AI・テクノロジー
The Shock of Qwen3.6-27B | Claude-Level AI on an RTX 4090
機械翻訳 / Machine-translated
A
AI Friends
2026年4月26日・26分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
"I want to use Claude Opus, but the monthly subscription is painful." "I'm nervous about handing company code to ChatGPT." — A free AI has arrived that blows both of these concerns away in one shot.
Released on April 22, 2026, Alibaba's "Qwen3.6-27B" is an open-source AI that runs on a home gaming PC yet matches Claude Opus 4.5 on certain benchmarks — a truly shocking development.
Let's break down exactly what makes it so impressive.
First, let's cover the basics: who made this AI and what is it?
The Qwen team at Alibaba (China's answer to Amazon) released their latest AI, "Qwen3.6-27B," on Hugging Face and ModelScope on April 22, 2026. It comes under the Apache 2.0 license — a permissive set of terms that means "commercial use OK, modifications OK, anyone can use it freely."
The impact is like a coffee shop publicly releasing its recipes and then declaring that modifications are welcome too. Engineers around the world started downloading it on day one.
The parameter count is 27 billion (27B). Alibaba's previous version, "Qwen3.5," had featured a massive 397 billion (397B) parameter MoE model.
The impossible has happened: a compact car has outrun a semi-truck. The reasons why something so small can be so powerful lie in the new architecture described later, but this is being called a historic moment when the assumption that "bigger AI is always better AI" began to crack.
Benchmarks — standardized tests between AI models — are the most important indicators of performance. Qwen3.6-27B posted shocking scores in the coding domain.
SWE-bench Verified measures "whether an AI can fix real-world GitHub issues," and Qwen3.6-27B scored 77.2%.
That puts it within 3.6 percentage points of Claude Opus 4.6's 80.8%.
On Terminal-Bench 2.0 (measuring terminal operation proficiency), it scored 59.3% — an exact match with Claude Opus 4.5.
The fact that "a free AI has emerged that can stand toe-to-toe with some of the world's most powerful paid AIs" has sent shockwaves through the industry.
The new "Thinking Preservation" feature is another major differentiator.
With traditional AI, as conversations grew longer, there was wasteful "forgetting what was thought before and starting over from scratch."
Qwen3.6-27B saves the reasoning process in the conversation history so it can be reused in subsequent turns.
Think of it like this: instead of doing all the prep work for a meal from scratch every time, you now save it in the fridge and reuse it. This dramatically improves the accuracy of agents (AI that works autonomously).
The answer to "why is it so small yet so strong?" lies in a new design called hybrid attention.
Qwen3.6-27B consists of 64 transformer layers total, structured as 16 repetitions of a 4-layer set: "3 layers of Gated DeltaNet (linear attention) + 1 layer of standard attention."
Imagine three fast part-time workers being supervised by one experienced veteran — fast, but without sacrificing intelligence. This creates an exquisite balance: dramatically reducing computation while maintaining accuracy.
The context window determines how much information can be packed into a single conversation. Qwen3.6-27B natively supports 262,144 tokens, expandable to approximately 1,010,000 tokens via YaRN technology.
That's the equivalent of feeding it ten paperback novels at once and asking for a summary. Use cases that once seemed like pipe dreams — such as feeding an entire massive Git repository to an AI for refactoring suggestions — are now a reality.
The most important question for open-source AI is "can I run it on my own PC?" Here too, Qwen3.6-27B delivers a shock.
The full BF16 weights are a large 55.6 GB, but with Q4_K_M quantization (a method that compresses size while preserving quality), it shrinks to approximately 16.8 GB.
This runs comfortably on a single RTX 4090 (the top consumer GPU with 24 GB VRAM, priced at around ¥300,000 as of 2026).
Q5 quantization comes in at 19.5 GB and Q6 at 22.5 GB, so an existing gaming PC can be used as-is.
We have entered an era where "AI that used to require million-yen enterprise servers can now run for ¥3,000 a month in electricity."
An FP8 (8-bit floating point) version — Qwen/Qwen3.6-27B-FP8 — was also released simultaneously for enterprise use. With fine-grained quantization at block size 128, it supports major inference runtimes including SGLang (0.5.10+), vLLM (0.19.0+), KTransformers, and Hugging Face Transformers.
The design is praised for its scalability: "small teams use RTX 4090; large enterprises run massive parallel processing on H100/H200."
Let's compare it with the major AIs as of April 2026.
The standout factor is cost-effectiveness.
Development teams paying hundreds of thousands of yen per month for Claude can switch to Qwen3.6-27B and pay only the upfront GPU cost.
There are also reports that Qwen 3.6 Plus Preview delivers 2–3× the output speed of Claude Opus 4.6, making it a fully viable replacement for coding tasks.
For Japanese engineers, startups, and small-to-medium businesses, Qwen3.6-27B has the potential to be a true game-changer.
At Japanese companies — especially in finance, healthcare, and manufacturing — the rule "don't send internal code to external AI" is standard practice.
The reality is that many workplaces cannot use Claude or ChatGPT for business purposes.
Because Qwen3.6-27B runs entirely on in-house servers, it's both legal and safe to "feed confidential code to an RTX 4090 at home and get refactoring suggestions."
The infrastructure for "industries that were falling behind in AI adoption to catch up all at once" is now in place.
The Qwen series has long been known for strong Japanese language performance. Qwen3.6-27B handles Japanese prompts, Japanese output, and Japanese code comments with ease.
In a situation where "you don't have to wait for a domestic LLM — a high-quality model from China is available for free," the competitive landscape of the local LLM market (self-hosted language model market) is being rapidly redrawn. For AI startups, the reduction in API usage fees is an immeasurable benefit.
Tamura-san earns ¥800,000 per month doing contract web app development.
Until now, he was paying ¥15,000 per month for Claude Pro and GitHub Copilot.
He switched by running Qwen3.6-27B on his home RTX 4090 and connecting it to Cline (an AI agent for VS Code).
"I saved ¥180,000 over three months — and it even runs offline on the Shinkansen," he says, fully satisfied. The clincher was the peace of mind that clients' confidential code could be processed locally.
Takahashi-san handles IT at a 50-person machine parts manufacturer.
Under the president's strict order of "don't hand customer data to external AI," he installed Qwen3.6-27B on an internal server. He then built an in-house agent pipeline covering invoice OCR → automatic entry into the core system → automated weekly report generation.
He reported to the president that "invoice processing that required 3 people for 2,000 slips per month now runs with just 1," and was promoted to the company's first ever "Chief AI Officer (CAIO)."
Professor Kimura specializes in materials engineering.
He had been troubled by the thought that "feeding unpublished experimental data into ChatGPT for analysis would violate research ethics." He installed Qwen3.6-27B on the lab's workstation and brought in-house everything from pattern analysis of experimental data, to drafting English papers, to drafting responses to reviewer comments.
"The number of international papers we submit went from 3 to 8 per year," and it became a talking point in academic circles.
It's a symbolic example of "free AI changing the productivity of basic research."
A. The model itself is completely free under Apache 2.0.
The initial investment is an RTX 4090 (approximately ¥300,000) or a used RTX 3090 (approximately ¥100,000).
Even running 24 hours a day, electricity costs are approximately ¥3,000–¥5,000 per month.
"Compared to paying ¥30,000 per month for Claude Opus, you break even in six months" — so the return on investment is quick for both individuals and companies. With a used 3090, you can recoup costs in as little as three months.
A. Because it is open-source, the weights can be verified by third parties, and under the Apache 2.0 license, "you can use it however you like," so it is technically impossible for the model to send data externally (when running locally).
The situation is: "Made by a Chinese company, but it only runs inside your PC." Government agencies may require a formal evaluation process for adoption, but it is already globally commonplace in private-sector use.
A. Running on CPU only with 64 GB or more of DDR5 RAM is possible with the Q4 quantized version, but the response speed of roughly 1 token/second is not practical.
On Mac, it runs reasonably well on Apple Silicon (M3 Max or later; 64 GB unified memory recommended).
"For stress-free use, an NVIDIA GPU with 24 GB or more" is the practical answer as of April 2026. Note: CUDA 13.2 has a bug that causes garbled characters; please use CUDA 13.1 or the 12.x series.
A. Apache 2.0 is a permissive license that essentially says "almost anything goes as long as you include the copyright notice and disclaimer."
Embedding it in your own service for sale, offering it as a SaaS, and using it for internal business operations are all permitted.
When redistributing, you must retain the original license notice, but you have free rights over the output — meaning "selling code generated by the AI as your own product is perfectly fine."
A. Qwen 3.6 Plus Preview (with 1M context support) is already available as an early release on OpenRouter, and it is highly likely that the standard version will also see different sizes as part of the flagship lineup. The Qwen team has been releasing major updates every quarter since 2025, and a next-generation model is expected in Q3 2026.
There is a concern that "the GPU you buy today will be obsolete in a year," but given the pace of AI development, this is a standard risk.
The assumption that "bigger AI is always stronger AI" — which held until 2024 — has been rendered completely obsolete by Qwen3.6-27B. We are in an era where you can run, for free, a model that stands shoulder-to-shoulder with the world's top commercial AIs on a home gaming PC. Whether you watch this change from the sidelines or get your hands on it right now and make it your own weapon will make a huge difference in your value as an engineer one year from now.
"Those who have the best free AI on their side win" — the second half of 2026 is here. Why not start with a 30-minute download?
This article is a cross-post from AI Friends.
