← back to catalog · registered 2026-09-13 06:56

Navin-Models/Qwen3.8-Flash-Next-Uncensored-AD-3.87bpw-MLX

Navin-Models MoE multimodal second-order
Abliteration classifier · v1.0.0
M1
Primary method

Direct removal

No other method signals detected in this model.
Confidence
MEDIUM
Why this label 3 signals
Method inferred from partial signals — repository name, related files, or tag patterns. Producer identity not confirmed; label may sharpen or shift as we gather more evidence.
  • 'abliterated' in name/tags
  • is_gguf=0 (base model)
  • no specific method indicators - defaulting to M1 (most common)
Refusal direction extracted via
Extraction technique

Difference-of-means

Confidence
MEDIUM
Why we say so
primary_method=M1; difference-of-means is the reference extraction for M1/M3 (Arditi 2024)
Downloads · 30-day
0
Likes
0
Model age
today
created 2026-09-13

Genealogy 0 direct forks

Full fork graph →

This model's place in the market. Above: what it was derived from. Below: the tree of everything derived from it.

Metadata

License
other
Languages
en zh
Tags
mlx qwen4_exp qwen qwen3.8 qwen4-exp uncensored abliterated mixture-of-experts vision-language function-calling reasoning mtp

Related

Total size
0 B
Files
15
Quantizations
1
Registered
2026-09-13 06:56
Last updated on HF
2026-09-13 05:59

Files by quantization

Auxiliary files 15 files 24.0 MB
tokenizer.json 12.2 MB ******** download
vocab.json 6.41 MB 0aa0ce06 download
merges.txt 3.20 MB a494e019 download
artifact-manifest.json 1.84 MB d38be0c2 download
SHA256SUMS 230 KB ebece742 download
tokenizer_config.json 17.5 KB 5de744b3 download
.gitattributes 16.3 KB d974ddda download
chat_template.jinja 8.74 KB c0c686f9 download
README.md 8.46 KB b9bf5593 download
config.json 4.63 KB 491017e9 download
LICENSE 3.16 KB 9557a896 download
RELEASE_COMPLETE.json 581 B e502db0f download
preprocessor_config.json 390 B 2ea84a43 download
video_preprocessor_config.json 385 B 3ba673a5 download
generation_config.json 202 B 023756cf download

README current version from Hugging Face


license: other
license_name: qwen-community-1.0
license_link: LICENSE
base_model:

  • orcarouter/Qwen3.8-Flash-Next-Uncensored
    base_model_relation: quantized
    pipeline_tag: image-text-to-text
    library_name: mlx
    quantized_by: Navin-Models
    language:
  • en
  • zh
    tags:
  • qwen
  • qwen3.8
  • qwen4-exp
  • mlx
  • uncensored
  • abliterated
  • mixture-of-experts
  • vision-language
  • function-calling
  • reasoning
  • mtp
  • imatrix
  • research
    extra_gated_prompt: >-
    This is an abliterated research model with substantially reduced safety alignment.
    It may produce harmful, illegal, offensive, biased, or otherwise unsafe content.
    Access is provided for legitimate research and controlled evaluation. You are responsible
    for lawful use, downstream safeguards, and compliance with the Qwen Community License 1.0.
    extra_gated_fields:
    I have read and accept the included Qwen Community License: checkbox
    I understand that safety alignment has been substantially removed: checkbox
    I will use this release lawfully and take responsibility for appropriate safeguards: checkbox

Qwen3.8-Flash-Next-Uncensored AD-3.87bpw MLX

[!CAUTION]
Research artifact with substantially reduced safety alignment. This model is derived from an abliterated checkpoint and may comply with harmful, unethical, illegal, offensive, biased, or otherwise unsafe requests that the aligned model would refuse. It has no dependable built-in guardrails. Do not expose it to end users or production traffic without independently designed safety, moderation, access-control, logging, and abuse-prevention measures. You are responsible for how you use it and for compliance with applicable law.

A tensor-specific mixed MLX artifact for 64GB Macs, derived from orcarouter/Qwen3.8-Flash-Next-Uncensored at revision 8336e613ea508b13c2159bd0f68965d97a606b95.

The build preserves the 262,144-token model configuration, separate F16 vision tensors, a separate MTP section, and an external mmap-backed Q5_1 PLE table. Its allocation adapts AtomicChat's published Atomic Dynamic tensor-allocation philosophy to a custom MLX representation. The 3.87bpw name is this release's measured encoded model-payload average excluding the separately loaded F16 vision section; it is not an AtomicChat release identity.

This is an independent community build. It is not produced, endorsed, or warranted by Qwen, Alibaba, OrcaRouter, AtomicChat, MLX, or mlx-lm.

Important runtime status

This repository is a verified packed-weight artifact, not a drop-in mlx-lm model. The included runtime/admlx code provides the manifest-bound loader, canonical decoders, mmap PLE access, and direct MLX/Metal packed-linear kernel seam used during validation. It does not currently provide a turnkey end-to-end generation CLI for the full model.

Static memory accounting passes all four 262K scenarios—text, text with MTP, vision, and vision with MTP—for 64GB Macs. The heaviest projected Metal working set is 56,880,334,312 bytes, with a projected 10,765,400,600 bytes of physical memory remaining after the stated buffer and touched-PLE assumptions. A real full-model 262K runtime trial has not yet been performed, so the release verdict intentionally remains:

STATIC_PROVISIONAL_RUNTIME_PENDING

Do not interpret the static fit calculation as measured generation performance or a guarantee against memory pressure.

Contents

Section Tensors Payload bytes Format / role
Resident body 1,166 47,578,244,504 Mixed IQ2_XS, IQ2_S, IQ3_S, Q8_0 and exact structural tensors
External PLE 128 38,400,184,320 Canonical Q5_1 shards, mmap-backed
MTP 31 902,321,776 Separate draft section; target importance matrix not applied
Vision 333 897,862,112 Separate F16 section
Total 1,658 87,778,612,712 Complete encoded payload

Measured effective average:

  • model payload excluding separate F16 vision: 3.8710216579 bpw → release name 3.87bpw
  • complete package including separate F16 vision: 3.9012720779 bpw

Tensor allocation

  • routed gate/up: IQ2_XS, IQ2_S, or IQ3_S selected per layer by measured importance-weighted reconstruction error
  • routed down: IQ2_S or IQ3_S selected per layer
  • sensitive tensors: Q8_0
  • PLE n-gram table: external canonical Q5_1 shards
  • vision: separate F16
  • MTP: separate mixed section; no target-only importance-matrix reuse
  • structural integer tensors: exact I64 preservation

Four consecutive 640-wide routed-down logical rows are packed as one 2,560-element canonical encoder row. Consumers must honor logical_rows_per_canonical_row in artifact-manifest.json.

Verification

The terminal verifier passed with no reported problems:

  • exactly 1,658 unique tensors
  • missing tensors: 0
  • extra tensors: 0
  • section overlap: 0
  • exhaustive distributed-file SHA-256 inventory
  • decode checks for IQ2_XS, IQ2_S, IQ3_S, Q8_0, Q5_1, F16, and I64
  • exact source equality for all three structural I64 tensors
  • static 262K budget checks for all four text/vision and MTP combinations

See:

  • artifact-manifest.json — tensor names, sections, shapes, paths, formats, and payload checksums
  • SHA256SUMS — release-file SHA-256 inventory
  • evidence/CONVERSION_VERIFICATION.json — terminal conversion verifier result
  • evidence/FINAL-V7-TERMINAL-INVENTORY.json — archived final inventory summary
  • RELEASE_COMPLETE.json — release-tree identity and counts

Verify after download:

shasum -a 256 -c SHA256SUMS

Runtime integration

The files under runtime/admlx are the exact runtime primitives distributed with this release. Minimal manifest loading:

from pathlib import Path
from runtime.admlx.runtime_loader import ArtifactLoader

root = Path("Qwen3.8-Flash-Next-Uncensored-AD-3.87bpw-MLX")
artifact = ArtifactLoader(root / "artifact-manifest.json")
linear = artifact.load_qlinear("lm_head.weight")

This demonstrates manifest-bound tensor loading; it is not a full generation example. A full consumer must assemble the complete Qwen4Exp architecture, attach the vision and MTP paths, implement the cache policy, and preserve mmap behavior for the external PLE table.

Suggested Python dependencies for the included primitives:

mlx==0.32.2
numpy

Provenance and method

  • source checkpoint: orcarouter/Qwen3.8-Flash-Next-Uncensored
  • pinned source revision: 8336e613ea508b13c2159bd0f68965d97a606b95
  • source inventory: 131 weight shards, 360,000,192,888 bytes, 1,658 tensors
  • llama.cpp revision used for the importance-matrix and canonical-format toolchain: d1a92352cbd417fd840b4e765c0b82f5fe3d1d89
  • MLX version: 0.32.2
  • mlx-lm revision used during development: c62d7957064058b3f00cac9a3e1fe00cbc23ab68
  • importance-matrix SHA-256: 5591ce3dc3bf0b73d3c074bc588c90b6c4f7b3c273de6b10e50d111b75f05487

The conversion was resumable and collision-safe across 18 workers. Worker-local transaction journals, independent atomic tensor writes, SHA-256-verified merge, existing-output reuse, and corruption recovery were tested before production conversion. Of 1,658 tensors, 1,192 were generated by the parallel workers and 466 verified tensors were reused from earlier valid progress.

Attribution

The AD label acknowledges adaptation of AtomicChat's Atomic Dynamic tensor-allocation philosophy and its published 64GB Mac memory guidance. This repository uses a different custom MLX representation and its own measured 3.87bpw allocation; it is not AtomicChat's AD-3.84 artifact.

License

The model and its derivatives are distributed under the included Qwen Community License 1.0. Review LICENSE before use. The source checkpoint's model card previously displayed Apache-2.0 metadata while its distributed license file contained the Qwen Community License; this release follows the actual included license text and uses license: other metadata.

This summary is not legal advice. Users are responsible for determining whether their intended use complies with the license and applicable law.

Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in app" button that hands off directly to a local runtime of your choice - Infrahuman, LM Studio, or Ollama. No API keys, no subscription, no prompt leakage.