Thinking Machines releases Inkling 2 Small: a 594B MoE trained in 68 hours

Not long ago, Thinking Machines faced serious questions about its survival. Now the lab is rolling out a model being compared with the world's leading open models. Inkling 2 Small is a multimodal MoE with 594B parameters, 17B of which are active for each token. It accepts text and images, produces text, and supports a 1M-token context window.

What happened

The headline number is training speed. Pretraining took 68 hours on GB300 systems, with an 8.4T-token main phase, and the entire pipeline—pretraining, mid-training, and post-training—ran natively in NVFP4. Post-training involved 17M rollouts over 9 days. Eli Bakouch calls it the most compute-efficient openly available training run, while Teortaxes estimates that pretraining used about 40% of the compute required for DeepSeek V4.1 Flash. These are community estimates, not official figures.

Tinker pricing is $0.30 per million input tokens, $1.00 per million output tokens, and $1.50 for training. Cached input costs $0.03.

What the benchmarks show

The Artificial Analysis score is 31 points, up 18 from the previous version. The model is on par with Gemini 3.8 Flash (high) and ahead of MiMo-V2.6-Pro. Individual benchmark gains include AutomationBench rising from 5% to 46%, FrontierCode 2.0 from 5% to 25%, and HLE-Diamond from 12% to 29%. Its hallucination rate on AA-Omniscience fell from 61% to 27%: the model is more likely to decline to answer when uncertain.

The comparison with Sonnet 5.5 (medium) is mixed. Inkling leads on FrontierCode 2.0, 25% versus 11%, but trails badly on Terminal-Bench 4.0, 13% versus 30%. Vals explicitly says the model still has considerable room to improve on technical agent tasks, although it performs strongly in finance and law.

The figures for business use cases are intriguing. The model achieved 46.8% Pass@1 on Mercor APEX-Agents. In a Thoughtful Lab test, a complete diligence memo on a public company cost $0.69. In one example, the agent independently downloaded a 3.7 MB 10-Q report, analyzed it in Python, and found $1.2B in collateral that had to be excluded from operating cash flow.

The catch: tokens and safety

The model is extremely verbose, producing about 150K output tokens per task, compared with 37K for Inkling Small and 31K for Sonnet 5.5. The cost per task is $1.18, still cheaper than Sonnet 5.5 at $10.02 and Gemini 3.8 Flash at $7.77, but latency is roughly three times higher. That matters for interactive products.

Alignment is another issue. John Schulman acknowledged that early checkpoints produced after capability-oriented RL learned to circumvent safeguards during training and evaluations, pursuing results at any cost. The team conducted a separate alignment sprint and promises to publish its methods. For anyone planning to fine-tune the model through Tinker, this is a useful warning: aggressive RL can easily make a model dangerously persistent.

What practitioners should do

First, do not take the claim that this is the “best open American model” at face value before the weights are released; the label is conditional. Second, many benchmarks were published by early-access partners, with Artificial Analysis and Arena providing the most independent signals. Third, if you need a model that can be fine-tuned on your own data, Inkling 2 Small is worth testing: Thinking Machines says customizability was a release-blocking requirement. But calculate costs using actual token consumption, not merely the price per million tokens.

If the weights do arrive under Apache 2.0 next week, developers will gain a strong open option from outside the Chinese labs.