Instella-Prometheus 🔥

Stole code-fire from the gods. Gave it to the people.

No system prompt. No thinking tags. No decoder bans. No adapter dependencies. Just a bare user prompt and clean Python.

stamsam/Instella-Prometheus — 16B total / 2.8B active parameters — fine-tuned on powerhouse Python code traces from four teacher models and fused permanently into the base weights. It loads with a single from_pretrained() call and zero extra dependencies.

Four Ways to Run It

This repository contains four downloadable weight variants. Choose the one that fits your hardware and runtime:

Variant	File	Approx. size	Best for
Full Transformers weights	7 model-0000X-of-00007.safetensors shards	~30 GB	transformers, training, maximum flexibility
Q8_0 GGUF	Instella-Prometheus-Q8_0.gguf	~16.9 GB	Near-full-quality llama.cpp inference
Q4_K_M GGUF	Instella-Prometheus-Q4_K_M.gguf	~9.4 GB	Best quality/size balance
Q3_K_M GGUF	Instella-Prometheus-Q3_K_M.gguf	~8.2 GB	Lower-memory llama.cpp inference

The three .gguf files are in the repository root so Hugging Face can identify them as quantized variants. The GGUF files are standalone: they include the model metadata needed by llama.cpp and do not require the Transformers weights.

The Theft

The gods kept their fire locked away — DeepSeek V4, Qwen 3 Coder 480B, Claude Opus, and the competitive programming pantheon. Each held a piece of what it means to write clean, functional Python. Prometheus stole from all of them, distilled their fire into 26,028 records of AST-valid, think-free, code-first answers, and baked it into a single 30GB artifact that answers to no system prompt and bows to no decoder ban.

The Proof

Metric	Base Instella	Instella-Prometheus	Bar
Code blocks generated	2 / 12	12 / 12	—
Syntax valid	2 / 12	12 / 12	—
Functional passes	2 / 12	10 / 12	≥9 ✅
Think tag leak	12 / 12	0 / 12	≤3 ✅
Natural EOS	2 / 12	12 / 12	—

The base Instella-MoE-SFT produces almost no code under a bare prompt — it defaults to thinking tags and narrative text. Prometheus writes clean, functional Python directly, with zero thinking leakage and perfect EOS completion. That's a 5× improvement in functional passes and a complete elimination of think-tag leakage — no system prompt, no decoder bans, no adapter dependencies.

The Training

Three phases, each building on the last:

1. v1 — 4,000 records from agentic coding traces (kimi-k3, fable-5, hermes-agent). Proved the concept: 4/12 functional, 11/12 think leak.
2. v2 — 300 steps continuation from v1 checkpoint-500. Same data, bare prompt. Hit 11/12 functional, 0/12 think — but the dataset was too small to bake it in permanently.
3. v3 (powerhouse) — 300 steps from v2 checkpoint-100 on a 26,028-record powerhouse pool. The LoRA was merged permanently into the base weights. The result: 10/12 functional, 0/12 think, 12/12 EOS — standalone, no crutches.

Source	Teacher	Records	License
Competitive-Programming-python-blend	Contest problems (multi)	15,014	Permissive mix
qwen3-coder-480b-distill-mini	Qwen 3 Coder 480B	8,303	Apache-2.0
DeepSeek-V4-Distill-8000x	DeepSeek V4	2,148	MIT
Claude-Opus 4.6/4.7 (code split)	Claude Opus	454	Apache-2.0
v1 unique carryover	(kimi-k3, fable-5, hermes)	109	CC-BY-4.0

Every record: AST-valid Python, no thinking tags, code-first, provenance preserved.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "stamsam/Instella-Prometheus",
    trust_remote_code=True,
    torch_dtype="auto",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(
    "stamsam/Instella-Prometheus",
    trust_remote_code=True,
)

# No system prompt. No thinking preamble. Just the task.
prompt = "Write a function to merge overlapping intervals."
messages = [{"role": "user", "content": prompt}]
inputs = tokenizer.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
).to(model.device)

output = model.generate(
    inputs,
    max_new_tokens=1024,
    do_sample=False,
    eos_token_id=tokenizer.eos_token_id,
    pad_token_id=tokenizer.eos_token_id,
)
print(tokenizer.decode(output[0][inputs.shape[1]:], skip_special_tokens=True))

GGUF Quantizations

The three GGUF quantizations are in the repository root and are suitable for llama.cpp and its bindings.

File	Size	BPW
Instella-Prometheus-Q8_0.gguf	16.9 GB	8.94
Instella-Prometheus-Q4_K_M.gguf	9.4 GB	5.28
Instella-Prometheus-Q3_K_M.gguf	8.2 GB	4.14

For the original Transformers format, use the seven model-0000X-of-00007.safetensors shards plus model.safetensors.index.json.

License

Apache-2.0. The base model (Instella-MoE-16B-A3B-SFT) is released by AMD under a research-permissive license. Training data sources carry their own licenses (CC-BY-4.0, Apache-2.0, MIT, BSD) — all permissive for redistribution.

Acknowledgements

Built on the shoulders of giants: AMD (Instella-MoE), greghavens (coding traces), Jackrong (distillation datasets), angrygiraffe (Opus reasoning traces), and the open-source community. The fire belongs to them. The torch is yours.