Skip to content

Search wahloon.net

Search models, news, guides and FAQs.

All news
News#catalog#deepseek#glm#gemma#qwen#roundup

Fourteen new abliterated builds join the catalog

DeepSeek V4.1, GLM 5.3, MiniCPM-V, a 122B Qwen and more: the newest abliterated and uncensored entries, grouped by what they need to run.

5 min readMarkdown
On this page

The catalog grew from 18 to 32 models this week. The new entries came from searching Hugging Face for abliterated and uncensored, then keeping the builds that have a real quantization path and enough model-card detail to judge. Every entry links to its source repository and records its method, formats, approximate Q4 memory and license.

The additions are not one story. There are two new DeepSeek V4.1 Flash builds, two GLM 5.3 Flash builds, a 122B Qwen, several 35B-A3B mixture-of-experts models, and a handful of small multimodal options. Here is how they break down.

The headline additions

DeepSeek V4.1 Flash and GLM 5.3 Flash are the largest bases in the batch. Both ship under permissive licenses, and both have an abliterated version plus a separately tuned uncensored version.

BuildMethodApprox. VRAM at Q4License
Huihui-DeepSeek-V4.1-Flash-1017-abliteratedAbliterationVery highMIT
DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCOFine-tuneVery highMIT
Huihui-GLM-5.3-Flash-abliteratedAbliterationHighMIT
GLM-5.3-Flash-UncensoredFine-tuneHighMIT

“Very high” is deliberate: these are large mixture-of-experts models, and the catalog does not pretend they fit on a single consumer card. They are worth tracking if you have multi-GPU hardware or a high-memory workstation. GLM 5.3 Flash is the more attainable of the two, but check the model card for the exact parameter count and quant list before planning a build.

The batch also includes Huihui-Qwen3.8-Flash-Next-abliterated-v2. It has passed a million downloads in the last 30 days and is one of the most-used entries in the catalog. It is a large MoE with a custom license, so read the card before commercial use.

The 35B-A3B cluster

Qwen3.6-35B-A3B is becoming the default “big model that runs like a small one” base. The catalog now tracks three builds:

  • Huihui-Qwen3.6-35B-A3B-abliterated-MTP — the popular baseline, about 21 GB at Q4.
  • Qwen3.6-35B-A3B-uncensored-heretic-MTP — a Heretic edit that preserves native MTP.
  • MiniCPM-V-4.7-35B-A3B-Uncensored — the same size class with vision support.

All three still need room for every expert. The active parameter count is small, which helps token speed, but it does not shrink the weights. Budget roughly 21–22 GB at Q4, plus context.

MiniCPM-V is the interesting one for multimodal work: it is an uncensored refusal-direction edit with image input, not just a text model with a vision projector bolted on. If you already run a 24 GB card, it is a reasonable alternative to a separate vision model.

The practical mid-range

Three additions target the 14–18 GB window:

BuildMethodApprox. VRAM at Q4Notes
LFM2.5-24B-A2B-UncensoredAbliteration~14–16 GBMoE; GGUF for llama.cpp
Underdog-Saluki-27B-1.0-MTP-AbliteratedAbliteration~16–18 GBTool use; non-commercial license
Qwen3.8-27B-Defiance-Uncensored-MTPFine-tune~16–18 GBReasoning and roleplay

LFM2.5-24B-A2B is the smallest full-size option in the group and the easiest to fit on a 16 GB card. Saluki is tuned for agent workflows and tool calls, with MTP support for speculative decoding, but its base is CC-BY-NC-4.0 — fine for personal and research use, not for a commercial product. Defiance is a Qwen3.8 fine-tune aimed at reasoning and roleplay, with the usual caveat that aggressive tuning can drift from the base model.

Small and multimodal

The last group is for smaller cards and vision experiments:

  • Gemma-4-26B-A4B-Uncensored-HauhauCS-Balanced-MTP — QAT, vision, about 16–18 GB at Q4.
  • Gemma-4-E4B-Uncensored-HauhauCS-Aggressive — vision and audio, about 3–4 GB at Q4.
  • Huihui-Qwen2.5-VL-7B-Instruct-abliterated — Transformers format, about 15 GB at BF16.

The E4B build is the one to try on a 6 GB card or a laptop. The Qwen2.5-VL entry is older and Transformers-only, so it is less convenient than a GGUF, but it is a useful comparison point for local vision.

What to run first

Your hardwareStart with
6–8 GBGemma 4 E4B at Q4
12 GBLFM2.5 24B-A2B at Q3/IQ3, or a 12B Gemma build
16 GBLFM2.5 24B-A2B at Q4, or a 27B at Q3
24 GBQwen3.8-27B or Defiance at Q4, or MiniCPM-V at Q4
48 GB+35B-A3B at Q5/Q6, or a 70B at Q4
Multi-GPU / high memoryDeepSeek V4.1 Flash, GLM 5.3 Flash, or the 122B Qwen

These are starting points, not guarantees. Context length, KV cache and your backend all move the number.

Licenses and caveats

  • MIT: the DeepSeek V4.1 Flash and GLM 5.3 Flash entries.
  • Apache-2.0: most Qwen, MiniCPM-V and LFM builds in this batch.
  • Gemma: the Gemma 4 entries. Read the Gemma terms before commercial use.
  • CC-BY-NC-4.0: Saluki. Non-commercial only.
  • Other: Flash-Next v2 and LFM2.5 have custom terms; check the model card.

Download counts are approximate 30-day figures and change constantly. They are a popularity signal, not a quality score. The catalog stores the last-updated date for each entry so you can spot a stale build.

Next steps

3 minNews

MoE abliteration arrives for Qwen3.6-35B-A3B

The 35B-A3B MoE base is now the most interesting size/quality trade-off for local inference. Here are the abliterated and uncensored builds worth testing.

#qwen#moe#35b-a3b
Read