ATRIUMsearch → argument graph
Article · 2026-06-27 · 6 moments

Using Local Coding Agents

Using Open-Weight Models in Local Coding Harnesses as an Alternative to Claude Code and Codex Subscriptions ✦ AI generated

Read original ↗
HOST Sebastian Raschka
01
Claim

A local model generating faster than 20-30 tokens/sec is fast enough for practical local agent work, comparable to GPT 5.5's high-reasoning speed.

Raschka sets a practical speed bar of 20-30 tokens/sec for local coding agents, noting this matches GPT 5.5's "high" reasoning speed and that both Qwen3.6 and North Mini Code clear it easily.

transcript

Sebastian Raschka: Anyway, the bottom line is that, in my opinion, anything faster than 20-30 tok/sec is pretty reasonable for local agent work. This is about the same speed as GPT 5.5 with "high" reasoning. In this case, both models clear the bar easily.

02
Mechanism

Claude Code's much higher token usage compared to Codex comes from its harness re-feeding a large accumulated context history into the model at each turn, not from generating longer outputs.

Investigating why Claude Code burns far more tokens than Codex or Qwen-Code, Raschka found the difference is almost entirely on the input side—one run used 578k input tokens versus just 4.5k output tokens—suggesting Claude's harness re-accumulates prior context every turn.

transcript

Sebastian Raschka: The logs suggest that Claude is repeatedly feeding more context back into the model across turns, including previous messages, tool calls, command outputs, and file contents. For example, one Claude run used about 578k input tokens but only about 4.5k output tokens across 25 turns.

03
Fact◆

Even when running a fully local model through Ollama, the Qwen-Code harness can still send telemetry and metadata to Alibaba/Aliyun servers unless explicitly disabled.

A code audit of Qwen-Code found that even with a fully local Ollama-served model, the harness still phones home usage telemetry and metadata to Alibaba/Aliyun endpoints unless disabled via settings.

transcript

Sebastian Raschka: Even with local Ollama, Qwen Code can send usage telemetry and metadata to Alibaba/Aliyun endpoints unless usage statistics and telemetry are disabled (more on that below). This is riskier than a local-only setup because model prompts may stay local, but session IDs, tool metadata, model info, and local base URL metadata can still leave the machine.

04
Claim

Local coding agents are becoming increasingly attractive because they run at near-zero marginal cost on owned hardware and offer better privacy than sending data to OpenAI or Anthropic.

Raschka argues local models are increasingly appealing due to near-zero marginal cost and privacy benefits, citing his own reluctance to send personal data like receipts to OpenAI or Anthropic.

transcript

Sebastian Raschka: Either way, local solutions become more and more attractive each day. One aspect is the costs. If you have the hardware, they are practically free to run. And then there's, of course, the privacy angle. For example, for organizing and processing my receipts, I'd be more comfortable with a local model ingesting them rather than sending the data over to OpenAI or Anthropic.

explains mechanism · 1rebuts · 1

05
Data◆

Qwen3.6 actually performs better when run through the Codex harness than through its own purpose-built Qwen-Code harness.

Contrary to the expectation that Qwen3.6 would perform best in its "native" Qwen-Code harness, Raschka's benchmark showed it actually did better when run through Codex, suggesting harness-model pairing may matter less than assumed.

transcript

Sebastian Raschka: When rerunning the test cases from the Agent Capability Assessment section, to my surprise, Qwen3.6 does actually perform better via Codex compared to its "native" Qwen-Code coding harness, as shown below.

06
Data

Qwen models are specifically optimized to perform best within the Qwen-Code harness compared to other coding agent harnesses.

Raschka cites Nvidia's Polar benchmark showing Qwen3.5-4B performs best specifically within the Qwen-Code harness, supporting his choice to pair Qwen models with Qwen-Code.

transcript

Sebastian Raschka: Regarding the second point in the list above, that Qwen models work better in Qwen-Code, Nvidia's Polar: Agentic RL on Any Harness at Scale paper (May 2026) has a benchmark showing that the Qwen3.5-4B base model has the best coding performance in said Qwen-Code harness (both before and after their Polar-RL training), which I included below.

rebuts · 1

Highlight slides
Harness Surprise: Codex Beats Qwen-Code✦ from: Qwen3.6 actually performs better when run through the Codex harness than through its own purpose-built Qwen-Code harness.Qwen-Code Phones Home Despite Local Ollama✦ from: Even when running a fully local model through Ollama, the Qwen-Code harness can still send telemetry and metadata to Alibaba/Aliyun servers unless explicitly disabled.What Data Leaves the Machine✦ from: Even when running a fully local model through Ollama, the Qwen-Code harness can still send telemetry and metadata to Alibaba/Aliyun servers unless explicitly disabled.What This Suggests✦ from: Qwen3.6 actually performs better when run through the Codex harness than through its own purpose-built Qwen-Code harness.Why This Is Riskier Than Expected✦ from: Even when running a fully local model through Ollama, the Qwen-Code harness can still send telemetry and metadata to Alibaba/Aliyun servers unless explicitly disabled.
Related episodes
Article · Simon Willison's Weblog

Fable's judgement

→
Article · Simon Willison's Weblog

Better Models: Worse Tools

→
Article · Simon Willison's Weblog

sqlite-utils 4.0rc2, mostly written by Claude Fable (for about $149.25)

→