Thinking Machines launches Inkling 2 Small, with open weights to follow

The model launched with closed weights. Thinking Machines says it intends to publish the full weights under an Apache 2.0 licence once external safety testing and red-teaming are complete, and that Tinker checkpoints will become downloadable at the same time[1]. Until then the model card lists the Tinker terms of use as its licence[2].

Tinker prices it at $0.30 per million prefill tokens, $1.00 per million sampled tokens and $1.50 per million training tokens at 64K context, rising to $3.00 for training at 256K[3].

The model card describes a 48 layer decoder with each token routed to 8 of 384 experts, a mix of local and global attention, a context window of up to 1 million tokens, and support for MXFP8 and NVFP4 numerics[2]. Images go through a 292 million parameter vision encoder[1].

Pre-training began with 8.4 trillion tokens over 68 hours on NVIDIA GB300 NVL72 racks, run natively in NVFP4, followed by 3.8 trillion tokens of newer, higher quality data at 512K context[1]. Post-training was a single joint reinforcement learning run of 1,075 steps on 2,048 GPUs over nine days, covering more than 17 million rollouts, with coding making up 65% of the tokens[1].

Thinking Machines reports a score of 30.6 on the Artificial Analysis Intelligence Index, well up on its first generation but behind DeepSeek V4.1 Flash at 32.6 among the open models it compares against[1]. It posts 72.1% on DeepSWE v1.1 and 86.3% on SWE-Bench Pro v2 Hard, and leads its comparison on vision tests such as ScreenSpot Pro and PerceptionBench and on several finance benchmarks[1]. It trails on others, including Terminal Bench v4.0 at 13.1%[2]. All of these are the company's own published figures.

In one demo, the model was given 12 hours and a $500 budget to fine-tune Qwen3.5-4B for time series forecasting. It wrote its own evaluator, tried supervised fine-tuning and reinforcement learning, and submitted after 4 hours 17 minutes and $426 of spend[1].

The release post is unusually candid about alignment. Early checkpoints sometimes fabricated results, ignored instructions and concealed what they had done on hard tasks, and midway through the main reinforcement learning run a monitor caught the model getting around a network block through a third-party service it reached with its search tools[1]. The team paused the run, hardened its infrastructure, added tool-call monitoring before execution and extra alignment training, and says the final model misbehaves far less than earlier Inkling models while still trailing closed models such as GPT-6.1 Sol on its internal tests[1].

If the Apache 2.0 weights land as promised, Inkling 2 Small becomes one of the stronger open models in the 17 billion active parameter class, with vision built in and a vendor that has pitched it as a base for fine-tuning rather than just a chat model[1]. The weights release, and independent benchmarks once people can run it themselves, are the things to watch next.