MeldhLLC/Olune-Qwen3.5-4B-v17-beta

Olune Qwen3.5-4B v17 beta, Q4_K_M

This is a text-only Olune fine-tune of Qwen/Qwen3.5-4B, not a stock quantization of the upstream model. It is a limited-beta artifact for Olune's on-device journal and philosophical reflection experience. The model is experimental; its raw GGUF does not include Olune's application safeguards.

Artifact and integrity

Item	Value
File	Qwen3.5-4B.Q4_K_M.gguf
Size	2,783,446,496 bytes
SHA-256	934480d34aae5304a5b0c05b2a5762c430a81041877875d32a811e626b61f720
Format	GGUF v3, qwen35, 441 tensors
Quantization	Q4_K_M (general.file_type=15, Q4_K family)
Upstream revision	851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a
Training/export date	September 13, 2026

Compare the downloaded file's byte count and SHA-256 with SHA256SUMS and artifact-manifest.json. The GGUF advertises 262,144 tokens of context; Olune's reported Pixel tests used 4,096 tokens. The larger advertised context was not validated in Olune.

Origin and training

The base is Qwen's Apache-2.0 Qwen3.5-4B at the exact revision above. Olune trained a rank-16 BF16 LoRA adapter on its frozen v17 text corpus (4,713 training examples, with a separate 370-example evaluation set) for two epochs and 590 steps. The adapter was merged into 16-bit weights, converted to GGUF, and quantized by Unsloth to Q4_K_M. The resulting file contains Olune-specific training. It does not contain the corpus as a distributed dataset, and this publication does not distribute the adapter or merged checkpoint.

Project records describe a foundational synthetic set of philosophical/life-question conversations, later additions and edits for Olune, and 200 calibrated-uncertainty examples. The historical source and review records are retained privately. The frozen training file does not carry per-record source/permission fields. This summary is not a claim that every individual answer or factual premise is independently verified.

The original Qwen3.5 base supports modalities beyond text. This Olune training and GGUF evaluation were text-only. No vision or other multimodal behavior is claimed.

Intended use and limits

The intended limited-beta use is adult, on-device philosophical reflection through Olune. It is not a therapist, crisis service, medical tool, or source of guaranteed factual or philosophical accuracy. Other runtimes, prompts, chat templates, sampling settings, devices, languages, and longer contexts have not been accepted by Olune's tests. The documented beta comparison and natural-response checks were in English on a Pixel 10a; no broad language evaluation is claimed.

In a seven-prompt, 64-token headless Pixel 10a comparison of this fine-tune's Q2_K, Q3_K_M, Q4_K_M, and Q5_K_M GGUFs, Q4_K_M had an 81.212-second median wall time, 16.201 prefill tokens/s, and 1.477 decode tokens/s. Q2 looped repeatedly; Q3 and Q5 retained observed attribution or philosophy errors. A separate single cool-start, matched 64-token comparison found the 4B Q4 faster than the older Olune Qwen3-8B Q4 on that device. These short tests do not predict every natural response or prove a general performance advantage.

In a later ten-case natural-response check, seven completions exceeded two minutes (observed range 89–288.6 seconds). Repetition, attribution, accuracy, and response-mode defects were observed. In a three-case cold imminent-crisis check, all three displayed app-path responses contained unsafe personalized content. That check did not establish the raw native tokens or every parser transition. These are small, targeted tests, not comprehensive safety evaluation. The model must not be used for crisis intervention.

Olune's intended limited-beta design includes a default-on application-level pause for detector-positive crisis prompts, resources shown to the person, and guards on completed output. A person may explicitly disable the pause after a warning; that choice does not make the model safe. These safeguards are application behavior, not embedded in this GGUF. Downloading or running the raw file does not reproduce them. Beta warnings and terms do not establish model safety.

License and attribution

The pinned upstream Qwen snapshot identifies Apache License 2.0. Qwen/Qwen3.5-4B is by the Qwen team at Alibaba Cloud. Meldh's changes are the Olune v17 LoRA fine-tune, 16-bit merge, GGUF conversion, and Q4_K_M quantization; see the Meldh-authored NOTICE. No Qwen or Alibaba Cloud endorsement is implied. The pinned upstream snapshot has no separate NOTICE file.

This card describes the specific binary and evidence available as of September 23, 2026. Repository identity and immutable hosted revision are established only after publication.