RLSPICERun 1
6 waiting for training to startin the watch party
Four open models train live to beat the frontier at SPICE.
RLSPICE is a reinforcement-learning run you can watch. Each model gets a component’s datasheet, builds a SPICE model of the part and checks it in a circuit simulator. Then it’s graded on curves it never saw. The bar is GPT‑6.1 Sol, the frontier model behind Coppr’s model builder today.
2 November 2026, 22:00 UTC
Training starts 2 November 2026, 22:00 UTC.
Training started 2 November 2026, 22:00 UTC. Results post to this page as they come in.
We’ll email you when training starts.
Figure 1How an attempt is graded
The contenders
Same tasks, same tools, same simulator, same held-back tests. Each model is scored before training and again after it, so you can see what training added.
| Model | From | Parameters | Architecture | Licence |
|---|---|---|---|---|
| Muse Glimmer-30B | Meta | 29.6B | Dense | Apache 2.0 |
| Gemma 4 31B | 30.7B | Dense | Apache 2.0 | |
| Gemma 4 26B-A4B | 25.2B, 3.8B active | Mixture of experts | Apache 2.0 | |
| Qwen3.8-27B | Qwen | 27B | Dense | Apache 2.0 |
| GPT-6.1 SolCoppr’s model builder today | OpenAI | Undisclosed | Undisclosed | Proprietary, API only |
If a model can’t run the environment at all, it’s withdrawn and we publish the reason. No model gets a different prompt, more attempts or a bigger budget than the others.
How each attempt is scored
An attempt is one part, start to finish. Training repeats it thousands of times, and the model is rewarded only for what the simulator confirms.
-
The datasheet
The model gets the part’s typical curves and table values as plain numbers: forward voltage, leakage, capacitance, each at its stated temperature. Some curves are held back.
-
The work
It sets parameters, fits them with a numerical optimiser, runs the simulator and reads the result. It can try as often as it likes within a fixed number of tool calls.
-
The submission
It submits one model it actually simulated, and states what the datasheet can’t tell it.
-
The grade
The simulator runs the submitted model on the held-back curves and measures the error against each curve’s tolerance. A simulation that fails scores zero.
Why the score means something
A model under training will exploit any gap in how it’s graded. These rules close the gaps we know about, and the rest is published so you can look for more.
Graded on what it never saw
Matching the curves it was shown earns little. Most of the score comes from the held-back ones, which only a model with the right physics predicts.
“Not enough evidence” beats a confident guess
When the datasheet really doesn’t contain what’s needed, saying so earns partial credit. Claiming ignorance when the evidence is there earns nothing.
The test can’t shrink
Each task fixes what gets graded. Nothing a model writes in its submission changes that.
The referee was tested first
Coppr’s verification caught ngspice, the open-source simulator, reporting an AC gain of 4.6×1015 in a plain resistor divider. Our fix was merged upstream. The run uses that patched build, pinned.
Practice parts, then real ones
Training uses synthetic parts whose physics we know exactly. Final scores come from real datasheets the models never trained on.
Every number goes public
Training reward and held-back score sit side by side, so a model that games its reward shows up. Runs that fail stay on the page.
What you’ll see during the run
- The leaderboardEach model’s score on held-back curves, updated as new checkpoints are graded.
- Attempt replaysOne part, start to finish: every tool call, every simulation, the grade.
- Daily notesWhat improved, what broke, and what we changed, with the reason.
We’ll email you when training starts.