Ember 8B (GGUF)

Ember is the AI partner built into RusticEngine, a game engine based on Godot 4.7 where games are written in Rust (godot-rust 0.5).

Ember is meant to work with a creator rather than build a whole game from one prompt. It asks about the creator's idea and inspirations before making creative choices, explains what it builds when the creator wants to learn, and guides people to make their own art in RusticEngine's Studio instead of making it for them.

New in RusticEngine 2.0: Ember also teaches Blender, from modeling concepts to keyboard shortcuts, and can answer questions from inside Blender. That comes from RusticEngine's built-in Blender library and guides, not from retraining: this is the same model as in 1.0.

RusticEngine downloads this model automatically the first time you open the AI Guide. You only need this page if you want to run Ember yourself.

What it is

  • Base: Qwen3-8B (Apache-2.0). Fine-tuned with QLoRA (rank 16) on several hundred RusticEngine agent traces: real tool calls (reading and editing code, cargo builds with compiler feedback, headless playtests, Godot API lookups), plus collaboration conversations (kickoff questions, learning mode, concept art, Studio guidance).
  • File: ember.gguf (SHA-256 5b0017088220f714af249405a5ba5db5a93bb5880bb29072e727b127cc9fa02c).
  • Format: GGUF, mixed 4/5-bit (attention and ffn_down in Q5_0, ffn_gate/up and embeddings in Q4_0, output in Q8_0), about 5.5 GB.
  • Chat format: Qwen3 ChatML with <tool_call> tool calls, non-thinking mode (an empty <think></think> block before each reply).

How well it does

On RusticEngine's own evaluation (32 unseen godot-rust coding tasks, scored by whether the code builds and the game runs), Ember with RusticEngine's agent harness built about 17 of 32 and ran about 16 of 32 per pass. Base Qwen3-8B scored roughly 7 of 16 on an earlier, smaller version of the same evaluation. On held-out Studio and Mechanics questions it gives grounded, step-by-step guidance (34/34 and 9/9 on the current probes). Asked for new game mechanics it hasn't seen (a dash, a door, checkpoints, a patrolling enemy...), it delivers ones that build and do what was asked in most cases (6 of 8 in the latest full run, up from 0 of 5), by starting from RusticEngine's tested mechanic variants.

It is an 8B model, so it still invents Godot methods sometimes. RusticEngine's harness compensates: builds feed compiler errors back to it, missing imports are fixed automatically, and API lookups are offered before it writes code. For harder work, RusticEngine can also use a frontier model (Claude, OpenAI and others) in the AI Guide settings.

Running it yourself

  • RusticEngine: built in, nothing to do.
  • llama.cpp: llama-server -m ember.gguf --jinja. Turn thinking off by ending prompts with an empty <think>\n\n</think>\n\n block, as in training.
  • Ollama: create a model from this GGUF using Qwen3's template, with think: false.

The tools Ember was trained with come from RusticEngine's agent (rustic-agent --print-tools). RusticEngine can also act as an MCP server (rustic-agent --mcp), so other agents can use the same tools.

License

Apache-2.0, like the Qwen3 base model. RusticEngine itself is GPL-3.0.

Downloads last month
297
GGUF
Model size
8B params
Architecture
qwen3
Hardware compatibility
Log In to add your hardware

We're not able to determine the quantization variants.

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for johangdarkdev/Ember-8B-GGUF

Finetuned
Qwen/Qwen3-8B
Quantized
(522)
this model