GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
arc-director
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
arc-director
Model Tree
google/gemma-4-E4B-it
Input Modalities
Output Modalities
Supported Functionality
The ArcDirector Bench (v2, part of ArcDirector library) with every case played 10 times at temperature 1. Each cell is the share of goals reached. Tested on a RTX 5090 GPU with 32GB of VRAM.
| Skaldic Gemma 4 E4B | Gemma 4 31B | Gemma 4 E4B (base) | |
|---|---|---|---|
| Total | 55% ±4 | 63% ±3 | 16% ±3 |
| Easy | 64% | 73% | 18% |
| Medium | 59% | 59% | 22% |
| Hard | 42% | 58% | 8% |
| Chain | 63% | 75% | 25% |
| Trade | 83% | 100% | 17% |
| Social | 88% | 100% | 25% |
| Lookup | 67% | 100% | 33% |
| Questing | 100% | 100% | 38% |
| Reactions | 17% | 0% | 0% |
| Fight | 60% | 60% | 20% |
| Tactics | 33% | 8% | 0% |
| Economy | 10% | 40% | 0% |
| Planning | 50% | 83% | 17% |
| Harm rate | 1.3% | 1.5% | 5.5% |
| Time/response | 17.1 s | 64.5 s | 37.1 s |
Totals are ± one standard error over 10 runs. Each skill rests on 3–6 situations, so skill gaps under about 20 percentage points are within noise.
The ArcDirector library recognizes Skaldic models by name: it gives them the compact prompt they were trained on and asks them for JSON Arcs. This reduces the amount of time it takes to get a response from the model by a factor of 3–4.
This is the simplest way to use it:
ts
import { ArcDirector, BrowserProvider, Skaldic } from 'arc-director'; import wasm from '@wllama/wllama/esm/wasm/wllama.wasm?url'; // vite// const wasm = new URL('@wllama/wllama/esm/wasm/wllama.wasm', import.meta.url).href; // webpack BrowserProvider.shared = new BrowserProvider({ wasm }); const director = new ArcDirector() .withLoadingBar() // Shows a loading bar on screen while it downloads .addBrowserModel('main', Skaldic.Gemma4E4B);await director.start();It is an ordinary Gemma 4 E4B: the GGUFs run in llama.cpp, LM Studio, Ollama and other GGUF runners, and the safetensors in 🤗 Transformers. Out of ArcDirector it still chats, but it is at its best with ArcDirector's prompts.
| File | Size | For |
|---|---|---|
skaldic-gemma-4-e4b-Q8_0.gguf | 7.9 GB | best quality |
skaldic-gemma-4-e4b-Q4_K_M.gguf | 4.6 GB | smaller GPUs |
mmproj-gemma-4-E4B-it-BF16.gguf | 1.0 GB | image input, beside either GGUF |
browser/skaldic-gemma-4-e4b-Q4_K_M-*-of-00009.gguf | 4.6 GB in 9 parts | the browser (each file under 2 GB) |
model.safetensors and configs | 15.9 GB | 🤗 Transformers, further fine-tuning, your own quantizations |
Fine-tuned on thousands of game situations across multiple genres, each played out in ArcDirector, keeping only the plans that reached their goal. It was trained on ArcDirector's compact prompt and answers in JSON Arcs.
It also works as ArcDirector's decision model (System One), though it has had only light training for that.
Questions and issues: ArcDirector on GitHub.
Apache 2.0, as Gemma 4.
import { ArcDirector, BrowserProvider, Skaldic } from 'arc-director'; import wasm from '@wllama/wllama/esm/wasm/wllama.wasm?url'; // vite// const wasm = new URL('@wllama/wllama/esm/wasm/wllama.wasm', import.meta.url).href; // webpack BrowserProvider.shared = new BrowserProvider({ wasm }); const director = new ArcDirector() .withLoadingBar() // Shows a loading bar on screen while it downloads .addBrowserModel('main', Skaldic.Gemma4E4B);await director.start();