Qwen3.5: The 397B Open-Weight Model Priced to Disrupt GPT-5.2

Alibaba released Qwen3.5 on February 16, 2026 — Chinese New Year’s Eve — and the 397-billion parameter headline is designed to mislead you. Only 17 billion of those parameters activate per forward pass. The model trades punches with GPT-5.2 on most benchmarks, runs on a Mac with enough RAM, and costs about 5x less than Gemini 3 Pro via API. That last number is the one that matters.

Released under Apache 2.0 on Hugging Face, Qwen3.5 is the largest open-weight model to reach frontier-competitive performance. But the benchmark comparison is almost beside the point. At $0.40 per million input tokens on Alibaba Cloud, the pricing changes the math for every developer currently paying Anthropic or OpenAI rates. Qwen3.5 does not need to be better than GPT-5.2 or Claude. It just needs to be good enough — and at that price, the threshold for “good enough” is remarkably low.

What Alibaba Actually Built

The timing was deliberate symbolism from the world’s most prolific open-weight lab. Qwen3.5 packs 397 billion total parameters into 512 Mixture-of-Experts modules, routing 10 experts plus 1 shared expert per token. That 4.3% activation ratio — 17 billion active parameters out of 397 billion — means the “397B” headline describes storage, not compute. The distinction matters: this model’s per-token cost is closer to a 17B dense model than to anything that sounds like 397 billion parameters.

Unlike previous Qwen releases that shipped text and vision as separate models, Qwen3.5 is natively multimodal. Images up to 1344×1344 resolution, video up to 2 hours, and text feed into a single unified architecture — no separate VL variant required. The native context window spans 262K tokens, extensible to 1 million via YaRN scaling.

Gated Delta Networks: Why Qwen3.5 Doesn’t Do Attention Like Everyone Else

The technical innovation that separates Qwen3.5 from “just another large MoE model” is its hybrid attention architecture. Across 60 layers, the model uses a 3:1 ratio of Gated Delta Networks (linear attention derived from Mamba2) to standard Gated Attention. Those GDN layers scale near-linearly with sequence length rather than quadratically — which is why the model handles 262K context natively without memory explosion.

The practical results per Alibaba’s benchmarks: 8.6x faster decoding at 32K context and 19x faster at 256K context compared to Qwen3-Max. NVIDIA has separately documented the predecessor architecture and its efficiency gains on Blackwell hardware. This approach diverges from DeepSeek’s Multi-Head Latent Attention and Zhipu’s Deep Sparse Attention. A year ago the industry debated MoE versus dense models. Now the question is how to implement attention efficiently — and three major Chinese labs have chosen three different answers.

For readers tracking this architectural divergence, we covered the broader shift toward hybrid linear-attention architectures in January. For local deployment, a 4-bit quantized version weighing 214GB runs on a 256GB M3 Ultra Mac at usable speeds, per Unsloth’s deployment guide. Full BF16 precision requires 8x H100 80GB GPUs at approximately 45 tokens per second.

Qwen3.5 Benchmarks: Where It Leads, Where It Trails

Alibaba claims Qwen3.5 outperforms GPT-5.2, Claude Opus 4.5, and Gemini 3 Pro on “80% of evaluated benchmarks.” That claim is technically defensible and strategically misleading. An independent analysis by mlabonne on HuggingFace provides a clearer picture.

Where Qwen3.5 genuinely leads: instruction following. IFBench scores hit 76.5 versus GPT-5.2’s 75.4 and Claude Opus 4.5’s 58.0. MultiChallenge shows an even wider gap: 67.6 versus GPT-5.2’s 57.9. Multimodal vision tasks are similarly strong, with MMMU at 85.0 and OmniDocBench at 90.8. Coding sits in a dead heat: SWE-bench Verified 76.4, within fractions of Kimi K2.5 (76.8) and Gemini 3 Pro (76.2).

Where it trails meaningfully: math reasoning (AIME26: 91.3 versus GPT-5.2’s 96.7), competitive programming (LiveCodeBench v6: 83.6 versus 87.7), and agentic tasks (TAU2-Bench: 86.7 versus Claude Opus 4.5’s 91.6). Qwen3.5’s BrowseComp score swings from 69.0 to 78.6 based solely on scaffolding strategy — not model changes. As mlabonne noted, “agentic scores are increasingly a function of scaffolding and context management, not just raw intelligence.” Apply that skepticism to every lab’s agentic numbers, not just Alibaba’s.

CNBC reported on February 17 that it “could not independently verify” Alibaba’s comparative claims. Full third-party evaluation remains underway as of February 18.

The Pricing Is the Point

Qwen3.5-Plus API costs $0.40 per million input tokens internationally. Claude Opus 4.6 costs $5. GPT-5.2 runs approximately $1.75. Gemini 3 Pro sits at $2.00 for prompts under 200K tokens. Against Gemini that’s a 5x gap; against Claude it’s 12.5x. A task that costs $50 with Claude costs roughly $4 with Qwen3.5. For most production workloads, the performance gap is single-digit percentage points. The price gap is measured in multiples. For full context, see our comparison of AI API pricing across the major providers.

One developer reaction on X captured the emerging calculus: “If an open weights model is released that’s as capable at coding as Opus 4.5, then there’s very little reason not to offload the actual writing of code to open weight subagents running locally and stick strictly to planning with Opus 5.” That two-tier architecture — cheap open-weight models for volume execution, expensive closed models only for planning — is the structural shift Qwen3.5’s pricing enables.

But the pricing doesn’t survive scrutiny as pure engineering efficiency. The GDN architecture delivers roughly 8.6x throughput improvement — impressive, but that still leaves a gap engineering alone can’t explain. HelloChinaTech’s analysis called it plainly: “The pricing almost certainly carries a subsidy component.” The answer is Alibaba Cloud — the same loss-leader playbook Amazon ran with AWS in the early 2010s. The model is not the product. The platform is.

Bigger Picture: Qwen Has Overtaken Llama

Qwen has quietly displaced Meta’s Llama as the most-downloaded and most-used base model for fine-tuning on HuggingFace. More fine-tunes mean more community tooling, more deployment benchmarks, more adoption momentum. The network effect that kept Llama dominant for over a year now works for Qwen.

Qwen3.5 arrives as part of an unprecedented February 2026 wave from Chinese open-weight labs. Zhipu shipped GLM-5 (744B parameters), Moonshot released Kimi K2.5 (1 trillion), and Baidu’s ERNIE 5.0 reached 2.4 trillion. Latent Space positioned Qwen3.5 as potentially the last major Chinese open-weight release before DeepSeek v4. The open-weight ecosystem is now built primarily on Chinese infrastructure — not because of geopolitics, but because of release cadence and engineering merit.

The Subsidy Question Nobody Can Answer Yet

If Alibaba is subsidizing Qwen3.5 pricing to drive cloud adoption, what happens when that subsidy ends? Developers who optimize their stacks around $0.40 per million tokens may face the same rate-shock that AWS customers discovered once cloud lock-in was complete. The free 1-million-token trial and 90-day promotional pricing lower adoption friction to near-zero — and that is exactly the point.

The gap between open-weight and closed models is now small enough that a 5x-to-12.5x price difference does the persuading. Qwen3.5 does not need to win the benchmark race. It just needs to survive it — and the benchmark data suggests it already has. The more interesting question is whether “good enough at a fraction of the price” redraws the revenue map for OpenAI and Anthropic, who retain the hardest workloads but lose the volume that subsidizes them.

Every previous open-weight model that reached “good enough” performance eventually forced a pricing response from closed-model providers. Llama 2 pressured fine-tuning costs down. Mixtral compressed inference pricing. Qwen3.5 targets the API tier itself — and at 12.5x cheaper than Claude, the pressure arrives not as a technical threat but as a margin event. If OpenAI and Anthropic cut rates to compete, their revenue-per-token economics shift permanently. If they don’t, they cede the volume tier to an open-weight model hosted on a competitor’s cloud.