In January 2025 a Hangzhou research company released a reasoning model under the MIT licence, and the industry has been arguing about what it meant ever since. A year later Hugging Face published a three-part audit of what actually changed. I read all three parts; the interesting story is not which model won a leaderboard. It is what the licences and the file sizes did to the rest of us.
I have not run any of these models on my own hardware. This is an assessment built from the three articles themselves and the work they cite, and where a number comes from Hugging Face’s own platform I say so — a platform counting its own ecosystem is not a neutral witness.
What the series actually says
Part one (20 January 2026) is about strategy. Its argument is that R1 lowered three barriers at once. The technical one: by publishing its reasoning traces and how it post-trained the model, it turned advanced reasoning from something locked behind a paid API into an engineering asset you can download, distil and fine-tune. The adoption one: MIT terms meant companies could put it straight into production instead of running a special project. The third was psychological: as the series puts it, the question shifted from whether this could be done at all to how to do it well.
The measured consequence: Baidu published nothing open in 2024 and more than 100 repositories last year, while ByteDance and Tencent multiplied their own output 8 to 9 times. Moonshot’s Kimi K2 was widely called “another DeepSeek moment”. The line worth keeping from the whole series is its own explanation for the pattern: this is not coordination by agreement but alignment under constraint — the same compute, cost and compliance pressures pushing different companies down similar engineering roads.
Part two (27 January 2026) is about architecture, and it is the most technically useful of the three. Mixture-of-experts became the default shape for leading open models — Kimi K2, MiniMax M2, Qwen3 — for a reason that has nothing to do with benchmarks: it spends compute per request rather than per machine, which is what you do when you cannot buy your way out of a shortage. Open activity spread from text into video, audio, 3D and agents, and it shipped as bundles rather than weights alone: inference stacks, datasets, evaluation, toolchains. Small models did the heavy lifting, with 0.5B to 30B models described as simpler to run on ordinary hardware, to fine-tune, and to fold into real systems, while 100B to 700B models sat above them as teachers to be distilled down. Apache 2.0 became the near-default licence. And the hardware story turned: DeepSeek’s V3.2-Exp shipped with day-zero support for Huawei and Cambricon chips as reproducible inference pipelines rather than demos; Ant Group reported training domestic chips at roughly 20% less cost per trillion tokens; Baidu documented training on more than 5,000 of its own accelerators; two early-2026 models were announced as trained entirely on domestic silicon.
Part three (3 February 2026) is about who is doing it. DeepSeek is the most-followed organisation on Hugging Face and Qwen the fourth. Qwen is the base for over 113,000 models, against Llama’s 27,000 and DeepSeek’s 6,000. Baidu — whose chief executive had been openly sceptical of open source — reversed course through free access and an open model family. Application companies that once only consumed models began training their own. And the state context is explicit: the “East Data, West Compute” plan with 8 hubs and 10 data centre clusters, around 1,590 EFLOPS of total compute, an average power usage effectiveness near 1.46, and an August 2025 “AI+” plan that shifted policy from building foundations to large-scale deployment.
Why this matters if you own a computer and not a datacentre
Four things, and none of them needs a policy briefing.
Permissive licences are the whole point. MIT and Apache 2.0 are why R1 and its descendants ended up on ordinary machines instead of behind an API. A model whose weights you may look at but not modify, fine-tune or redistribute is not open in any sense that helps a person. If you take one idea from the series, take the licence.
Small models are the real beneficiaries. A pyramid with a few enormous teachers on top and a wide layer of distilled 0.5B to 30B models underneath is exactly what a home machine needs. The intellectual work happens at the top; the usefulness lands at the bottom.
Shipping for your own silicon is the sovereignty lesson. The day-zero support for domestic accelerators, delivered with the weights and the pipeline rather than as a promise, is the part I would hand to anyone building infrastructure. Whether you run NVIDIA, AMD or something else, a model that arrives with a reproducible way to run it beats a model that arrives with a score.
Hardware policy reaches your desk. Export controls and who may buy which accelerator decide what a household can run, not just what a cloud provider can rent. The series leaves the H200 question open, and it is the right question.
The case for taking the series seriously
In one year, a genuine second source of open weights appeared, competing on what matters to a self-hoster: cost and reproducibility rather than peak scores. Licences got easier, not harder. Serving stacks, quantisation formats and deployment tooling were open-sourced alongside the models, which raises the floor for everybody. And small models proliferated — the part that reaches a person rather than a procurement department.
The case against — and this is where I get sceptical
It is published by the platform that hosts the goods, using its own platform’s numbers. Every statistic about derivatives, downloads and popularity is Hugging Face measuring Hugging Face. That does not make it wrong; it makes it a well-argued case rather than an audit, and the difference is the reason to read the citations.
“Open” here means weights. Not the training data, not the full pipeline, and not reproducibility of the pretraining run. Those are three different claims and the articles sometimes let the word cover all of them at once.
It did not happen by accident. The compute hubs, the power build-out and the “AI+” plan all pre-date R1 — the plan’s date is 2017. The series is candid about this, and the candour is to its credit; but a story about a coordinated national build-out reads very differently from the story about a scrappy underdog that the model release inspired elsewhere.
And the admission that matters most sits in a subordinate clause: compute is still the binding constraint, with one named company reportedly rationing usage during a crunch. A year-long account of an ecosystem that cannot get enough chips is not a victory lap, and the series does not quite say that out loud.
Where this is likely heading
My judgement, not a forecast. Two tracks keep diverging: a closed frontier sold by the token, and open weights under permissive licences that get better every quarter without becoming cheaper to train. The local tier will keep improving mostly through distillation rather than new architecture — which is good news for anyone planning to buy a single used accelerator rather than a rack. “Open” will keep being stretched by marketing until it means whatever the speaker needs it to mean, which is why the licence line in a model card will keep being the only part that is worth trusting. And the fight will run through silicon and export paperwork, because that is where the constraint actually lives.
The thing I would watch over the next 12 months is not a leaderboard. It is whether permissively licensed models in the 1B to 30B range keep arriving at the same pace. That is the part that decides whether a person with a desktop and a graphics card has real choices, or a choice between one vendor’s cloud and another’s.
What you can do now
Check the licence before you download anything, and treat a bespoke licence as friction you are being asked to absorb. Prefer the small distilled variants of large models when you are running locally; they exist precisely for that. Size your expectations against your own memory and disk rather than against a benchmark table. And when you read platform statistics — mine included — ask who runs the platform.
The three parts are worth reading in full, and they are unusually well-sourced for the genre. A series written by the host of the ecosystem it describes is still the best account of that ecosystem I have found — provided you read it with your scepticism intact.
That is the same instinct that makes you read a licence before you accept it.