Qwen3.6-27B Open Source Release "Openclaw, Hermes First Choice": AI Performance Ties with Claude Opus 4.5, Cost Reduced by 14 Times

Alibaba Qwen Series Latest Flagship Qwen3.6-27B Officially Open-Sourced on the Evening of April 22, 2026. This 27B dense model scored 59.3 on Terminal-Bench 2.0, tying Claude 4.5 Opus, and with less than 1/14 of the parameters, achieved a score of 77.2 on SWE-bench Verified, surpassing the previous 397B MoE flagship’s 76.2. The full model is 55.6 GB, and with Q4_K_M quantization compressed to 16.8 GB, it can run on consumer-grade hardware, making OpenClaw, Hermes Agent, and other local agent frameworks the first to have a truly capable local brain.
(Background: After being blacklisted by Anthropic, OpenClaw recommends users switch to API keys or alternative models like Qwen or Kimi.)
(Additional context: Why does the US AI need “censorship” and stay in labs? China, on the other hand, fully embraces open-source models—why?)

Table of Contents

Toggle

• Benchmark Results: The Three Most Impressive Conclusions
• Specifications: Trillion-Parameter Performance Fit for Consumer Hardware
• Why Are OpenClaw and Hermes Agent Using It as Their Local Brain?
• Open Source vs Closed: Alibaba’s Strategic Layout and Geopolitical Landscape

On the evening of April 22, 2026, Alibaba’s Qwen team quietly pushed a bombshell to Hugging Face: Qwen3.6-27B is officially open-sourced under Apache 2.0 license, free for commercial use. The number seems ordinary, but its significance is extraordinary—this 27B dense architecture (not MoE) first matched Anthropic’s closed flagship Claude 4.5 Opus in terminal agent tests, and at only 55.6 GB, defeated the previous 397B MoE beast that required 807 GB of video memory to run fully. Local deployment, capable of running agents, on consumer-grade hardware—all three conditions are met by Qwen3.6-27B.

Benchmark Results: The Three Most Impressive Conclusions

Qwen team selected 10 benchmarks reflecting real agent programming ability, with results as follows:

Three key conclusions stand out: First, Terminal-Bench 2.0 score of 59.3 matching Claude 4.5 Opus—this is the first time a 27B dense model has caught up with Anthropic’s closed flagship in terminal agent tasks, with Qwen3.5-27B’s previous score only 41.6, a 17.7 point jump. Second, SWE-bench Verified score of 77.2 surpassing Qwen3.5-397B-A17B’s 76.2—the 27B dense model beats the previous 397B MoE flagship, shrinking the model size from 807 GB to 55.6 GB, over 14 times smaller. Third, SkillsBench jumped from 27.2 to 48.2 (+77%) and Claw-Eval Pass^3 exceeded Claude 4.5 Opus’s 59.6 with 60.6—multi-turn, multi-step consistency is the biggest upgrade, indicating the model’s improved ability to execute complex agent tasks continuously without crashing or deviating.

In knowledge and reasoning: MMLU-Pro 86.2, MMLU-Redux 93.5, GPQA Diamond 87.8, AIME 2026 94.1, LiveCodeBench v6 83.9, all surpassing previous models with similar parameters.

Specifications: Performance at the Trillion-Parameter Level on Consumer Hardware

Qwen3.6-27B is a pure dense architecture, with 27B parameters that are not active parameters of MoE but fully activated during inference. Native context length is 262,144 tokens, expandable via YaRN extension up to 1,010,000 tokens (about 1 million), essential for coding agents analyzing long documents or understanding across repositories. The full-precision model is 55.6 GB; with Q4_K_M quantization, compressed to 16.8 GB, it can be loaded directly on Macs with 24 GB VRAM or consumer GPUs. Licensed under Apache 2.0, no extra licensing fees for commercial use. Recommended deployment with SGLang ≥0.5.10 or vLLM ≥0.19.0; KTransformers and HF Transformers are also supported. Additionally, Qwen3.6-27B integrates a vision encoder, supporting image, text, and video understanding, not just pure text.

Why Are OpenClaw and Hermes Agent Using It as Their Local Brain?

The original news highlighted two agent frameworks: OpenClaw and Hermes Agent. OpenClaw is a popular local agent application similar to Claude Code, and was targeted by Anthropic for blacklisting—after Anthropic restricted accounts citing violation of terms, OpenClaw’s official advice was to switch to API keys or to models like Qwen or Kimi for local deployment. The open-source release of Qwen3.6-27B precisely meets this alternative route: capable of running locally, matching Claude-level benchmarks, and licensed under Apache 2.0 for free commercial use, all essential conditions.

Hermes Agent, led by NousResearch, emphasizes a “self-learning skill” loop—execute, evaluate, extract, refine, retrieve—allowing the agent to automatically distill skills after task execution and store them in memory, so similar future problems can be directly called. Compared to OpenClaw’s more straightforward approach, Hermes emphasizes long-term autonomous evolution, supporting integrations with NousPortal, OpenRouter, NVIDIA NIM, LM Studio, Ollama, and more. The common pain point for both frameworks is running a truly powerful model locally. Qwen3.6-27B scored an average of 72.4 on Claw-Eval (designed for coding agents) and 60.6 Pass^3, surpassing Claude 4.5 Opus, providing a serious local option.

Open Source vs Closed: Alibaba’s Strategic Layout and Geopolitical Landscape

The open-sourcing of Qwen3.6-27B is not an isolated event. Alibaba earlier released Qwen3.6-35B-A3B (MoE architecture, 35B total parameters, 3B active) on April 16, and this time, the 27B dense model fills the gap of “fully deployable locally without MoE sharding.” Meanwhile, Qwen3.6-Plus and Qwen3.5-Omni remain closed-source, commercialized via cloud APIs. One open, one closed—Alibaba’s strategic outline is clear: build an ecosystem and trust through open source, monetize flagship models through closed source.

A broader background is the shift in the US-China AI open-source competition. Earlier reports indicate Meta’s Zuckerberg has ordered a move away from “open-source AI,” switching to Alibaba’s Qwen-trained paid AI, Avocado—US tech giants are contracting open source, while Chinese companies are fully embracing open source. This reverse pattern is accelerating. For developers and those needing local deployment, the question is shifting from “whether to open source” to “which open-source model is good enough.” Qwen3.6-27B offers a very clear answer at this moment.