Ollama Launch + Claude Code + GLM Flash

Ollama has introduced “Ollama launch,” a one-command way to run Anthropic API–compatible coding assistants locally—making it possible to use Claude Code–style workflows with models like ZAI’s GLM 4.7 Flash. The setup is straightforward: update Ollama, pull the desired model, adjust the model’s context length, then launch Claude Code from the terminal. In practice, the workflow works, including tool use via MCP, but performance is the tradeoff.

The test centers on GLM 4.7 Flash, a smaller GLM 4.7 variant described as 30B parameters with 3B active parameters—positioned as similar in scale to mixture-of-experts models used in the Quen 3 family. With Ollama launch, the model can be run on a Mac, and the user confirms the GLM 4.7 Flash model loads correctly and can drive a “Claude Code” style experience (including plan mode and agent-like task execution).

A key requirement is context length. Ollama defaults to a 4,096-token context window, but the user found that Claude Code–style tooling and memory behavior breaks down without a larger window. The fix is to raise the context length to 64K in Ollama settings for the model. Without that change, the assistant “churns” and fails to retain enough context to use tools properly or save outputs reliably.

Even with the 64K setting, speed becomes the limiting factor. Running the model locally slows both the initial “prefill” and subsequent token decoding, and the user reports that results take noticeably longer than cloud-based Claude Code setups using larger Anthropic models. After about 90 minutes of testing on a Mac Mini Pro with 32GB RAM, the system could perform many Claude Code-supported tasks and detect MCP tools, but it sometimes produced incorrect tool arguments—behavior the user did not see with Opus 4.5. The likely causes are the quantized model variant and the very large context window.

Overall, the local approach is functional but not yet “prime time” for everyday Claude Code replacement on modest hardware. The user suggests it’s best suited for people with powerful machines (ideally with a capable GPU) and for those willing to accept slower coding-agent cycles. Still, the promise is clear: using smaller GLM/Gemma/“Quen 4s”–class models via the Anthropic API could enable cheaper, local “small agents” that run efficiently, and future coding models may make this setup practical for more users.

The takeaway is pragmatic: Ollama launch makes local Claude Code–style coding assistants easy to start, but hardware and latency determine whether it’s a viable alternative to paid Claude Code plans. The user encourages others to try it and share results, especially any setups that improve speed or tool accuracy.