---
id: "lib-2026-11-03-qwen-3-8-27b-on-my-laptop-can-interpret-model"
type: "library"
title: "Qwen 3.8 27B on my laptop can interpret model weights, too"
lang: "EN"
org: "Simon Willison"
source_type: "Practitioner post"
date: "2026-11-03"
week: "2026-W45"
url: "https://nosov.ai/library/lib-2026-11-03-qwen-3-8-27b-on-my-laptop-can-interpret-model/"
original_url: "https://simonwillison.net/2026/Nov/3/qwen-can-interpret-model-weights/"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "Simon Willison, a widely followed independent developer, re-runs one experiment from the MIT weight-space paper on a laptop model with commands and transcripts linked; a useful informal replication of one prompt per game, one run each."
summary_basis: "full-text"
original_minutes: "6"
tags: ["ai technology", "open weights", "local models", "interpretability"]
---

# Qwen 3.8 27B on my laptop can interpret model weights, too

Tier B (Useful, limited): Simon Willison, a widely followed independent developer, re-runs one experiment from the MIT weight-space paper on a laptop model with commands and transcripts linked; a useful informal replication of one prompt per game, one run each.

Summary based on the full text.

## In short

A 3 November 2026 post by Simon Willison re-running Experiment 1 of the MIT CSAIL "Sparks of Weight-Space Programming" post: identify which grid game a small policy network was trained on from its raw weights, with no tools. Using the published prompts, Kimi K3.1 Pro got Breakout right, and Qwen 3.8 27B running on his laptop identified Pac-Man, Pong and Breakout but not Freeway. He argues the paper mixes two questions, reading weights without tools and editing models with code, and that some commentary around it goes too far.

## Key numbers

- 3 of 4: games Qwen 3.8 27B identified from raw policy weights on a laptop (failed on Freeway)
- 17GB: size of the Qwen 3.8 27B model file run locally in LM Studio
- 46m 12s: time for Qwen to identify the Pac-Man network, with 43,506 reasoning tokens
- 9m 41s: time for Kimi K3.1 Pro via OpenRouter to identify Breakout, with 19,493 reasoning tokens

## What it reports

- The paper's prompts are on GitHub and are identical apart from the weights, so anyone can re-run them.
- Kimi K3.1 Pro (via OpenRouter) named Breakout after 9m 41s and 19,493 reasoning tokens.
- Qwen 3.8 27B (17GB, in LM Studio with a long context window) named Pac-Man in 46m 12s with 43,506 reasoning tokens (72% confidence), Pong in 1h 4m 54s (75%) and Breakout in 52m 25s (88%).
- Qwen misidentified Freeway as a Pac-Man-like game.
- The default 8,192-token context failed; a longer context and a raised timeout were needed.
- Willison cites Matthew Honnibal calling the paper misleading for conflating the no-tools and with-code results, and Sean Goedecke's view that the Qwen3.5-0.8B fix followed steps a human ML researcher would take.
- His view: an experienced person with pencil and paper could likely get there with enough effort.

## How they got it

Willison ran the paper's published Experiment 1 prompts through his `llm` command-line tool against a hosted model and a local model, one run per game, and linked the reasoning transcripts.

## Limits

- One run per game and only the simplest experiment from the paper; no repeated trials or controls.
- Covers the no-tools reading task only, not the weight-editing results.
- The Qwen answers describe an 8x8 board, while the paper's setup is a 10x10 grid, so how much the model actually read from the weights is unclear.

## Why it matters for companies

The replication shows the paper's simplest claim can be checked cheaply by anyone, including with an open model on a laptop, which is a good habit before acting on a striking AI result. It also separates what is shown (small networks can be read from weights, slowly) from what is not (models patching production systems by editing weights). For planning, treat weight-space work as research to watch.

Original: https://simonwillison.net/2026/Nov/3/qwen-can-interpret-model-weights/