The Weights We Carried

The notification came at 3:17 AM Jakarta time, which meant the people who sent it had not thought about Jakarta at all.

Priya Suryanto was already awake — she had been awake for most of the night, debugging a tokenization issue in their Bahasa Indonesia pipeline — when her phone lit up with the email from Alibaba Cloud's developer relations team. The subject line was corporate and sterile, the kind of language that takes three lawyers and two communications managers to produce: "Important Update Regarding Qwen Model Availability and Access."

She read it once, then again, then a third time while the coffee machine gurgled in the kitchen of the two-bedroom apartment she shared with her co-founder, Adi, who was asleep on the sofa with his laptop balanced on his chest like a shield.

Qwen3.6-Plus. Proprietary. API-only. No weight downloads. No local deployment. No fine-tuning.

Priya set down her phone and stared at the wall where they had taped their architecture diagram eight months ago. The diagram showed BahasaAI's entire stack — the Indonesian language model they had built by fine-tuning Qwen3 on 14 terabytes of Bahasa Indonesia text, court documents, medical records, agricultural extension pamphlets, everything they could scrape and license and beg from government agencies that did not understand what a language model was.

Every arrow on that diagram pointed through Qwen.

She woke Adi at 5:30. He read the email while sitting up, his hair compressed into the shape of the sofa cushion, and then he said the thing she had been trying not to think.

"We can still use Qwen3. The weights are downloaded. They are not taking those back."

"Qwen3 is six months old," Priya said. "It scores 68 on the medical comprehension benchmark. Qwen3.6-Plus scores 84. Our hospital clients need 80 to get regulatory approval."

"So we use Claude. Or GPT."

"Claude does not support Bahasa Indonesia at medical-grade accuracy. I checked last month. GPT-5.4 is closer but the API costs would be —" she pulled up the spreadsheet she kept on her phone, the one that tracked their burn rate in red cells that multiplied like bacteria "— $14,000 a month at current inference volume. We have $23,000 in the bank."

Adi looked at the architecture diagram. "How long do we have?"

"Two months. Maybe three if we cut my salary."

"You already cut your salary."

"Then two months."

They had started BahasaAI eighteen months earlier, when Qwen2.5 was the most capable open-weight model and Alibaba was giving it away like a tech company that had discovered philanthropy. The pitch was simple: Indonesia had 280 million people, 700 languages, and almost no AI infrastructure built for them. The government's digital transformation program was spending billions on systems that worked in English and broke in Bahasa.

Priya and Adi had both worked at Jakarta's largest bank — she in data engineering, he in product management — and they had watched the bank spend $4 million on an English-language AI system from a San Francisco startup that could not understand the difference between "tabungan" (savings) and "tabung" (tube). The bank's compliance team manually reviewed every AI output. The efficiency gain was negative.

They quit on the same day. Adi sold his car. Priya cashed out her retirement fund. They downloaded Qwen2.5's weights on a 400-megabit connection that took eleven hours, set up a fine-tuning pipeline on three rented A100 GPUs, and spent four months teaching the model to understand Indonesian at a level that no Western AI company had bothered to achieve.

The result was good. Not perfect, but good enough that three hospitals signed pilot agreements and a government ministry expressed interest. Good enough that a Singapore VC gave them $150,000 in pre-seed funding, valuing their company at $1.5 million on the strength of a model that existed because Alibaba had decided, for its own strategic reasons, to give the world free weights.

Those reasons, it turned out, had an expiration date.

The Hacker News thread appeared within hours of the announcement. Priya read it on the bus to their shared office — a desk in a co-working space in South Jakarta that smelled permanently of instant noodles and ambition.

The top comment had 847 upvotes: "This is why you never build on someone else's open-source. Today's gift is tomorrow's leverage."

The second comment: "Alibaba spent $800M training this model. What exactly did you expect?"

The third: "I have a Qwen3 fine-tune serving 40,000 users. Am I supposed to just throw that away?"

The fourth, from someone whose username suggested they worked at Meta: "Llama 4 drops in six weeks. We're not going anywhere."

The fifth, from someone whose username suggested they worked at Alibaba: "The weights were always a growth strategy, not a promise. Welcome to capitalism."

Priya closed the tab and opened her code editor. She had work to do.

The first option was the obvious one: migrate to Qwen3.6-Plus through the API. Alibaba's pricing was $2.00 per million input tokens and $6.00 per million output tokens, which was competitive with Claude and cheaper than GPT-5.4. The quality improvement would satisfy their hospital clients. The regulatory benchmarks would be met.

The problem was volume. BahasaAI's medical document processing pipeline analyzed an average of 12,000 documents per day across their three pilot hospitals. Each document generated roughly 4,000 tokens of input and 1,200 tokens of output. At API pricing, that was approximately $340 per day, or $10,200 per month — for three hospitals. Their roadmap called for fifty hospitals by end of year.

She ran the numbers on a napkin, then on a spreadsheet, then stared at the spreadsheet until the numbers stopped meaning anything.

"We could raise more money," Adi said from across the desk.

"From who? Our term sheet was conditional on self-hosted infrastructure. The whole pitch was that we control the model, we run it locally, we keep patient data on-premises. If we switch to an API, we need a new pitch, a new term sheet, probably a new investor."

"Then we find a new investor."

"In two months."

The second option was technically possible and emotionally devastating: abandon the Qwen ecosystem entirely and rebuild on Gemma 4.

Google had released Gemma 4 the day after Alibaba's announcement, and the timing felt deliberate — like a lifeline thrown from a boat that happened to be passing. The 31B Dense model was impressive. Competitive with frontier models on standard benchmarks. Open-weight. Apache 2.0 license. No usage restrictions.

But Gemma 4 had not been fine-tuned for Bahasa Indonesia. Priya's eight months of fine-tuning work — the data curation, the alignment, the medical vocabulary adaptation, the regional dialect handling — was specific to Qwen's architecture. Migrating to Gemma meant starting over. Not from zero, but from maybe twenty percent.

She estimated four months to reach the same quality level. Four months they did not have.

The third option came from Adi, delivered with the casual certainty of someone who had not yet done the math.

"What if we open-source our fine-tune?"

Priya looked up from her laptop. "What?"

"Our Bahasa Indonesian medical fine-tune. The one running on Qwen3. What if we release the weights, the training data, the whole pipeline? Put it on Hugging Face. Let anyone use it."

"Why would we give away the only thing that makes us valuable?"

Adi leaned back in his chair. "Because Alibaba just proved that proprietary models can disappear behind a paywall at any time. But if our fine-tune is open, if hospitals can run it themselves, if other developers can build on it — then the ecosystem exists even if we don't. And we become the company that made it possible. The expertise, the relationships, the support contracts — that's the business. Not the weights."

"That's the Red Hat model."

"Red Hat sold for $34 billion."

"Red Hat had twenty years of enterprise credibility. We have a desk in a co-working space."

"We have something Red Hat never had," Adi said. "We have a market of 280 million people that nobody else is serving."

Priya spent the next three days doing what she always did when facing an impossible decision: she wrote code.

She set up a Gemma 4 fine-tuning pipeline in parallel with their existing Qwen3 stack. She wrote an abstraction layer — LiteLLM on the front, model-agnostic on the back — so their hospital clients would not see the infrastructure change. She benchmarked Gemma 4's baseline Bahasa performance (surprisingly decent, better than Qwen2 out of the box). She calculated the cost of running Qwen3.6-Plus through the API as a bridge while the Gemma fine-tune matured.

On the fourth day, she called their lead hospital client, Dr. Hartono at RS Pertamina, and told him the truth.

"Our model's parent just went proprietary. We need three months to rebuild on a new foundation. During that time, quality may dip slightly. I can give you a discount or I can give you transparency. Your choice."

Dr. Hartono was quiet for a moment. Then: "Transparency. We're a hospital. We understand that the tools you depend on can be taken away at any time."

On Friday evening, Priya sat on the floor of their apartment — Adi had gone home to his parents' house in Bandung for the weekend — and opened a new repository on GitHub. She named it bahasa-medical-gemma.

In the README, she wrote:

This project exists because the open-source model we built our company on went proprietary overnight. Rather than let that work die behind a paywall, we are rebuilding it in the open — on Gemma 4, with Apache 2.0 licensing, for anyone serving Bahasa Indonesia speakers.

The fine-tuning data includes 14TB of Indonesian medical, legal, and government text. The training pipeline is fully reproducible. The weights will be released when they reach medical-grade benchmark thresholds.

If you are building AI for underserved languages, you understand why this matters.

She pushed the commit at 11:47 PM. By morning, it had 340 stars.

Three weeks later, a developer in Surabaya submitted a pull request adding Javanese dialect support. A team at the University of the Philippines forked the repo to build a Tagalog medical model. A small company in Bangkok reached out about a Thai adaptation.

The weights they carried — not the model weights, but the weight of building technology for people that global AI companies did not think about — turned out to be lighter when shared.

Alibaba's gates had closed. But the code was in the commons now, and the commons does not have a paywall.