Thinking Machines drops a 12B-active model that outsmarts its predecessor
Thinking Machines just launched Inkling-Small, an open-weights MoE model with 276B total parameters and only 12B active. Operating at roughly a quarter of the compute cost, it actually edges out the flagship Inkling across reasoning and agentic coding benchmarks like Humanity's Last Exam. Why it matters: You get multimodal reasoning and a 1M context window without the heavyweight inference tax.
Inkling-Small is an open-weights multimodal Mixture-of-Experts model from Thinking Machines that packs 276B total parameters while firing just 12B at inference. Built to slash compute costs to roughly a quarter of the original Inkling flagship, it actually ends up beating its bigger sibling across several core reasoning and coding workloads while maintaining native support for text, images, audio, and a 1M token context window.
The complete model weights are out in the wild right now, alongside hosted fine-tuning and playground access on Tinker.
What it can do
The real punch here is in reasoning density and agentic tool use.
On the text-only, tool-free Humanity's Last Exam (HLE), Inkling-Small notched a 31.6% score, edging out the flagship Inkling at 29.7%. It also cleared 80.2% on SWEBench-Verified and hit 64.7% on Terminal Bench 2.1, showing serious stamina for terminal-based coding harnesses.
Like the larger model, it supports variable test-time compute. You can dial its thinking budget up or down depending on how much latency and cost your task can tolerate.
For multimodal inputs, Inkling-Small relies on an encoder-free native design. Audio feeds directly in as dMel spectrograms, while images are sliced into 40×40 patches and processed by a four-layer hMLP alongside standard text tokens. It can even invoke Python scripts to zoom, crop, and inspect dense charts or blurry text on the fly.
Specs, weights, and how to run it
Thinking Machines pulled this off by revamping its pre-training recipe, running on-policy distillation using the original Inkling as a teacher, and finishing with two weeks of continuous RL focused on agentic coding. The tradeoff is factuality: larger models still beat it on broad knowledge tests like SimpleQA.
Deploying it locally takes serious iron.
Even though active parameters sit at a lean 12B, the total parameter footprint means the downloadable weights clock in at around 532 GB. You will need a multi-GPU cluster or enterprise hardware like stacked DGX boxes to host it yourself.
If you want to kick the tires without downloading half a terabyte, Tinker offers fine-tuning and a web chat playground for text, audio, and vision, while raw weights are hosted on Hugging Face.
Why it matters
If you are building autonomous coding agents or need local audio-visual reasoning, Inkling-Small fills an awkward gap in the open-weights world. Most lightweight open models stumble on native audio intelligence or complex repo-level code edits.
We see this release as clear proof that targeted teacher distillation and tool-use RL can push active-sparse MoEs way beyond their raw active weight class. You lose some encyclopedic fact-checking, but for agentic utility per dollar, it punches far above its weight.
Sources (2)
- thinkingmachines.ai(opens in a new tab)Originalthinkingmachines.ai
- youtube.com(opens in a new tab)Reportyoutube.com