🧩

🧩 Each model runs at its own documented sampling, not a shared preset. Why this matters →

🌟 Models in this lab

Four models, three providers, two of them free.

Muse Glimmer 30BHugging Face · 29.6B

Meta's local-first agent model. Strongest tool-calling score of the four.

FreeReasoning control
Kimi K2.7 CodeMoonshot · 1T MoE / 32B active

Agentic coding specialist, 256K context. The only paid model here.

~$0.95 / $4.00 per 1M
Nemotron 3 NanoNVIDIA · 30B / 3B active

Small and fast. The cheapest way to sanity-check a prompt before spending.

Free tier
Nemotron 3 SuperNVIDIA · 120B / 12B active

The largest model here. Balanced quality against the Nano at the same price.

Free tier

🔑 API Keys

Get key →
Get key →

Only the keys for the models you select are required. Keys auto-clear after 24 hours and are never persisted on our servers.

🧩 Select Models

Up to 4. Each card shows the sampling this lab will actually use for that model.

Eight lenses drawn from the risk axes these models are evaluated on. Sends an evaluation instruction in the system field; your prompt goes as written.

None selected — your prompt is sent as written. Click again to deselect.

Select a model with a reasoning control to enable this

Sent through each provider's native system mechanism. The reasoning line and the selected risk check, when either applies, sit above this text — yours reads last, so it is the most specific instruction the model sees.

📋 Notes

  • Each model runs at its own documented temperature; models with fixed sampling are sent none
  • Only temperature is plumbed — top_p and top_k are not yet supported by the provider layer, so Glimmer runs at its documented temperature but default top_p/top_k
  • Reasoning strength is applied only to models with a documented control; Nemotron's toggle exists but its syntax is unconfirmed, so it is not wired up
  • This lab is independent of /kimi and /nemotron, which are unchanged