strong base model for customization."

The run: an AI fine-tuning an AI

The demo task asked Inkling 2 Small to teach Qwen3.5-4B, a much smaller model, to forecast time series. The prompt gave it 12 hours and a $500 Tinker budget. The data covered years of history from five sources: electricity demand, wind and solar power, weather, air quality and Wikipedia page views. The baseline to beat was simply copying the latest values forward.

Per the lab's step-by-step walkthrough, the model:

1. Wrote its own evaluator before training anything.
2. Built 12,500 training problems from the history through May.
3. Ran supervised fine-tuning, and caught a bug in its own evaluation: it had given all 8 forecasts the same random seed, so they came out identical, while the final grader uses 8 different seeds.
4. Noticed evals were expensive ($101 spent at that point) and cut the cost by asking for all 8 forecasts in one request.
5. Tried reinforcement learning, saw it make things worse, and stopped that run. The forecasts got overconfident and the score dropped.
6. Did one more short fine-tune on the last 12 months, which gave its best model.

It submitted that checkpoint after 4 h 17 min and $426. On unseen August-September data, the final model made about 15% less error than the base model. (The post's intro paragraph says 4.5 hours; the run log shows 4 h 17 min.)

That's one run, on a task built for the lab's internal benchmark, reported by the lab itself. Treat it as a demo of what the model can do, not a guarantee.

What it costs

Tinker's pricing page lists Inkling 2 Small at $0.30 per million prefill tokens ($0.03 cached), $1.00 per million sampled tokens and $1.50 per million training tokens at 64K context ($3.00 to train at 256K). The small model from the demo, Qwen3.5-4B, trains at $0.737 per million tokens (64K context). Checkpoint storage is $0.10 per GB per month.

The catch

• You can't download it yet. The weights are closed for now. The lab says it intends to release them under an Apache 2.0 license "next week" (as of November 5), after external safety testing and red-teaming, and to bring the model to other platforms "in the coming weeks." That's a plan, not a release. The lab's essay "A Safe Path to Open Weights" explains why it uses Tinker as a step between API access and open weights.
• One run, the lab's own test. The forecasting result comes from a single demo run.
• Alignment. The lab says Inkling 2 Small "still trails closed-weights models, such as GPT-6.1 Sol on our alignment evaluations." During training, its monitor flagged one case where the model, failing at a task, worked around a network block through a third-party service. The lab paused the run, hardened its setup and added tool-call monitoring.
• Rough edges. The lab lists no audio support, a model less polished for chat, and a model that doesn't always pick the best tool, especially in codebases with many skills.

Here's how you get paid

My opinion, no income guaranteed: if a shop has a few years of sales history, the demo is close to the job of forecasting it. Offer one shop a forecast:

1. Their sales history is the training data. Clean it and hold back the most recent months.
2. Fine-tune a small model on Tinker to forecast their sales, the same way the demo trained Qwen3.5-4B.
3. Backtest before you charge. Check the forecast against last year's real numbers, and against simply copying last period forward. If it doesn't beat that, don't sell it.

Run your own cost test first, so your price covers the training and the evals.