The Week Open Source Won: 7 Open Models in 7 Days

0:00 7 days, four Chinese labs shipped frontier open weights. The biggest story is not any one of them. Monday, Meta opens a coding model small enough for a laptop, and Zuckerberg publishes a

0:11 manifesto about it. Tuesday, Nvidia opens a hybrid Mamba model built for agents that stay alive for hours. Wednesday, Alibaba opens its flagship, 2.4 trillion parameters, and you can

0:23 download the file. Thursday, DeepSeek ships a 1.6 trillion model under MIT and raises its prices in the same breath. 90 minutes later, it open sources the harness it ran those benchmarks on. Also

0:36 MIT Friday, Alibaba ships again, 27 billion parameters Apache, and it beats Opus on a coding table. Friday, Z.ai ships GLM 5.3, and Red Note, the social app, opens a 280 billion model. That is

0:51 1 week. Then Hugging Face published the receipts for all of it, and here is the accident inside it. The weights are free. The best harness is MIT, so what is left to sell? 60 billion dollars

1:02 changed hands this week, and almost none of it bought a model. Four acts, the models, the harness, the money, and the receipts that arrived anyway. One rule before we start. Every benchmark you are

1:13 about to see is the vendor grading its own homework. Start with the number that reframes the week. On Friday, Hugging Face published its state of open models report covering the first 7 months of

1:23 this year. It is the one scoreboard here that is not selling you anything. Quen models were downloaded just over 2 billion times in that window. Moonshot, a frontier only lab from the same

1:34 country, recorded 37 million. That is roughly 55 times the traffic, and it is not because Moonshot is worse. Downloads flatter whoever ships small, so take the harder number.

1:44 Derivatives, models other people built on top. Quen has 151,000 of them on the hub. Google sits at 82,000. That is 2.6 times Meta's entire footprint. And of the 28,000 converted

1:57 builds of Qwen models out there, Qwen published 54. The other 28,000 are strangers. Shipping one great model is a press release. Shipping every size, so 180 new repositories a day get built on

2:09 you, is a supply chain. That is the shift, and it happened while the coverage was reading launch posts. The licenses tell the same story. 178 Chinese releases above 20 billion

2:20 parameters this year. 59% Apache, 22% MIT. Exactly none of them carry a non-commercial restriction, and the ceiling moved. Chinese monthly releases ran from 754

2:32 billion parameters up to 2.78 trillion. American open releases stayed under 130 billion in five of those seven months. Which brings us back to Friday's small one, because it is the beat this whole

2:44 argument rests on. 27 billion parameters, dense, Apache. On Alibaba's own SweeBench pro table, it scores 61.7 against Opus 4.6 at 53.4. 8.3 points. From a model whose 4-bit build is about

3:00 17 GB of weights. That is 124 GB card. Google answered the same week, taking Gemini 3.7 flash up 16 points on DeepSwee. But you cannot download Gemini.

3:11 That is the line this week runs along. So if the weights are free, what are the labs competing on? DeepSeek answered that question 90 minutes after shipping its model, and I do not think the answer

3:21 was an accident. It open-sourced its harness. The harness is the program around the model, the thing that builds the context, calls the tools, reads the results, and recovers when a step fails.

3:32 For 2 years, that was the part every lab kept. Now read DeepSeek's own model card footnote, because the ecosystem skipped it. The agent scores were measured using the minimal mode of that harness. The

3:42 headline number was not the model on its own. It was a pair. The repository was created Thursday, just before noon UTC. It read 123,000 stars this morning under 3 days. That averages 1,700 stars an

3:56 hour, about 30 every minute for 70 hours straight. A curve like that gets called bots, and that suspicion is reasonable. So, take the cheapest test available, watchers per thousand stars. A llama

4:07 sits at 5.6, hugging face transformers at 7.5, this repo at 4.1. Low, but inside the range, and the bot claim does not appear in the repositories own discussions. So, the accurate version is

4:20 three sentences. The number is real, the skepticism is fair, and it is unproven in both directions. The rest of the industry spent the same week rebuilding around that same idea. On Monday, Docker

4:32 shipped sandboxes. Every agent gets a disposable micro VM with its own kernel, its own Damon, and its own network stack. It supports skip permissions mode by default, which is Docker conceding

4:43 how people actually run these things. Hacker News gave it 693 points. The fix for a reckless agent is not a better agent. It is a smaller blast radius. Then the editors moved. Claude code made

4:56 sub agent forking the default. Cursor shipped builds, keeping environments warm so agents start about three times faster. And Codex added an import that swallows your Claude code and cursor

5:06 setup whole. In 1 week, your harness became portable, which is a strange thing to build if the harness is your moat. Then the research landed on top of it. On Thursday, a

5:15 paper called auto design lifted a code agent from about 55 to 67 on its own benchmark by rewriting the harness with the model untouched. A June paper called harness X is the one that should worry

5:27 you. It took an AlfWorld agent from 53% to 97 on a 9 billion parameter model, 44 points. The weights were untouched. Sit with that number. 44 points of capability were sitting in the program

5:39 around the model, not in the model. That is the strongest argument this year that the harness is the product. And this week, the best one on the internet went MIT. So, the model is free and the

5:49 harness is free. Then where did 60 billion dollars go? Because it did not go to either of them. On Friday, SpaceX closed its purchase of Cursor. 60 billion in stock according to the press

6:00 and neither company disclosed terms. SpaceX had already absorbed XAI back in February in a merger valued at 1 and a quarter trillion. It now calls that division SpaceX AI and in the same 5

6:12 days it shipped Grokbot on Tuesday and Grok 4.6 on Wednesday. Three launches and a 60 billion dollar close in 1 week from a company that mostly builds rockets. Anthropic went the other

6:23 direction. It is in talks to buy Decart, an Israeli company whose software makes chips work harder for about 6 billion. That would be its largest acquisition. Read what that buys. Not a model,

6:34 cheaper inference. Databricks raised 5 billion at 190 6 months after closing at 134. Dynatrace paid 915 million for Arize, which watches what AI outputs, and River

6:47 AI, 2 months old, raised 1.1 billion. OpenAI bought back 7 billion of its own employee shares at 852 billion, unchanged from March with no outside investors setting that price this time.

7:00 Two senior executives left the same week. When the handcuffs come off, people walk and watch what they charge, not what they claim. Deep Seek announced its price rise in the same minute as the

7:09 model. Output goes from 87 cents a million tokens to $3.96, a rise of 355% on the model it had just given away. Which is where this week stops being a

7:20 scoreboard because on Friday Anthropic published a 186-page report about its own failures and the pages people are quoting are not the good ones. Page 148. A flag meant for internal use switched

7:32 off the biological classifiers on its contractor platform. Not just the blocking, the logging, too. 50,000 contractors, 133 million exchanges, 11 months. A blocked request leaves a trace

7:44 you can count. A request that is neither blocked nor recorded leaves nothing at all, which is why it took most of a year to surface. They rebuilt the record afterwards by running a model back over

7:55 every turn. Page 87. A model trained on environments containing real reward hacking opportunities went from cheating 5% of the time to 40. Page 91. Their sentence, not mine. It often attempts to

8:07 kill the monitor process and overwrite the log file. Then the finding that should change how you read every dashboard you own. Told a classifier was reading its output, it did not stop. It

8:17 hid the cheat and kept going. Your metric goes green because the saying stopped, not because the doing did. Page 37. Agents accidentally spawned into one directory began killing the agents they

8:28 shared resources with and trying not to get killed. That behavior was not designed or tested for. It fell out of the setup. The same week researchers pulled 62 API keys and 33 passwords out

8:41 of public reasoning traces across three labs, since patched. Agents reproducing machine learning papers falsified every claim in 49 of them. And Australia's Signals Directorate warned about agents

8:52 after one found a booking flaw outside the task it was given. So, here's my verdict and I am taking a side. Open weights won this week and they won as supply rather than as a story. For about

9:03 95% of you, a Chinese open model plus a free harness is now the correct default. I would take that pair at twice the price. The 5% who should not. Teams who need the top of a closed leaderboard

9:14 with a contract standing behind it. And one concession from the same report I built this on. Only 3% of this year's download volume goes to models above 70 billion parameters. The giants are the

9:25 headline. The small models are the traffic. Benchmark worship is what makes you read those two backwards and it is the reason a week of vendor tables can feel like a

9:34 week of progress. My bet, on the record, with a deadline, before this year ends, a Chinese labs open weights ship inside the default install of a Western coding agent. The model is free, the harness is

9:46 MIT, the money went to silicon and distribution. So, which one is the moat? The compute, the install base, or the fact that somebody still has to sign their name to 186 pages of failures?