Thinking Machines has released the open-source Inkling-Small model, with a total of 276 billion parameters and 12 billion activated parameters, featuring a MoE architecture and native multimodal design. Despite being only one-quarter the size, it outperforms the original Inkling model on benchmarks including mathematics, reasoning, and agent coding. Performance surpass was achieved through distillation and two weeks of reinforcement learning training, significantly lowering the barrier to inference and enabling mid-sized teams to customize the model. Despite four of the six co-founders having left, the team has maintained a high release cadence amid the turmoil.

Article author and source: 36Kr

Thinking Machines releases Inkling-Small, an open-source model that surpasses the original.

After the debut of the first model, Inkling, a new AI has taken over social media!

Today, Thinking Machines officially releases its second major model—Inkling-Small.

276B total parameters, 12B activated parameters, native multimodal, 1 million token context

Half a month ago, the first open-source 975B Inkling debuted, causing a major stir in the AI community.

Inkling-Small achieved superior performance to the "trillion-dollar giant" using only one-quarter of the scale.

On ARC-AGI-2, Inkling-Small sets a new open-source SOTA.

Looking back at the past few days of Thinking Machines, the plot has been nothing short of dramatic.

Just two days after co-founder Weng Li announced her departure, OpenAI confirmed her return—leading the effort on Recursive Self-Improvement (RSI).

On the third day, the company, having just lost one of its key figures, released the full weights of the new model.

Introducing Inkling-Small: 276B outperforms trillions.

In the official blog introduction, Inkling-Small is a MoE-based model with a total of 276 billion parameters and 12 billion parameters activated per inference.

Compared to the first-generation Inkling, the size has been directly reduced to one-quarter.

As PyTorch expert Horace He said, "Inkling-Small doesn't require the effort of the entire village—it can simply build on previous technologies."

Most notably, although its scale has shrunk, its overall strength has significantly improved.

In Inkling-Small, performance matches or exceeds that of Inkling and DeepSeek V4 Flash in mathematics, reasoning, agent coding, and multimodal benchmarks.

On the three core metrics—HLE, Terminal-Bench, and IFBench—Inkling-Small delivers higher performance per FLOP than Inkling.

As you can see, the Inkling-Small curve remains consistently below the Inkling curve.

On the multimodal side, Inkling-Small also held its ground nearly as well, achieving performance "close" to the original model in evaluations including chart understanding, visual reasoning, speech understanding, and long-audio reasoning—at a much lower cost.

Other benchmark results are as follows:

A model only one-quarter the size crushed its larger predecessor in reasoning and agent tasks.

As for how to do it, the official guide details the process thoroughly, with two key steps.

Inkling-Small began training later than Inkling, so it fully absorbed all the pitfalls from the previous round—

The pre-training data ratio has been adjusted, and the machine learning model has been refined accordingly.

More importantly, it's a two-step process.

First, use Inkling as a teacher for on-policy distillation and generate a preview checkpoint.

Step two: We ran agent encoding reinforcement learning for a full two weeks starting from this checkpoint, resulting in the "student surpassing the teacher" outcome.

In just two weeks, we closed the gap and even surpassed them.

This means Thinking Machines has directly established a production line capable of repeatedly generating models. The first model is a work; the second model is proof of production capacity.

12 billion activated is the true watershed moment.

For most developers, what truly matters is whether they can actually use it.

The original Inkling threshold is very high; the minimum requirement for BF16 checkpoints is 2 TB of aggregated VRAM—

The sample configuration is 8 B300s or 16 H200s; even with the NVFP4 quantized version, you still need at least 600 GB.

Inkling-Small has significantly lowered this barrier.

LMSYS's statement is straightforward: 276B total parameters with 12B activated is the sweet spot for reinforcement learning, making both LoRA and full-parameter training "achievable."

Previously, model customization was only accessible to large corporations, but now even a mid-sized team has the opportunity to truly train it into their own solution.

Real-world testing shows that, under the same environment with 8 B200, TP8, NVFP4, and batch size 1, running Inkling-Small with SGLang achieves 648 tok/s decoding with DSpark enabled, and 288 tok/s without DSpark.

In February last year, Thinking Machines made a stunning debut with a six-member team dubbed the "OpenAI dream team":

Mira Murati, John Schulman, Barret Zoph, Weng Li, Andrew Tulloch, Luke Metz.

At the time, they had an extraordinary deal: a $2 billion seed round with a $12 billion valuation.

Who could have imagined that the $5 billion funding round in January this year, which ultimately fell through, became the beginning of a turning point?

Now, only Murati and Schulman remain from the original six.

More interestingly, among the four who left, Zoph and Metz returned to OpenAI one after another, and now Weng Li is the third.

A company founded by "defectors" from OpenAI ultimately couldn't escape OpenAI's gravitational pull.

But what was truly dramatic was the team's response during times of turmoil.

On July 15, Thinking Machines, which had been silent for a full 17 months, finally unveiled its first large model, Inkling.

Less than two weeks later, Weng Li announced her departure; just three days after that, a second open-source model with full weighting was released.

Talent is accelerating away, while models are racing forward.

These two diverging curves have become the most iconic moment in the AI industry.

Perhaps this is the inevitable path to ASI: no matter how few remain, the wheels of progress will only turn faster.

Thinking Machines has released the open-source Inkling-Small model, with a total of 276 billion parameters and 12 billion activated parameters, featuring a MoE architecture and native multimodal design. Despite being only one-quarter the size, it outperforms the original Inkling model on benchmarks including mathematics, reasoning, and agent coding. Performance surpass was achieved through distillation and two weeks of reinforcement learning training, significantly lowering the barrier to inference and enabling mid-sized teams to customize the model. Despite four of the six co-founders having left, the team has maintained a high release cadence amid the turmoil.

Source: AI New Era