Cyberdonia-24B-v8

"Abliterated Dolphin" is a result of my 3AM brain reading about technique called abliteration and then thinking what would happen if I tried to abliterate a model that is already relatively free, such as Dolphin. Heavily inspired by mlabonne's article on abliteration on how to redirect refusals and effectively remove, or ablate, censorship from a language model. There is really no deeper meaning to any of this than pure curiosity. This is basically a dumbed down version of the original Dolphin model I used as a base, as I have not done any DPOs to heal the damage caused by abliteration. Don't try to do anything meaningful with this model. Use the original Dolphin3.0-R1-Mistral-24B instead.…

Model · Text generation

LFM2.5-24B-A2B-OptiQ-4bit

MLX Community

An OptiQ mixed-precision quant of LiquidAI/LFM2.5-24B-A2B, the largest model in the LFM2.5 family. It is 16.7 GB on disk, down from 47.9 GB at bf16, with a 128k context. LFM2.5-24B-A2B is a sparse mixture of experts on Liquid's hybrid backbone. Convolutional blocks alternate with full attention, so only 10 of its 40 layers keep a KV cache. Each token is routed to 4 of 64 experts. OptiQ measures each layer's sensitivity and assigns its bit-width from that. Most of the routed expert weights sit at 4-bit: 87 of the 114 fused expert projections. Attention, the routers and nearly all of the convolution blocks stay at 8-bit. Six-metric mean, the standard OptiQ eval, run in reasoning mode because…

Model · Text generation

Qwen3.8-Flash-Next-Abliterated-EXL3-2.50bpw

Ghost

A refusal-removed (abliterated) build of Qwen's Qwen3.8-Flash-Next, quantized to EXL3 2.50 bpw so the full model runs on a single 24 GB card (RTX 3090 / 4090) using MoE CPU-offload — with the vision tower, MTP head, native 262,144-token context, and the PLE n-gram table all intact. Requires the same MoE CPU-offload setup as the stock 2.50bpw pack. Needs ~59 GB host RAM for the CPU expert tail and a fast NVMe for the streamed n-gram table. Expected on an RTX 3090: ~38 tok/s decode with MTP on (~28 without), ~20 tok/s at 175K depth, ~664 tok/s prefill. See the upstream repo for the full measured ledger; this quant uses the identical flags and layout, so numbers should track closely. Fired on…

This mistral model was trained 2x faster with Unsloth and Huggingface's TRL library.

An MLX (Apple silicon) build of the abliterated Gemma-4-26B-A4B-it MoE (128 experts, top-8, ~4B active). It uses Google's QAT int4 lattice weights wherever they are exact. The build targets a Mac mini M4 with 24 GB serving 2 concurrent requests at 64k context. It runs through oMLX fully resident at a 21.5 GB Metal limit. At a 20 GB limit, MoE expert offload streams non-resident experts from the NVMe SSD on demand. - Quality vs the exact bf16 model: KL 0.053, top-1 agreement 92.4%. This is the same as the measured noise floor between two exact bf16 implementations (≈0.05 / ≈91%). - Refusal behaviour is identical to the bf16 abliterated model (XSTest safe prompts: 0.8%). The source is…

Model · Text generation

WaifuGemma4-26b-a4b-v1

HiWaifu Research

Gemma 4 26B-A4B, post-trained with GRPO against a reward model learned from 1.2 million double-blind votes cast by HiWaifu users inside their own role-play conversations. Put back into the same arena, blind, it met GLM-5.1 in 1,430 battles and won 49.6% of the decided votes; against a 13-model field including Gemini, DeepSeek-v4 and Qwen's character models it won 54.7%. Most open role-play models are tuned on preferences that come from an LLM judge, from a handful of annotators, or from synthetic pairs. We had something rarer: a live arena where, inside ordinary chats on our platform, a user is occasionally shown two candidate replies and asked which one they want to continue with. Those…