AETHER-Mythos-1 is a specialist agentic coding model with strong internal reasoning. It is designed for on-device / local deployment: low latency, modest VRAM/RAM, and high-signal tool-use + planning behavior.

Philosophy

The highest-leverage path to a small coding agent is not more web text — it is distilling the best long-horizon agent trajectories (think → tool → observe → verify) into an efficient backbone. We prioritize:

Fable 5 traces (Glint-Research/Fable-5-traces) as the primary high-signal source of Claude Fable 5 thinking + tool-use coding sessions.
Complementary elite CoT coding / reasoning data to reinforce planning and verification without drowning the mix in noise.
LiquidAI LFM2.5 as the substrate: hybrid architecture, strong edge speed, long context, and Unsloth-friendly fine-tuning.

Base model

Base: LiquidAI/LFM2.5-1.2B-Thinking
Context trained: up to 32768 tokens (packing enabled)
Method: LoRA (r=128, alpha=256)

Data mixture

fable5_cot (Glint-Research/Fable-5-traces) weight=0.65 — Primary identity and agent trace signal
opencode_reasoning (nvidia/OpenCodeReasoning) weight=0.18 — prompt_completion
open_r1_codeforces (open-r1/codeforces-cots) weight=0.07 — messages
openthoughts_code (open-thoughts/OpenThoughts-114k) weight=0.10 — messages

Data provenance & licenses

Source	Role	License (as published on Hub)
`Glint-Research/Fable-5-traces`	Primary agent CoT + tool traces (`fable5_cot_merged.jsonl`)	AGPL-3.0
Complementary CoT coding sets (see mixture above)	Secondary planning / verification signal	Per-dataset Hub terms

AGPL-3.0 notice: A substantial fraction of training signal comes from AGPL-licensed agent traces. Distributing model weights derived primarily from AGPL data may trigger strong copyleft obligations (source disclosure for network use in some interpretations). Do not treat this model as Apache/MIT-clean. Review AGPL compatibility with counsel before commercial or proprietary deployment. The base model (LiquidAI/LFM2.5-1.2B-Thinking) remains under Liquid AI’s LFM license terms.

Training setup

Setting	Value
GPU	L40S (Modal)
Effective batch size	16
Learning rate	8e-05
Schedule	cosine
Epochs / max steps	1.0 / 100
Packing	True
Optim	adamw_8bit
Grad checkpointing	unsloth
Seed	3407

Stack: Unsloth + TRL SFT on Modal with persistent volumes for dataset cache and checkpoints.

Intended use

Local coding agents (tool-use loops: shell, edit, read, write)
Planning + verification style reasoning before code changes
Edge / laptop / NPU-friendly deployments via GGUF / MLX / llama.cpp

Not intended for: unconstrained autonomous operation on production systems without human oversight; high-stakes decisions; generating malware or disallowed content.

Chat & thinking format

AETHER-Mythos-1 follows LFM2.5 ChatML-style templates. Assistant turns may include:

<think>
... internal reasoning ...
</think>
final answer or tool call

Tool calls use LFM tokens:

<|tool_call_start|>[tool_name(arg="value")]<|tool_call_end|>

Inference tips (LFM2.5 Thinking defaults)

temperature ≈ 0.05
top_k = 50
repetition_penalty ≈ 1.05

Limitations

Distilled from agent traces; may inherit tool schemas and path conventions from source data.
Context rows in Fable-5 merged JSONL may be truncated at the source.
Small models can still hallucinate APIs, file state, or test results — always verify.

Citation

bibtex

@misc{aether-mythos-1-2026,
  title = {AETHER-Mythos-1: Efficient Agentic Coding via Fable 5 Distillation on LFM2.5},
  year = {2026},
  howpublished = {\url{https://huggingface.co/enosislabs/AETHER-Mythos-1-1.2B}}
}

Acknowledgements

Liquid AI — LFM2.5 family
Glint Research / TeichAI ecosystem — Fable 5 trace corpora
Unsloth — efficient fine-tuning
Modal — GPU infrastructure

Trained with the open AETHER Mythos / Fableveil pipeline.

enosislabs/AETHER-Mythos-1-1.2B-gguf

AETHER-Mythos-1 — AETHER-Mythos-1-1.2B