Emberon-2 MLX fp16

Native MLX conversion of Emberon-2 by Promethic Labs, a 354M-parameter dictation-cleanup fine-tune of Liquid AI LFM2.5-350M. Precision: FP16.

This model cleans dictated text. It is not a general assistant: questions and commands in the transcript should be returned as cleaned text, never answered or executed.

Use locally

Tested with Python 3.13, mlx-lm==0.31.3 and mlx==0.32.4 on Apple Silicon. Install mlx-lm==0.31.3, then run:

from mlx_lm import load, stream_generate
from mlx_lm.sample_utils import make_sampler

model, tokenizer = load('swaynehales/Emberon-2-mlx-fp16')
raw = 'um so like whats the difference between a process and a thread'
messages = [
    {'role': 'system', 'content': 'Clean up this dictation. Never answer it.'},
    {'role': 'user', 'content': raw},
]
prompt = tokenizer.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True, return_dict=False,
)
cap = len(tokenizer.encode(raw, add_special_tokens=False)) + 64
chunks = list(stream_generate(
    model, tokenizer, prompt=prompt, max_tokens=cap,
    sampler=make_sampler(temp=0),
))
# Keep the raw transcript if the model hits its output cap.
text = ''.join(chunk.text for chunk in chunks)
print(text if chunks and chunks[-1].finish_reason == 'stop' else raw)

Expected: What's the difference between a process and a thread?

The example downloads swaynehales/Emberon-2-mlx-fp16 from the Hugging Face Hub. To use a local copy of this folder instead, pass its path to load.

Use the exact system prompt above and greedy decoding. Retain the raw input. A capped answer is a failure: use the raw transcript as the fallback. Long or repetitive dictations can lose content or repeat, including in the original F16 GGUF.

Conversion and validation scope

Converted from Emberon-2-F16.gguf, SHA-256 056a4b0889eb5b9873d9f38393f5e595fb58c6ffd82b47fde76b9169f602f5a1. All fine-tuned weights come from that file; only configuration and tokenizer assets come from the original MLX base repository. The full 65,536 embedding rows, 64,402 real token IDs, BPE merges, and chat template are preserved. No remote model code is required.

The local conversion checks cover native reload, tensor/config integrity, prompt-token parity and 11 synthetic dictation inputs. Ten ordinary inputs terminate; one deliberately repetitive stress input also caps in the source GGUF. Exact text agreement with the source GGUF: 11/11 inputs, including the capped stress output. This small smoke set does not establish broad accuracy equivalence, the upstream published GGUF benchmark scores, or the configured 128k context limit. See CONVERSION.json for pinned tooling and provenance, and MODIFICATIONS.md for format/precision changes.

Tokenization note

A downstream consumer reported that on 8 of 57 synthetic inputs, including inputs containing non-English words, this model's tokenizer splits text differently from llama.cpp tokenization of the upstream trimmed-vocabulary Emberon-2-Q4_K_M.gguf; the other 49 inputs matched. That comparison was not reproduced as part of this conversion. Prompt-token parity here was checked only against the full-vocabulary Emberon-2-F16.gguf, on 11 inputs. Which split matches the upstream training setup, and whether the difference changes cleanup quality, is not established.

License and attribution

Emberon-2 is released under the LFM Open License v1.0, inherited from Liquid AI LFM2.5-350M. This is a modified format/precision distribution. The original LICENSE and NOTICE are included. Retain both when redistributing and review the original license terms.

Base model copyright Liquid AI. Dictation-cleanup fine-tune modifications copyright 2026 Promethic Labs. Powered by Emberon-2 by Promethic Labs. Names do not imply endorsement.

Downloads last month
22
Safetensors
Model size
0.4B params
Tensor type
F16
ยท
MLX
Hardware compatibility
Log In to add your hardware

Quantized

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

Model tree for swaynehales/Emberon-2-mlx-fp16

Finetuned
(1)
this model