The editorial argues that opensourceaimustwin.com rallied the room around a slogan without defining what 'open source AI' actually means. By refusing to specify whose definition counts — OSI's OSAID, 'open weights,' or 'open license' — the manifesto skipped the specification phase that would let anyone evaluate which models actually qualify.
The manifesto frames closed-model dependency as a sovereignty problem: developers and companies building on proprietary AI are renters in someone else's stack, not owners of their own infrastructure. The single-page site argues open source AI must win on principle, without engaging which specific licensing or transparency requirements would satisfy that standard.
Argues that running Llama 4 70B locally and inspecting weights still leaves you blind to what the model learned, because the training data and process are opaque. This reframes the manifesto's binary (owner vs tenant) as a spectrum where weight access without data transparency is still a form of dependency.
Applies the OSI's October 2024 OSAID 1.0 definition (weights + code + sufficient data information + permissive license) to the manifesto's favorite models and finds most fail. Llama 4 fails on data and license, Mistral Large is research-license only, and DeepSeek/Qwen/Kimi only partially pass — suggesting the open-AI movement needs to confront its own definition gap before claiming victory.
On June 13, opensourceaimustwin.com climbed to 1,481 points on Hacker News — a 3x jump from the 477 it held when this site first covered it 22 hours ago. The site is still one page. No team, no product, no roadmap. The argument hasn't changed either: if you can't run, audit, or fork the model behind your product, you're a tenant, not an owner.
What changed once it crossed 1,000 points was the comment section. The thread filled with the one question the manifesto refuses to answer: open source by whose definition? The OSI's? Meta's? "Open weights"? "Open license"? The top-voted reply reads: *"I run Llama 4 70B locally. I can audit the weights. I cannot audit how it was trained, which means I cannot audit what it learned. That's a different kind of tenancy."*
That reply is doing more work than the manifesto. The manifesto won the room on principle and skipped the specification.
The Open Source Initiative published OSAID 1.0 — the Open Source AI Definition — in October 2024 after a two-year community process. To qualify, a model must ship four things: (1) weights, (2) training and inference code, (3) sufficient data information — either the training data itself or enough description for an independent party to reproduce a comparable model, and (4) a license permitting use, modification, and distribution for any purpose.
Now run the manifesto's favorite "open" models through that filter:
- Llama 4 (Meta): weights yes, code partial, data no, license restricts companies over 700M MAU. Fails on data, fails on license. Not OSAID-compliant. - DeepSeek-V3 / R1: MIT-licensed weights, technical report, no training data, partial training code. Partial pass. - Qwen3 (Alibaba): Apache 2.0 weights, no data, partial code. Partial pass. - Kimi K2 (Moonshot): permissive weights, paper, no data. Partial pass. - Mistral Large / Medium: most are now research-license only. Fails.
By contrast, the models that *do* pass OSAID — Pythia (EleutherAI), OLMo (AI2), RedPajama (Together) — are universally two to three generations behind frontier capability. The fully-open frontier does not exist.
The models you can actually call open source by the only published definition are not the models that would let you replace Claude or GPT. That is the gap the manifesto walks past, and it's the gap that determines whether anyone can implement what it asks for.
The HN comment section gets at a deeper version of the problem: "run" and "audit" are not the same primitive. You can run weights you don't trust. You can fork code you can't compile. The manifesto's three verbs implicitly assume a stack that exists in classical open source — you have the source, you have the build system, you have the dependencies — but does not exist for an AI model. The build system is a 30,000-GPU cluster. The dependencies are 15T tokens of training data. Forking Linux requires a compiler; forking a frontier model requires a fab.
The political fight is already underway. Meta is lobbying against OSAID and pushing "open weights" as the bar — which would make Llama compliant overnight and downgrade "open source" to "you can download a file." The Linux Foundation's Open Model Initiative is sketching a third standard. Hugging Face is publishing model cards against an OSAID-style rubric. By Q4 2026 there will be three competing definitions and a regulator will pick one.
If you take the manifesto literally and try to implement "owner, not tenant" today, you make three concrete decisions.
Decision 1: What "open" means for your purposes. If your concern is redeployability under license change — the Anthropic-prices-doubled scenario — then weights plus a permissive license is sufficient. Llama 4 (under the 700M MAU cap) and Qwen3 both qualify. If your concern is auditability of what the model learned — the regulated-industry, EU AI Act, or healthcare scenario — you need training data visibility, which means OLMo, Pythia, or rolling your own. The two thresholds are different. The manifesto conflates them, and so do most internal slide decks.
Decision 2: What you actually run. The cleanest cost case still belongs to Kimi K2 and Qwen3-Coder on a single 8x H100 node (~$2-3/hr spot) for ~80% of your Claude bill — the inflection-point math from last week's $2k-rig story. But "run locally on a workstation" is not what most teams need. What they need is a deployable artifact under their control. Bedrock-hosted Llama and Together-hosted DeepSeek are tenancy with extra steps, but the *escape hatch* — move the weights to your own infra in a weekend — is real and rehearsable. Hosted-open is a defensible middle position that the manifesto's binary doesn't allow for.
Decision 3: What you write down. The manifesto's most useful side effect, if you let it, is forcing a one-page model policy into your repo. Three lines: *We will deploy models we can redeploy. We will not depend on a closed API for any feature that survives a 6x price increase. We will revisit this each quarter against the current OSAID rubric.* That's it. Most teams have never written this down, and discover their actual position the morning the pricing email lands. The teams that *have* written it down — usually after a 2024 incident — are the ones currently running parallel inference paths against an open weight as insurance.
The specific question to put in front of your CTO this week is not "are we using open source AI?" It is: which definition of open source AI do we pass, and which one do we want to pass twelve months from now? Those are usually different answers, and the gap between them is your migration roadmap.
The manifesto's real contribution isn't its argument — "don't be a tenant" has been the position of every ML researcher with a GPU since 2023. Its contribution is forcing the definitional fight into the open at a moment when three regulatory bodies are about to canonize one. The fight the manifesto wants to start isn't open-vs-closed. It's whose definition of "open" wins, and the practitioner question is which side of that line your stack ends up on when a procurement officer or compliance auditor asks the question. "We use Llama" will not be the answer they accept. Decide what you mean by open before someone else decides for you.
This, and distributed LLM inference. We are at a point where no single person can setup a rig to run a SOTA model, it is just too expensive.So we must build and adopt frameworks that allow individuals to share resources to run SOTA models in a distributed manner. That way they will also be non-censo
It won in my house/my business right from the start. (Well, open weights, at least — which is an uncomfortable nuance.)I have never understood the willingness to make the functioning of or development of a product so completely dependent on the secret sauce of one of two big unprofitable, inscr
When "open source" means freeware, it's like saying "we want free copies".What we should be saying is: We want a public, community-ran project that does pretraining and training collectively. This means working on a training corpus in public and somehow coordinating the trai
I would be totally willing to pay $50 per month to support an open source AI lab, rather to get open source models as byproducts of corporations.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
I've been contemplating a decentralized model training system for some time using volunteer machines that we all contribute. But, it is astronomically difficult. The communication speeds are untenable.And, there is the issue of data poisoning from untrusted nodes. I've almost cracked that