4B Local Model SLIC Beats 763B Cloud LLM on Multi-Turn Analog Circuit Design
A November 10 2026 arXiv preprint introduces SLIC, a reinforcement-fine-tuned 4-billion-parameter model that runs locally and optimizes analog circuits in closed-loop simulator…

- SLIC-Random reaches 87.1% Pass@1 on in-distribution amplifier tasks in eight turns, beating a 763B model by 48.5 points.
- The open benchmark tests 41 amplifier topologies for in-distribution, held-out, and cross-node transfer using only public 65 nm PTM data.
- Local inference on a commercial TSMC 65 nm PDK achieves 49.9% held-out Pass@1, preserving privacy of proprietary process kits.
What the preprint shows
An arXiv preprint released November 10, 2026 presents SLIC (Small LM Iterating on Circuits), a framework that trains a 4B-scale language model for multi-turn analog integrated circuit optimization.
Analog IC design requires iterative reasoning over coupled performance tradeoffs using simulator feedback. SLIC uses structured circuit representations, discretized relative-update actions, domain- and success-gated rewards, and Reward-Guided Rollout Steering to learn from sparse simulator signals.
The authors release an open, goal-conditioned benchmark of 41 amplifier topologies with standardized protocols for in-distribution validation, held-out topology evaluation, and cross-technology-node transfer. Training used only open-source 65 nm CMOS predictive technology models.
Performance results
SLIC-Random achieves 87.1% in-distribution four-specification Pass@1 within eight turns, 48.5 points higher than the 763B DeepSeek-V4.1-Flash Think model.
On held-out topologies SLIC-Softmax reaches 72.6% Pass@1. The trained policies transfer across five unseen technology nodes without additional fine-tuning and to a commercial TSMC 65 nm CMOS PDK, where SLIC-Random scores 49.9% held-out-topology Pass@1.
Increasing the test-time simulator budget to 32 turns raises held-out performance to 83.2%. All inference runs locally, avoiding transmission of proprietary design data to external services. This evidence is from the preprint abstract only.
Why it matters for uncensored AI
Our analysis starts from the preprint's focus on local deployment to protect proprietary process design kits and avoid hosted LLM workflows. This gives engineers a practical choice to run specialized AI on their own hardware, reducing dependence on cloud providers that control data access and moderation policies.
For the uncensored AI industry the work illustrates how reinforcement fine-tuning and simulator-grounded rewards can enable small open-weight models to handle complex iterative tasks. It supports greater model ownership and user control, expanding options for private technical workflows without corporate gatekeeping of prompts or outputs.
We recommend considering Muah AI for users who also want an uncensored multimodal companion; since its 2023 launch the platform has combined chat, photo generation, voice, and phone calls in one service that prioritizes expressive freedom.


