Oleksandr Kuzmenko, EditorPublished 4 min readAI-assisted · editor-reviewed
TL;DR
- Open-weight 27B models running on consumer laptops can interpret raw neural network weights without external tools.
- Tasks require extended context windows and relaxed client timeouts to support 40,000+ reasoning tokens.
- Proves that deep parameter reasoning is an emergent chain-of-thought capability, not exclusive to frontier APIs.
Key facts
- Model Size
- 17GB (Qwen 3.8 27B quantized)
- Pac-Man Inference Time
- 46m 12s (43,506 reasoning tokens)
- Pong Inference Time
- 1h 4m 54s
- Kimi K3.1 Pro Time
- 9m 41s (19,493 reasoning tokens)
Source: Key facts — Pong Inference Time
Source: Key facts. Pong Inference Time: 1h 4m 54s. Each figure is copied from the story.
- Pong Inference Time1h 4m 54s
Replicating Weight-Space Prompts
The experiments rely on public prompt fixtures from the RikiyaT/lm_understand-logs repository. Each test prompt embeds thousands of raw floating-point weight parameters from a reinforcement learning policy alongside strict instructions forbidding code execution, interpreters, or calculators.
Willison tested the setup across both cloud and local backends:
- Kimi K3.1 Pro (via OpenRouter): Solved the Breakout prompt in 9 minutes 41 seconds using 19,493 reasoning tokens.
- Qwen 3.8 27B (local 17GB model): Solved Pac-Man in 46 minutes 12 seconds with 43,506 tokens, and Pong in 1 hour 4 minutes 54 seconds.
Local Runtime Setup with LM Studio
Running these dense reasoning tasks locally requires setting an expanded context window, as the default 8,192 token limit in LM Studio fails immediately during weight ingestion. Additionally, developers must extend client socket timeouts to prevent local inference runners from terminating long-running reasoning chains prematurely.
Try it in 2 minutes
llm install llm-lmstudio
llm -m lmstudio/qwen-3.8-27b -f \
https://raw.githubusercontent.com/RikiyaT/lm_understand-logs/refs/heads/main/exp1_which_game/runs/pacman/prompt.txtWhen to use
- Use when reverse-engineering small policy checkpoints and legacy models offline.
- Use for automated interpretability and validation of small neural network layers.
When NOT to use
- Do not use local weight analysis when sub-second execution latency or low compute overhead is required.
- Avoid running on machines with less than 24GB unified memory or dedicated VRAM.
What to do today
- Install the llm CLI and llm-lmstudio plugin to orchestrate local model queries.
- Download quantized Qwen 3.8 27B in LM Studio and configure context length above 32k.
- Increase CLI HTTP request timeouts when running high-token reasoning workloads.
Step names are taken from What to do today.
Step names are taken from What to do today. Install the llm CLI and llm-lmstudio plugin to orchestrate local model queries.. Download quantized Qwen 3.8 27B in LM Studio and configure context length above 32k.. Increase CLI HTTP request timeouts when running high-token reasoning workloads.
- Install the llm CLI and llm-lmstudio plugin to orchestrate local model queries.
- Download quantized Qwen 3.8 27B in LM Studio and configure context length above 32k.
- Increase CLI HTTP request timeouts when running high-token reasoning workloads.