Most AI labs that brag about their reasoning models keep the recipe locked away. The weights get released, maybe. The training data and the code that produced them almost never do. Kunlun Tech, the Beijing company behind the TianGong (天工) assistant, took the opposite bet.
In early 2025 the company shipped a family of reasoning models and, unlike most of its peers, walked the whole kitchen out into the open.
Key takeaways
- Skywork-OR1 is a series of three open-source reasoning models aimed at math and code, the two tasks where "showing your work" matters most.
- Kunlun released not just model weights but also the training dataset and the full training code on GitHub and Hugging Face.
- The move extends Skywork-o1, which the company says was the first Chinese logic-reasoning large model (大模型) when it appeared in February 2025.
- For ordinary developers, it means a capable reasoning model they can download, audit, and fine-tune without asking permission or paying a license.
What a "reasoning model" actually is
A normal chatbot answers fast and often confidently wrong. A reasoning model is trained to slow down, break a hard problem into steps, and check itself before answering. Think of the difference between a student who blurts an answer and one who writes out the proof.
This matters most in math and programming, where a single logical slip breaks everything. That is exactly where Kunlun pointed the OR1 series.
Three models, one idea
The April 2025 release came as three siblings:
- Skywork-OR1-Math-7B — a math specialist that also writes decent code.
- Skywork-OR1-7B-Preview — a generalist blending math and code ability.
- Skywork-OR1-32B-Preview — the larger flagship for harder problems.
The "7B" and "32B" refer to the number of parameters, the dials the model tunes while learning. A 7-billion-parameter model is light enough to run on a single serious workstation; the 32B version wants more hardware but reasons more reliably.
The part competitors usually hide
Here is the unusual bit. Kunlun did not just post the finished weights. The team published the training dataset and the complete training code, with a technical write-up explaining how the data was prepared and what they learned. The stated goal was full reproducibility — another lab should be able to walk the same path and get a similar result.
That is rare. Most open releases are "here is the model, good luck figuring out how we built it." Publishing the data and code turns a product drop into a teaching exercise, which is why the OR1 release got attention inside the research community even though Kunlun is not one of the usual China AI headline names like DeepSeek or Qwen.
How they measure "good at thinking"
Kunlun leaned on two well-known benchmarks: AIME for competition-level math and LiveCodeBench for real programming problems. More interesting is the metric they pushed, called avg@k. Traditional pass@k asks "did the model solve it at least once in k tries?" avg@k instead averages how often it succeeds across those tries, which the team argues better reflects steady, useful performance rather than a lucky shot.
The company framed OR1 as competitive with much larger models at the same size, but readers should treat head-to-head claims against named rival models with caution — those comparisons come from the lab itself, not an independent audit.
Why a company would give this away
Open-sourcing a reasoning model is not pure charity. It builds developer trust, seeds a community that files bugs and writes tutorials, and positions Kunlun's commercial TianGong products as the "supported" option next to the free research release. It is the same playbook open-source software has used for decades: give the engine away, sell the service.
For China's broader AI scene, releases like this chip away at the idea that the best reasoning research lives only behind corporate walls in the US. A model anyone can download and inspect is a model a university in Nairobi or a startup in São Paulo can actually use.
Honest limitations
The reasoning-quality claims rest on Kunlun's own benchmarks and blog posts, not a neutral third-party evaluation, so independent verification is thin. The 32B model still needs real GPU hardware and is not something you run on a laptop. "Open" here means open weights and code, not open training compute — reproducing it from scratch would still cost serious money. And reasoning models, even good ones, remain prone to confident errors on questions outside math and code, where the step-by-step training helps less.
What readers can do now
- If you build software, pull the OR1 weights from Hugging Face and test them on your own math or coding tasks instead of trusting marketing numbers.
- If you teach or study AI, use the released training code as a rare concrete example of how a reasoning model is actually built.
- If you just use chatbots, watch for "reasoning mode" toggles in products — OR1 is a sign that careful step-by-step thinking is becoming a free, standard feature rather than a premium one.