Qwen3.5 Medium Models Just Ran Circles Around Claude Sonnet 4.5 — On a Desktop PC

Here is something that would have sounded completely made up two years ago: a team at Alibaba just released a set of AI models that can match or beat some of the most expensive, cloud-only AI systems in the world — and you can run the best one on a gaming PC with a decent GPU.

No server farm. No six-figure cloud bill. Just your computer, a 32GB graphics card, and one of the most capable language models ever built for local hardware.

That is the short version of what the Qwen team dropped on February 25, 2026. The longer version is even more interesting, so let’s get into it.

What Is Qwen3.5 Medium, and Why Should You Care?

Alibaba has been releasing AI models under the “Qwen” name for a while now, and each release has been pushing further into territory that was previously reserved for big American tech companies. The Qwen3.5 Medium series is the latest chapter in that story.

The team released four new models as part of this series. Three of them are fully open source under the Apache 2.0 license, which basically means you can download them, use them commercially, modify them, and build products on top of them — for free:

• Qwen3.5–35B-A3B — The flagship local model, designed to run on consumer GPUs
• Qwen3.5–122B-A10B — A heavier model built for server-grade hardware (80GB VRAM)
• Qwen3.5–27B — Optimized for efficiency, supporting context windows over 800,000 tokens
• Qwen3.5-Flash — A proprietary hosted version available through Alibaba Cloud’s API

You can grab the open source versions right now on Hugging Face and ModelScope.

The fourth model, Qwen3.5-Flash, sits behind Alibaba’s cloud API rather than being freely downloadable. But even that one is competitively priced in a way that makes most Western alternatives look overpriced by comparison (more on that below).

Now here’s the part that actually matters: these models are not just “pretty good for their size.” On third-party benchmark tests, the Qwen3.5–35B-A3B specifically outperforms OpenAI’s GPT-5-mini and — this is the headline — Anthropic’s Claude Sonnet 4.5. A model that only came out five months ago and costs $18 per million tokens to run through the API.

The Qwen3.5–35B-A3B can do that on your desktop.

Understanding the Tech Without Going Cross-Eyed

You do not need to be a machine learning engineer to appreciate what Alibaba did here. But a quick explanation of the architecture helps you understand why this matters.

The Mixture-of-Experts Trick

Most AI models work by activating all of their “neurons” (parameters) every single time they process a word or phrase. That is like asking every employee in a 10,000-person company to show up to every single meeting, even if only five people actually need to be there.

Qwen3.5 uses something called a Mixture-of-Experts (MoE) architecture instead. Think of it more like a specialist referral system. The model has 35 billion parameters in total, but for any given input, it only wakes up about 3 billion of them — the specific “experts” most relevant to that particular task.

The result? You get performance that competes with much larger models, while using a fraction of the compute. It is one of the smartest efficiency tricks in modern AI development.

For the Qwen3.5–35B-A3B specifically, the design breaks down like this:

Specification Detail Total Parameters 35 billion Active Parameters per Token 3 billion Number of Experts in MoE Layer 256 Routed Experts per Token 8 Shared Experts 1 Maximum Context Length 1 million+ tokens (on 32GB VRAM GPU) Quantization Support Near-lossless at 4-bit weights License Apache 2.0 (open source)

That 256-expert pool with only 8 routed per token is particularly clever. It creates enormous diversity in what the model can handle while keeping inference fast and memory usage down.

Gated Delta Networks: The Secret Sauce

On top of the MoE system, Qwen3.5 integrates something called Gated Delta Networks into its architecture. Standard AI models use what are called Transformer blocks to process information. Qwen3.5 combines those traditional Transformer blocks with these newer Gated Delta Networks, which are better at tracking information over very long sequences of text.

This hybrid design is a big part of why these models can handle context windows over 800,000 to 1 million tokens without falling apart. For context (pun intended): 1 million tokens is roughly 750,000 words, or about the length of seven average novels. Being able to hold that much information in memory while still reasoning accurately is an engineering challenge that most models fumble.

Near-Lossless Quantization

Here is another thing worth knowing. “Quantization” is the process of compressing a model so it takes up less storage and memory. It works by representing the model’s internal numbers with less precision — kind of like saving an image as a low-quality JPEG instead of a RAW file.

The problem is that most models lose noticeable accuracy when you quantize them aggressively. Qwen3.5 was specifically engineered to resist this degradation at 4-bit quantization, which is the most aggressive common setting. The Qwen team calls it “near-lossless” accuracy under these conditions.

That matters for local deployment because it means you can compress the model significantly to fit on consumer hardware without the model suddenly becoming dumb. You keep most of the performance while slashing the memory footprint.

The “Thinking Mode” Feature

One of the standout product decisions in Qwen3.5 is that it ships with a native “Thinking Mode” turned on by default.

Before giving you a final answer to a complex question, the model generates an internal reasoning chain — marked by <think> tags — where it works through the logic step by step. This is similar to what OpenAI calls "reasoning" in its o-series models and what Anthropic has built into newer versions of Claude.

The difference here is that it is baked into the base model and available locally. You do not need to pay extra for a “reasoning” version. You do not need to be connected to a cloud API. Your local Qwen3.5–35B-A3B instance will think through problems before answering them, right on your own machine.

For things like math problems, coding tasks, multi-step logical questions, and document analysis, this makes a very real difference in answer quality. Models that think before answering tend to catch their own mistakes and arrive at more reliable conclusions.

How Does It Actually Perform? The Benchmark Numbers

Benchmark tests are always a bit of a contested topic in the AI world — no single test tells the full story, and companies have been known to optimize specifically for benchmarks they plan to publicize. That caveat noted, the results for Qwen3.5–35B-A3B are genuinely hard to dismiss.

Head-to-Head Comparison Table

The 35B-A3B model does not win everything. Claude Sonnet 4.5 edges it out on the Harvard-MIT math test. But across most categories, especially the agentic ones — where the model needs to use tools, browse the web, or interact with systems autonomously — Qwen3.5–35B-A3B is ahead.

The agentic tool use score of 67.3 vs. Sonnet 4.5’s 55.5 is particularly notable. That gap is not small. It suggests that for anyone building AI agents or automated workflows, this open source model might actually outperform a model that costs $18 per million tokens to run.

And perhaps most strikingly, the 35B-A3B model even outperforms Alibaba’s own previous flagship, Qwen3–235B, on several of these benchmarks. A model with 35 billion parameters beating a model with 235 billion parameters at various tasks — that is the whole story of the MoE architecture in one sentence.

The Pricing Reality Check

Let’s be honest about something: price matters. A lot. Especially if you are a student, an independent developer, a startup, or anyone who is not backed by enterprise-level funding.

The AI API market right now is a wild place. You can pay anywhere from a few cents to nearly $200 per million tokens depending on which model you choose. Here is how Qwen3.5-Flash stacks up against the competition when accessed via API:

AI Model API Pricing Comparison (Per 1M Tokens, Input + Output)

Qwen3.5-Flash at $0.50 total vs. Claude Sonnet 4.5 at $18.00 total — while delivering comparable or better benchmark scores — is a 36x cost difference. That is not a minor pricing adjustment. That is a different category entirely.

For API users, Alibaba Cloud’s Model Studio also offers granular tool-calling pricing on top of the base rates:

• Web Search: $10 per 1,000 calls
• Code Interpreter: Free for a limited time

If you want to test these models or build something without spending a lot of money, Qwen3.5-Flash through the API or the open source models running locally are both serious options.

Running a Frontier-Level Model on Your Own PC

Let’s talk about what “runs on consumer hardware” actually means in practice.

The flagship Qwen3.5–35B-A3B can handle a context window exceeding 1 million tokens on a GPU with 32GB of VRAM. That is the kind of GPU you find in high-end gaming setups and workstations — things like the NVIDIA RTX 4090 or the RTX 5090. Not cheap, but far more accessible than renting server infrastructure.

For reference, running a model that performs comparably to Claude Sonnet 4.5 through the API costs $18 per million tokens. If you are doing serious work and generating a lot of tokens, that adds up fast. A researcher or developer running 100 million tokens per month would be looking at $1,800 just in API costs — before any other business expenses.

Owning the hardware and running Qwen3.5–35B-A3B locally sidesteps that cost entirely after the upfront GPU investment. It also sidesteps the privacy concerns that come with sending your data to a third-party API.

The Privacy Angle Is Real

This is something that does not get talked about enough in most AI coverage. When you use a cloud API — any cloud API, from any company — your prompts and data are sent to servers you do not control. For individuals, that is often fine. For companies handling sensitive information, medical records, legal documents, customer data, or proprietary research, it is a genuine problem.

Running Qwen3.5 locally means your data never leaves your machine. You can ingest entire document archives, process hours of transcribed audio, or analyze sensitive internal reports without any of it touching an external server. That is a meaningful advantage for a lot of real-world use cases, and it is now achievable at performance levels that were previously impossible without server farms.

What the 122B and 27B Models Bring to the Table

The 35B-A3B gets most of the spotlight, but the other open source models in this series are worth understanding too.

Qwen3.5–122B-A10B: The Server-Grade Option

This one is not for your desktop. It requires 80GB of VRAM, which puts it in the data center GPU category (NVIDIA H100s, A100s, that kind of thing). But if you have access to that hardware — through a university, a company, or a cloud computing service — it narrows the gap with the absolute largest frontier models available.

With 122 billion total parameters and 10 billion active per token, it handles 1 million token contexts as a baseline. The benchmark results suggest it sits between the 35B model and the very largest models from OpenAI and Google. For organizations that need better-than-Sonnet performance without paying frontier API prices, this is an interesting option.

Qwen3.5–27B: The Efficiency Champion

The 27B model trades some raw performance for exceptional efficiency. It supports context windows over 800,000 tokens, which is still massive, and it is specifically tuned to deliver strong results per compute dollar. For applications where you are running many parallel instances or have tighter hardware budgets, this model hits a nice balance point.

Alibaba also open-sourced the base (non-instruct-tuned) version of the 35B-A3B model, which is a specific gift to the research community. Researchers who want to fine-tune the model for specialized domains — medicine, law, scientific literature, specific languages — can start from the base weights rather than the instruction-following version. That is a meaningful contribution to the broader AI research world.

Why This Matters Beyond the Benchmarks

Numbers on benchmark charts are easy to dismiss as marketing material. So let’s think about what this release actually means for people building things with AI.

The Gap Between Open Source and Proprietary Is Closing Fast

Not long ago, the honest answer to “should I use an open source model or a closed-source API?” was: “Use the API if you need real performance.” Open source models were good for research and experimentation, but when you needed production-quality results on hard tasks, you paid for OpenAI or Anthropic.

That calculus is shifting. The Qwen team’s release, alongside similar efforts from DeepSeek and others, is showing that open source development can now produce models that genuinely compete with the best proprietary systems. The performance gap has not closed entirely, but it has shrunk enough that for many real-world applications, open source is now the practical choice.

Agentic AI Just Got More Accessible

One of the most interesting scoring categories in the benchmark data is the agentic task category — specifically BFCL V4 (tool use) and BrowseComp (web search). These tests measure whether the model can actually do things, not just answer questions. Using tools, browsing websites, writing and executing code, completing multi-step tasks autonomously.

Qwen3.5–35B-A3B scored 67.3 on BFCL V4, compared to 55.5 for Claude Sonnet 4.5. That is a bigger gap than most single-category differences in these benchmarks. For developers building AI agents — systems that take actions on behalf of users rather than just generating text — this matters a lot.

Early users on Hugging Face specifically flagged the agentic performance as a standout feature, noting that it competes in territory where previously only the largest closed models could hold their ground.

The Thinking Mode Changes What “Local AI” Can Do

Before models with built-in reasoning chains became common, running AI locally meant running fast but shallow models. They were useful for autocomplete, summarization, basic question answering. But ask them to work through a complex problem with multiple steps, and they tended to make mistakes that a smarter model would catch.

Qwen3.5’s default Thinking Mode changes that. The model reasoning through its own chain of thought before answering — locally, on your hardware — brings a different quality of output to on-device AI. You are not just getting local compute for speed and privacy; you are getting local compute for genuinely sophisticated reasoning.

How to Actually Get These Models

If you want to try these yourself, the process is fairly accessible:

Option 1: Download from Hugging Face or ModelScope

The three open source models are available at:

• https://huggingface.co/collections/Qwen/qwen35
• https://modelscope.cn/collections/Qwen/Qwen35

You will need a compatible runtime like Ollama, LM Studio, or vLLM to run them locally. The hardware requirements vary by model, with the 35B-A3B needing a 32GB VRAM GPU for its full 1M token context window (though it can run with reduced context on less VRAM with quantization).

Option 2: Use the Qwen3.5-Flash API

If you do not have the hardware but want to test the capabilities, Alibaba Cloud Model Studio offers API access to Qwen3.5-Flash at $0.10 input / $0.40 output per million tokens. At those prices, you can do a lot of experimentation without spending much.

Option 3: Fine-tune from the Base Model

For researchers and ML practitioners, Alibaba released the Qwen3.5–35B-A3B-Base model — the untuned foundation that the instruct model was built from. This is the starting point for custom fine-tuning, and it gives the research community something valuable to work with.

The Bigger Picture: What Alibaba Is Proving

There is a broader story here that goes beyond any single model release. Alibaba’s Qwen team has been on a consistent path of demonstrating that you do not need to be an American company with billions in funding to build world-class AI models.

The Qwen3.5 Medium release follows recent history in the same direction: Qwen3–235B, Qwen3.5–397B-A17B, and now this. Each release has pushed the performance-per-cost ratio further, and each has been matched with open source releases that put the technology in the hands of anyone with reasonable hardware.

This is creating real competitive pressure on the pricing of Western models. When an open source model running locally outperforms a $18-per-million-token API on most benchmark categories, the pricing of that API becomes difficult to justify unless there are specific capabilities or reliability guarantees that justify it.

For people who are just trying to build things and solve problems, more capable open source models are simply a good outcome. Better tools, lower barriers, more options. The fact that they are coming from Alibaba rather than from San Francisco is irrelevant to whether they actually work.

Things Worth Watching

A few questions remain open with this release.

Real-world performance vs. benchmark scores. Benchmarks measure specific, standardized tasks. They do not always predict how a model performs in messier, more varied real-world conditions. The early Hugging Face community response has been positive, but larger-scale real-world testing will tell a more complete story.

Quantization quality in practice. The Qwen team claims near-lossless accuracy under 4-bit quantization. That is worth verifying independently at different tasks and use cases. “Near-lossless” can mean different things across different task types.

Long-context reliability. Supporting a 1 million token context window and reliably using information from across all of it are two different things. Models that technically support long contexts sometimes struggle to accurately retrieve and reason about information from the distant parts of a long document. Independent testing on real long-document tasks would help clarify this.

The Flash model’s proprietary nature. Three of the four models are open source, but Qwen3.5-Flash is cloud-only and proprietary. The decision to keep the most production-ready, tooling-complete version behind Alibaba Cloud creates a natural limitation for organizations that want the official tool integrations without depending on a Chinese cloud provider.

None of these are deal-breakers. They are just things that a thoughtful person building on this technology should keep in mind as they evaluate whether and how to use it.

Wrapping Up

The Qwen3.5 Medium series is a genuinely impressive release. Not because it is Chinese or open source or cheap — though all of those things matter — but because it demonstrates that the assumptions underlying the current AI market pricing structure are becoming harder to defend.

When a 35-billion-parameter model that runs on consumer hardware can beat Claude Sonnet 4.5 across most benchmark categories, it raises a legitimate question for anyone paying $18 per million tokens: what are you actually getting for that premium?

For some use cases, the answer is still “reliability, support, and specific capabilities.” For many others, the answer is becoming “not much.”

Whether you are a student experimenting with AI for the first time, a developer building an application, a researcher fine-tuning models for specialized domains, or a company evaluating AI infrastructure costs — the Qwen3.5 Medium series is worth your time.

The models are out there. You can download them today from Hugging Face or ModelScope, or test Qwen3.5-Flash through Alibaba Cloud’s API at a fraction of what you would pay for comparable proprietary options.

The gap between what you can run locally and what only the biggest cloud providers could offer before is smaller than it has ever been. And based on the trajectory of the Qwen team’s releases over the past year, it is probably going to keep shrinking.

All benchmark data sourced from Alibaba’s official Qwen3.5 release documentation. Pricing data accurate as of February 26, 2026.