Qwen 3.5-35B-A3B: Local LLM deployment requires significant scaffolding

This review analyzes the practical challenges and solutions for deploying Qwen 3.5-35B-A3B locally with OpenClaw, focusing on chat template, context management, and cron job pitfalls.

TL;DR

Best for: Developers seeking a high-performance local LLM (Qwen3.5-35B-A3B) for specific tasks, willing to invest significant effort in custom configuration and scaffolding for agentic workflows. Skip if: You require out-of-the-box agentic capabilities, seamless tool calling, or robust context management without extensive manual setup and debugging. Bottom line: Qwen3.5-35B-A3B offers strong raw performance for local inference but demands substantial engineering to function reliably in complex, agent-driven environments.

METHODOLOGY

This v0 review draws on the founder's published claims and detailed problem-solving narrative in a dev.to blog post. We analyzed the specific technical challenges encountered during the deployment of Qwen3.5-35B-A3B within the OpenClaw framework. The review covers the founder's reported configuration issues related to chat templates, tool calling, agent context management, and cron job misconfiguration. We extracted specific data points, version numbers, and the multi-step problem-solving narrative as presented in the source. This review does not include independent performance benchmarks, long-term workflow assessments, or edge-case testing. Update cadence: This tool will be re-tested and this review updated when claims diverge from observed behavior or when new, verifiable data becomes available.

• Tool name + version + date observed: Qwen3.5-35B-A3B, observed May 26, 2026 (based on ingestion date).
• Source signal URL: https://dev.to/carryologist/qwen-is-not-yet-ready-to-power-local-openclaw-deployments-5ha3
• What's covered in this review: Founder's reported experience with Qwen3.5-35B-A3B's chat template, tool calling, agent context management, and cron job integration within OpenClaw. Specific technical details and solutions are discussed.
• What's NOT covered: Independent performance benchmarks, long-term workflow stability, comprehensive error handling, or comparison against other local LLMs beyond the initial model showdown claim.

WHAT IT DOES

Qwen3.5-35B-A3B is presented as a high-performance large language model capable of local deployment, winning an initial model showdown with an 85.3 weighted score and 206 tok/s on an RTX 5090. However, its practical integration into agentic workflows like OpenClaw revealed several critical configuration gaps.

Chat template configuration

Initially, Qwen's responses in OpenClaw showed raw <think>...</think> tags and failed to produce proper OpenAI-compatible tool_calls objects, instead outputting plain text like create_workspace. The root cause was using a minimal --chat-template chatml that lacked tool-calling awareness. Qwen3.5 ships with a 154-line Jinja template designed for these features. The founder addressed this by switching to --chat-template-file pointing to a patched version of Qwen's native template. This patch modified a strict ordering check that previously raised exceptions if a system message appeared anywhere but the conversation's beginning, allowing Coder Agents' out-of-order system messages to render correctly.

Agent context management

Qwen demonstrated a tendency to follow instructions literally without inferring necessary context. For example, when asked to clone