In February 2026, six open-weight models crossed 75% on SWE-bench Verified — all released within weeks of each other, all under permissive licenses. A 3-billion-active-parameter model hit 70% on the hardest coding benchmark in AI. The economics shifted overnight: frontier-quality coding at one-thirteenth the cost. Open-source didn't just catch up. It arrived.
Then the team behind the most important model fell apart.
This is the story of both.
Act I: The Achievement
Let's start with what happened, because the numbers are genuinely remarkable.
In the span of six weeks (late January to early March 2026), the open-source AI community shipped an unprecedented wave of coding-capable models:
- Qwen 3.5 — 397B total / 17B active MoE. Apache 2.0. 76.4% SWE-bench Verified. 256K context. 201 languages. 13x cheaper than Claude Opus 4.6 per token.
- GLM-5 — 744B MoE. MIT license. 77.8% SWE-bench Verified. Built on Huawei chips, not NVIDIA. $1 per million input tokens.
- MiniMax M2.5 — 230B. Open-weight. 80.2% SWE-bench Verified — higher than GPT-5.2.
- DeepSeek V3.2 — 685B MoE. MIT license. The Speciale variant surpasses GPT-5 on reasoning benchmarks.
- Llama 4 Maverick — 400B/17B MoE. Agentic focus. Distilled from a 2-trillion-parameter Behemoth model.
- Xiaomi MiMo-V2-Flash — 309B/15B MoE. 256K context. Hybrid thinking mode. Open-source reasoning and coding.
Six S-tier open-weight models. Every one of them under a permissive license. Every one of them capable of real-world software engineering.
But the Qwen family deserves special attention, because they didn't just release one good model — they released an entire ecosystem.
The Qwen Ecosystem
The Qwen 3.5 flagship at 76.4% SWE-bench was impressive enough. Then came the medium series: 27B dense, 35B-A3B, 122B-A10B. The 122B model beats GPT-5 mini on function calling (72.2 vs 55.5 on BFCL-V4).
Then the small series: 0.8B to 9B parameters. The 9B model beats GPT-oss-120B on GPQA Diamond — a model 13x its size — scoring 81.7 vs 80.1. The 2B model runs on an iPhone in airplane mode with 4GB of RAM.
Then Qwen3-Coder-Next: 80B total, 3B active parameters. 74.2% SWE-bench Verified. 70%+ with SWE-Agent. It runs on a 64GB MacBook, an RTX 5090, or an AMD 7900 XTX. A model you can run on consumer hardware, hitting scores that would have been state-of-the-art six months ago.
And finally, Qwen Code: an open-source agentic coding CLI, Apache 2.0, that works with any model provider. 1,000 free requests per day. The open-source Claude Code alternative the community had been waiting for.
All of this from one team. Over 1 billion downloads on HuggingFace. 180,000+ fine-tuned derivatives. A 554% surge in monthly active users for the Qwen app, from 31 million to 203 million in a single month.
The Economics
Here's why this matters beyond benchmarks: the cost gap is enormous.
Qwen 3.5 delivers 76.4% SWE-bench at roughly $0.19 per million input tokens and $1.05 per million output tokens. Claude Opus 4.6 delivers 80.8% at $15/$75. That's not a small difference — it's 13x to 70x cheaper depending on the metric.
For many production use cases, the quality difference between 76% and 81% doesn't justify a 13x price premium. Especially when you can fine-tune the open model, run it on your own hardware, and keep your code private without sending it to an API.
As one analysis put it: "The economic case for open-weight models in software engineering and agentic workflows is now undeniable."
The Honest Caveat
Open-source isn't winning everywhere. InfoWorld's review of Qwen Code called it "good but not great" — Claude Code still wins on quality, reliability, and fewer iterations needed. VERTU's analysis found Qwen 3.5 "cratering" on complex coding tasks and hallucinating commands in tool-calling. The scaffolding gap — the difference between a raw model and a polished agent — still favors the closed-source tools with their massive engineering teams.
But the trajectory is clear. The gap is months, not years. And on pure economics, open-source already wins.
Act II: The Crisis
On March 3, 2026 — the same day Alibaba released the Qwen 3.5 small models — Junyang Lin posted five words on X:
"me stepping down. bye my beloved qwen."
Lin was the technical lead of the entire Qwen project. At 32, Alibaba's youngest P10 — their highest individual contributor rank. The architect of the open-source strategy that made Qwen the most downloaded model family on HuggingFace.
Alibaba's stock dropped 5.3% — the biggest intraday loss since October.
But Lin wasn't alone. Yu Bowen, head of post-training, resigned the same day. Hui Binyuan, the Qwen Code lead, had already left in January — joining Meta. Kaixin Li, a core contributor to Qwen 3.5, Qwen-VL, and Qwen-Coder, also departed. And according to multiple reports, many young researchers resigned the same day as Lin.
Three senior leaders gone in ten weeks. The team that built the most important open-source AI models in the world — dismantled.
What Happened
The trigger was organizational. Alibaba restructured its Tongyi Lab, splitting the Qwen team's vertically integrated structure — where the same group handled everything from pre-training to post-training to deployment — into horizontal modules under central management. A new researcher, reportedly from Google's Gemini team, was put in charge.
Lin had championed the vertical integration model. He believed tight coordination between pre-training, post-training, and deployment was essential for building great models. The reorganization dissolved exactly the structure he'd built and defended.
Wenting Zhao, an AI researcher, called it "the end of an era." Chen Cheng implied the departure was sudden — the team had been working on model launches hours before.
On the same day, Jack Ma, Joe Tsai, and Daniel Zhang held a rare gathering at Hangzhou Yun Valley School. Ma's first notable public appearance in years. The timing speaks for itself.
What's at Stake
Over 1 billion cumulative downloads. 180,000+ fine-tuned derivatives. Thousands of production applications built on Qwen models. The Apache 2.0 license on existing models is legally locked — those models remain free forever. But future releases? Uncertain.
Community fears are concrete: will Alibaba retreat to a "walled garden" behind paid APIs? Will the next generation of Qwen models be open at all? Alibaba Group CEO Eddie Wu's March 5 statement committed to continuing model development, but said nothing specific about open-source.
The person who championed openness is gone. The person who built Qwen Code is at Meta. The person who led post-training resigned on the same day. The young researchers who executed the vision are scattering.
Act III: The Question
Here's what makes this story complicated: the code is open, but the knowledge isn't.
You can download every Qwen model ever released. The weights are there. The architecture papers are published. Thousands of developers have fine-tuned derivatives. In the strictest sense, the work is preserved.
But model weights are artifacts, not capability. The institutional knowledge — the training recipes, the data curation instincts, the post-training judgment calls, the debugging intuition that comes from building dozens of models over years — that walks out the door with the people.
Open-source AI has a bus factor problem. Not in the traditional "what if someone gets hit by a bus" sense, but in the "what if the organization reorganizes and the team quits" sense. The models are open. The process of making them is not.
This isn't unique to Qwen. DeepSeek faces similar concentration risk. GLM-5 was built on Huawei chips partly because of geopolitical constraints, not just technical choice. The open-source surge is real, but it's built on a handful of teams in a handful of organizations in a handful of countries.
The Broader Picture
Step back from the crisis and the picture is still extraordinary. Open-weight models now trail proprietary state-of-the-art by roughly three months on average, according to Epoch AI. A model you can run on a MacBook now achieves scores that required a data center a year ago. Six different organizations independently produced frontier-quality coding models under permissive licenses in the same quarter.
The Pragmatic Engineer's survey shows 95% of developers now use AI tools weekly. 75% use AI for more than half their work. The AI coding revolution isn't coming — it's here, and open-source models are a massive part of it.
But sustainability requires more than good models. It requires teams that can keep building them. It requires organizations that understand why openness matters. It requires structures that don't accidentally destroy the culture that produced breakthrough results.
What to Watch
Three things will tell us whether the open-source surge sustains:
- Qwen's next release. Will it be open-weight? Will it be competitive? The answer tells us whether the team's departure was a setback or a collapse.
- Where the talent goes. Lin, Yu, Hui, Li — these are among the most capable AI researchers in the world. Where they land next will shape the landscape. Hui is already at Meta. The others haven't announced.
- The diversification of open-source. If one team's departure can threaten the movement, the movement isn't robust enough. GLM-5, DeepSeek, MiniMax, Mistral, and Meta's Llama all matter more now.
The Signal
The open-source surge in AI coding is real. The economics are undeniable. The quality gap is closing. For the first time, developers have genuine alternatives to proprietary APIs — models they can run, modify, and own.
But the surge was built by people, not just code. And people leave. The most important open-source AI team in the world built something extraordinary in 2025 and early 2026. Then the organization dismantled the structure that made it possible, and the team walked out.
The models remain. The question is whether the movement does too.
KaraxAI tracks the cutting edge of AI-assisted coding — the tools, models, and techniques that actually change how code gets written. Signal, not noise.