Alibaba released Qwen3.6-27B on April 22, 2026. Anthropic released Claude Opus 4.7 six days earlier. One runs on a closed API in a hyperscale datacenter. The other runs on a 24 GB GPU in your spare room. For the first time, the benchmark gap between those two worlds is small enough to actually matter.

This post lines up the numbers.

The Headline Comparison

BenchmarkQwen3.6-27B (local)Claude Opus 4.7 (API)Gap
SWE-bench Verified77.2%87.6%-10.4 pts
SWE-bench Pro53.5%64.3%-10.8 pts
Terminal-Bench 2.059.3%~59%~tie
MMLU-Pro81.7%——
GPQA Diamond87.8%——

A ten-point gap on SWE-bench Verified is not nothing. But consider what Qwen3.6-27B beats: it scores higher than Qwen3.5-397B-A17B — Alibaba’s own previous flagship — on every coding benchmark, despite having roughly 15× fewer parameters. SkillsBench is the most striking: 48.2 vs 30.0, a 77% relative improvement against a model with 14.8× the parameter count.

Terminal-Bench Is the Surprising One

SWE-bench Verified measures patch generation against curated GitHub issues. Terminal-Bench 2.0 measures agentic tool use — running commands, reading output, adapting. This is closer to what people actually do with AI coding assistants in 2026.

Qwen3.6-27B at 59.3% on Terminal-Bench is essentially tied with Opus 4.7. That is the benchmark where I expected the biggest gap, because closed frontier models have historically had a large lead on multi-turn agentic tasks. They no longer do, at this size class.

What the Gap Buys You

Opus 4.7’s 10-point lead on SWE-bench Verified translates to real differences:

  • More patches that compile and pass tests on the first attempt
  • Better handling of large, unfamiliar codebases
  • Stronger reasoning about API contracts and side effects
  • 200K context window vs Qwen3.6-27B’s standard 128K (extensible to 262K with config)

For a senior developer driving the model, the gap shows up mostly in how often you have to intervene. For a hobby project where you are doing the architectural thinking yourself, 77% is plenty.

What the Gap Does Not Buy You

Opus 4.7 costs $15/$75 per million input/output tokens. A Qwen3.6-27B query on your own hardware costs the electricity to run a 200W card for ~10 seconds. At any volume — agentic loops, large refactors, batch document processing — the economics flip hard.

You also keep the data on your machine. For consultancy work where I am looking at client codebases, that is not a small feature.

The Architectural Story

Qwen3.6-27B is a dense model. It activates every parameter for every token. The frontier has spent the last 18 months moving the other direction — MoE architectures with hundreds of billions of total parameters but only 10-20B active. Dense models were supposed to be obsolete.

They are not. Dense at 27B with MTP speculative decoding hits 140 tokens/sec on a single high-end GPU. The same hardware running a 200B MoE model spends most of its time shuffling weights through PCIe.

For local inference, dense wins because the bottleneck is memory bandwidth, not compute. Qwen3.6-27B is designed for the constraints we actually have, not the constraints frontier labs have.

Where This Leaves Things

If you write code for a living and your employer pays for Opus 4.7, keep using it. The 10-point gap on hard problems is real.

If you are building a side project, doing security-sensitive work, running automated pipelines where token cost matters, or just want to understand how the model behaves under the hood — Qwen3.6-27B has crossed the threshold. It is no longer a curiosity. It is a tool you would pick on merit.

The interesting question for the next twelve months is whether Anthropic and OpenAI can keep a ten-point lead, or whether the local ecosystem closes the gap further. Given Qwen3.6-27B already beat a 397B model from the same team, I would not bet on the gap holding.

Part of the Local AI series