Bud Decision Studio serves eleven local decision models on port 8420

Bud Ecosystem posted Bud Decision Studio on October 2, 2026. The repository is BudEcosystem/Bud-Decision-Engine. The README describes a desktop app for Macs with Apple Silicon, Windows 10 and 11, and Linux on x64 and ARM64.

GitHub’s license field is empty, and there was no LICENSE file at the repository root when we opened it. The post says the weights are open models, and the README says each model keeps the license on its own Hugging Face page.

While the app is open, the server listens at http://127.0.0.1:8420. POST /v1/systemone follows TypeSafe’s request shape, so a typesafe-sdk client points at that base URL and sends model laya. The README also lists OpenRouter’s decisions path, Vercel’s TypeSafe route, and the studio’s own routes under /v1/studio for templates and history.

Eleven models, then a training table

The catalog is Julia 1 at 144M, Laya Multilingual at 322M, Laya and Laya Typed-Decisions at 421M, Kev 0.5B, GLiNER2.5 Decide at 340M, Intern-Decision 4B, Kev 4B, Lev at 4B, CLM 8B, and Jev-Omni at 12B. Jev-Omni is the one the README marks for images, audio, and video, and it says that model needs a GPU. Weights stay in the Hugging Face cache.

The trainer’s table is the studio’s own runs on an NVIDIA GB10. Examples were held out before training. The README does not print how many. The “general decisions” column is a delta, and the README does not define the unit in the table itself.

Julia 1 on support tickets goes from 52 percent to 91 percent, general decisions -0.4. Laya on policy topics goes from 59 percent to 80 percent, +0.2. Laya Multilingual on business workflows goes from 34 percent to 62 percent, +12.7. Laya Typed-Decisions on policy topics goes from 61 percent to 81 percent, +0.2.

GLiNER2.5 Decide goes from 66 percent to 75 percent, +0.7. Kev 0.5B goes from 65 percent to 79 percent, +0.6. Kev 4B goes from 77 percent to 82 percent, +0.2. Intern-Decision 4B on 16 conversation emotions goes from 61 percent to 76 percent, +0.2. Lev goes from 75 percent to 82 percent, +2.5.

CLM 8B on business workflows goes from 39 percent to 68 percent, +11.2. Jev-Omni on business workflows goes from 62 percent to 77 percent, and the general-decisions cell is 0.0.

The README’s test list is 14 of 14 API conformance checks, 89 of 89 studio API tests, 21 of 21 browser checks across the eleven models, and 34 of 34 trainer core tests.

The desktop line says the first-run check ran on Linux ARM64 with an NVIDIA GB10. Training itself asks for an NVIDIA RTX 30 series or newer, including that GB10, or an Apple M2 or newer.