If you’re paying for Claude or GPT-5 API calls right now, here’s a number from Qwen that should bother you: $0.10 per million input tokens.

Not a promotional rate.

Not a crippled model.

That’s what Alibaba is charging for Qwen 3.5 Flash — a model that, according to their published benchmarks, beats GPT-5.2 on instruction following and tops every model tested on document understanding.

One-thirtieth (not one-third!) of the price of Claude Sonnet 4.6.

For a workload doing 10 million calls per month, the annual difference is roughly $12,000 versus $360,000.

You might ask: what’s the catch?

There isn’t one. The price is the architecture.

Qwen 3.5 has 397 billion parameters. It activates 17 billion per query — 4.3%. According to Alibaba’s own benchmarks, those 17 billion match Qwen3-Max, a model with over one trillion parameters. At 256K context length, it decodes 19 times faster than its predecessor.

On February 16 — Chinese New Year’s Eve — the Qwen team released the flagship. On February 24, the medium series dropped.

The Qwen3.5-35B-A3B, a model with just 3 billion active parameters, outperformed the previous generation’s 22-billion-active Qwen3-235B.

One-seventh the compute, beating its bigger sibling.

Today, March 3, the small models landed. The 9B variant runs on a single RTX 4090. Per Alibaba’s benchmarks, it scores 70.1 on MMMU-Pro visual reasoning — GPT-5-Nano manages 57.2.

Three releases in two weeks, and each one making the same argument: you are probably paying too much for intelligence.

A Chinese developer on 知乎 (Zhihu) described this with a phrase that stuck with me:

“暴力美学后的精细化手术” — a surgery after brute force.

I get my AI assistant to translate that, but it's pretty close.

The parameter arms race lasted two years. Alibaba just proved that architecture beats scale.

Here’s how.

Every modern AI model runs on a mechanism called attention.

It’s the engine that lets the model understand which words relate to which other words.

The problem of the traditional attention (called softmax attention) scales quadratically (for those who know time complexity, O(n^2)).

Double the context length, quadruple the compute.

This is why running long documents or complex agent workflows gets expensive fast.

Qwen 3.5 replaces most of that attention with something called Gated DeltaNet — a form of linear attention.

I will try my best to explain this … you can think of it this way:

Traditional attention is like running a meeting where every participant must talk to every other participant, every time a decision is made.

With 10 people, that’s 100 conversations. With 1,000 people, it’s a million. The room catches fire.

Gated DeltaNet works differently.

Instead of holding a massive all-hands meeting for every token, it maintains a compressed briefing document — a fixed-size state matrix. Each new piece of information updates the document (using what researchers call the “delta rule”), while a gating mechanism decides how much old information to keep and how much to forget. The meeting never grows. Whether you’re processing 1,000 tokens or 100,000, the per-token compute stays constant.

The technical lineage matters — because Gated DeltaNet isn't new. The original paper, "Gated Delta Networks: Improving Mamba2 with Delta Rule," by Songlin Yang (杨松林) and collaborators, was published well before Qwen 3.5.

Taiwanese engineer/CTO Oscar (倢愷) wrote an excellent deep-dive on Medium back in August 2025 walking through the mathematics of linear attention and KV cache optimization.

The technique existed. What nobody had done was stake a frontier model on it. Alibaba did.

But Qwen 3.5 doesn’t go all-in on linear attention. That would sacrifice the fine-grained token-to-token reasoning that full attention excels at. Instead, the Qwen team uses a hybrid 3:1 ratio: three layers of Gated DeltaNet followed by one layer of full attention (called Gated Attention, which adds output gating to standard softmax attention for training stability). The model repeats this pattern across all 60 layers.

The architecture specification from NVIDIA’s model card reveals the exact layout:

You might be curious what MoE (Mixture of Experts) is. I will talk about it in the next point.

That’s 45 layers of linear attention and 15 layers of full attention. The model gets the speed of linear attention for 75% of its computation, and the precision of full attention for the remaining 25%.

This hybrid design is the primary reason Qwen 3.5 achieves 19x faster throughput than Qwen3-Max at 256K context. The KV cache no longer grows linearly with context length for 75% of the layers — it stays fixed. For anyone building applications with long documents, multi-turn agent conversations, or large codebases, the cost and latency implications are significant.

If you read my breakdown of DeepSeek R1, you already understand Mixture of Experts.

Instead of activating all parameters for every query, MoE routes each token to a small subset of specialized sub-networks.

DeepSeek V3 — the model that shook the industry last year — uses 671 billion total parameters with 37 billion active (5.5% activation). That was already considered aggressive.

Qwen 3.5 goes further: 397 billion total, 17 billion active (4.3% activation). The model has 512 expert networks. For each token, only 10 routed experts plus 1 shared expert activate — 11 out of 512.

Here’s the design choice that caught my attention: each expert has an intermediate dimension of just 1,024.

That’s deliberately small. Instead of building a few large specialists, Alibaba built an army of tiny ones.

The efficiency comes not from any single expert being powerful, but from the routing network’s ability to assemble the right combination every time.

This extreme sparsity is what makes the pricing possible. When only 4.3% of parameters activate per token, inference cost drops proportionally. The Flash API at $0.10 per million input tokens isn’t a loss leader — it is a direct reflection of the architecture’s efficiency. For high-volume agentic workloads with frequent tool calls, this pricing changes the economics entirely.

Step back from Qwen 3.5 for a moment. Something bigger is happening across Chinese AI.

A year ago, the debate was “MoE or dense”?

That question is settled — MoE won.

Now the frontier labs are diverging on a more fundamental question:

Four major Chinese AI labs gave almost four different answers.

DeepSeek uses Multi-head Latent Attention (MLA) — compressing the key-value space into a lower-dimensional latent representation to shrink the KV cache. This is the approach that powered DeepSeek V3 and R1.

Kimi (Moonshot AI) takes DeepSeek’s foundation and refines it with Kimi Delta Attention (KDA) — kind of a variant of Gated DeltaNet but with channel-wise gating instead of scalar gating, giving finer-grained control over what each feature dimension remembers.

Qwen (Alibaba) bets on the Gated DeltaNet hybrid described above — linear attention as the default, full attention as the exception.

GLM (Zhipu AI) follows DeepSeek’s lead, adopting MLA plus DeepSeek’s Shared Attention (DSA) in GLM-5.

As Maxime Labonne (ML scientist) put it in his analysis on Hugging Face: “Nobody agrees on attention anymore.”

This divergence matters beyond academic interest. Each attention mechanism carries different trade-offs in inference cost, long-context performance, and KV cache requirements. For developers choosing which Chinese model to build on, the architecture choice determines your deployment economics — not just the benchmark scores.

I will map the full Chinese AI model landscape this/next week. But the attention architecture war is the thread connecting all of it.

Here’s how Qwen 3.5 Flash stacks up against the models you’re likely considering:

As I highlighted at the beginning of this article, for a workload making 10 million API calls per month with average 1K-token inputs, the annual cost difference between Qwen Flash and Claude Sonnet is roughly $12,000 versus $360,000.

The open-weight models (Qwen and DeepSeek) add another dimension: self-hosting eliminates per-token costs entirely, leaving only infrastructure expenses.

llama.cpp with Unsloth GGUFs is currently the most stable local option. Unsloth updated their Qwen 3.5 GGUFs on March 2 with chat template fixes that improve tool calling. The 9B model at Q4 quantization runs comfortably on 16GB RAM. The 35B-A3B needs approximately 22GB.

Qwen Flash API is the path of least resistance. OpenAI-compatible format, native tool calling, $0.10 per million tokens via Alibaba Cloud’s Model Studio (international pricing; domestic Chinese pricing runs even lower at ¥0.8/M tokens). If you need agentic capabilities today, this is it.

vLLM and SGLang are the production-grade options if you have GPU servers. Qwen’s official documentation recommends these frameworks, and both have Qwen 3.5 support.

While I decided to publish this article this morning, I received an email from Ollama:

Anyway, I want to publish my article in time (else procrastination kills me), so I will put up another shorter one for Ollama specifically on the usage of Qwen 3.5.

Important disclaimer: All benchmark numbers in this article are from Alibaba’s published evaluations unless otherwise noted. Independent third-party verification may vary. Let me know if these don’t align with your experience.

I was going to publish this yesterday. However, while I was writing this article, the story changed.

On March 4 — one day after the Qwen 3.5 small models dropped — Junyang Lin (林俊洋), Qwen’s technical lead, posted six words on X:

“me stepping down. bye my beloved qwen.”

The post hit multiple million views within hours. Ollama’s official account replied thanking him. The open-source community flooded the thread.

Then the picture got worse.

According to multiple sources on Chinese tech forums and X, Lin didn’t leave voluntarily. Qwen contributor Chen Cheng responded directly: “leaving wasn’t your choice.” Hyperbolic Labs CTO Yuchen Jin confirmed that Binyuan Hui, another core Qwen leader, also departed. Kaixin Li — a key contributor to Qwen 3.5, Qwen VL, and Qwen Coder — posted a farewell the same day. At least three to four core team members are confirmed gone.

The emerging picture from Chinese-language discussions on X and 知乎 points to internal politics at Alibaba Cloud.

Multiple people close to the team describe a shift in how Qwen was being evaluated: management began applying consumer-app metrics — daily active users — to a foundational model research team. Xinyu Yang (researcher at Deepseek) criticized the move publicly, noting that Alibaba Cloud reportedly brought in a non-core member from Google Gemini to replace Junyang’s position. But uhm, I did some research on that person, he was the senior staff research scientist at Google DeepMind.

Jiaxi Cui (Panda), a close observer of the Chinese AI lab ecosystem, offered a stark assessment: the Qwen 3.5 series may be the last masterwork from this team. He expects Alibaba to pivot toward commercialization — and away from the open-source frontier research that made Qwen matter.

Yuchen Jin’s concern cuts deeper: Qwen may stop shipping frontier open-weight models entirely, shifting to a closed commercial path.

The architecture I spent this entire article explaining — the Gated DeltaNet hybrid, the ultra-sparse MoE, the 19x throughput gains — was built by the people who just left. The $0.10 per million token pricing that opened this article exists because this team designed an architecture efficient enough to make it real.

Wtf.

Well, but this doesn’t change the technical analysis above. Qwen 3.5 is still the model described in this article, with the architecture and pricing and benchmarks as published. You can still use the Flash API today. The open-weight models are already on Hugging Face — they can’t be un-released.

But the roadmap? The next generation? The continued stream of frontier open-source models that developers worldwide were building on?

That’s an open question now…

I am still switching my pipeline to Qwen Flash (or whichever best after my research) this weekend — the model works regardless of who runs the team. But I will be watching the next few weeks closely. Where Junyang Lin and his team land will tell us more about the future of Chinese open-source AI than any benchmark ever could.

The architecture war hasn’t ended. But one of its best armies just lost its generals.

Cheers.

Zero Address covers Chinese technology for English readers. I read the Mandarin tech docs so you don’t have to.

If this reframed the story for you — share it with one person who is still reading the headline version. And of course, subscribe to receive free articles like this every week.