Standalone RL-MPQ (Reinforcement Learning Mixed-Precision Quantization) checkpoint for the Balanced scenario — a quantized variant of jetmoe/jetmoe-8b.

Field	Value
Base model	jetmoe/jetmoe-8b
Scenario	Balanced
Avg bits / weight	4.5
Compression vs FP16	3.5556×
WikiText-2 PPL	4.6648
Layers	24
Bit distribution	`{'4': 21, '8': 3}`
Format	Fake-quant FP16 + `rlmpq_policy.json`

Collection: RL-MPQ — JetMoE 8B — all five scenarios for JetMoE 8B.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "AvoCahDoe/jetmoe-8b-rlmpq-balanced"

model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="float16")
tokenizer = AutoTokenizer.from_pretrained(repo)

Other JetMoE 8B scenarios

Scenario	Avg bits	Compression	WikiText-2 PPL
Aggressive	3.5833	4.4651x	4.6648
Conservative	5.3333	3.0x	4.6648
Extreme Survival	2.625	6.0952x	4.6648
High Fidelity	6.8333	2.3415x	4.6648

Grouped archive (all scenarios in one repo): AvoCahDoe/jetmoe-8b-rlmpq

Method

Phase 3 — PPO agent assigns per-layer bit widths under the Balanced reward target.
Phase 4 — Policy replayed on real weights; WikiText-2 perplexity validates quality.
Export — Fake-quantized FP16 weights compatible with Hugging Face Transformers.

Files

File	Description
`config.json`	Llama architecture + RL-MPQ metadata
`model.safetensors`	Fake-quantized weights
`rlmpq_policy.json`	Per-layer bit-width policy
`rlmpq_metrics.json`	Validation & PPL summary

Citation

bibtex

@misc{rlmpq_jetmoe_8b_balanced_2026,
  title  = {RL-MPQ Balanced: JetMoE 8B Mixed-Precision Quantization},
  author = {AvoCahDoe},
  year   = {2026},
  url    = {https://huggingface.co/AvoCahDoe/jetmoe-8b-rlmpq-balanced}
}