"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.…
Open-weight model · Text generation
LFM2.5-24B-A2B-OptiQ-4bit
by MLX Community mlx-community/LFM2.5-24B-A2B-OptiQ-4bit
LFM2.5-24B-A2B-OptiQ-4bit is an open-weight model for text generation from MLX Community, released under other. It has 24B parameters and a 131,072-token context. At 16-bit it needs about 57.5 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index.
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.
Parameters24B
Context131,072
Weights16.7 GB
Licenseother
AccessOpen weights
Monthly Downloads—
Runs On
What it takes to serve LFM2.5-24B-A2B-OptiQ-4bit (24B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 47.9 GB | 57.5 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 24.0 GB | 28.8 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 12.0 GB | 14.4 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 18, 2026.
LFM2.5-24B-A2B-OptiQ-4bit on every accelerator the SAVRN Index prices, at every precision
Model Card
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…
Excerpt from the card by MLX Community, licensed other.
Configuration
- Architecture
- Lfm2MoeForCausalLM
- Context length (tokens)
- 131,072
- Layers
- 40
- Hidden size
- 2,048
- Feed-forward size
- 11,776
- Attention heads
- 32
- Key/value heads
- 8
- Vocabulary size
- 128,000
- Experts
- 64
- Experts active per token
- 4
- Model type
- lfm2_moe
Identity and Version
- Repository
- mlx-community/LFM2.5-24B-A2B-OptiQ-4bit
- Publisher
- MLX Community
- Task
- Text generation
- Modality
- Text
- Library
- mlx
- Parameters
- 24B parameters
- Languages
- mlx, moe, on-device
- Revision
- b5552f7f7e597e2ca02653f89098dea86b8818af
- First published
- 2026-10-16
- Last updated
- 2026-10-16
Files and Weights
15 files, 16.7 GB in total. The weights are 4 files totalling 16.7 GB in safetensors.
Weights4 files · 16.7 GB
Configuration6 files · 227.7 KB
Tokenizer2 files · 17.9 MB
Documentation1 file · 3.9 KB
Other1 file · 5.4 KB
Repository1 file · 1.6 KB
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model-00001-of-00004.safetensors | Weights | 5.3 GB | 4d7098a0724e |
| model-00002-of-00004.safetensors | Weights | 5.4 GB | 0618aa025a4d |
| model-00003-of-00004.safetensors | Weights | 5.3 GB | 8255962bf517 |
| model-00004-of-00004.safetensors | Weights | 697.5 MB | a1677ef46d75 |
| config.json | Configuration | 61.7 KB | — |
| generation_config.json | Configuration | 329 B | — |
| kv_config.json | Configuration | 660 B | — |
| model.safetensors.index.json | Configuration | 86.9 KB | — |
| optiq/metadata.json | Configuration | 24.3 KB | — |
| optiq/sensitivity.json | Configuration | 53.8 KB | — |
| README.md | Documentation | 3.9 KB | — |
| chat_template.jinja | Other | 5.4 KB | — |
| .gitattributes | Repository | 1.6 KB | — |
| tokenizer.json | Tokenizer | 17.9 MB | 695be7802a0e |
| tokenizer_config.json | Tokenizer | 363 B | — |
License and Download
- License
- other
- Access
- Open weights, no gate
- Download size
- 16.7 GB
Download from MLX Community
Released by MLX Community through its official repository on Hugging Face.
Built From
- Derived from LiquidAI/LFM2.5-24B-A2B
- Quantized from LiquidAI/LFM2.5-24B-A2B
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 16.7 GB |
| 16-bit | 47.9 GB |
| 8-bit | 24.0 GB |
| 4-bit | 12.0 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About LFM2.5-24B-A2B-OptiQ-4bit
How much GPU memory does LFM2.5-24B-A2B-OptiQ-4bit need?
About 57.5 GB at 16-bit and 14.4 GB at 4-bit: the weights (24B parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run LFM2.5-24B-A2B-OptiQ-4bit on?
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
What license is LFM2.5-24B-A2B-OptiQ-4bit released under?
other, as its publisher declares it. Read the license text before commercial use.
What is LFM2.5-24B-A2B-OptiQ-4bit's context length?
131,072 tokens, from the maximum position embeddings in its published configuration.
Similar Models
/ Cockpit & Void Palette / / Hyperspace & Instrument Lights / / Low-framerate flicker keyframes (stepped timing) / / Deep radial hyperspace tunnel backdrop through cockpit window glass / radial-gradient(circle at 50% 45%, rgba(255, 230, 150, 0.25) 0%, rgba(255, 42, 75, 0.25) 18%, rgba(189, 0, 255, 0.2) 36%, rgba(0, 229, 255, 0.15) 54%, transparent 75%), repeating-conic-gradient(from 0deg at 50% 45%, #05070a 0deg 10deg, #0e131d 15deg, #05070a 20deg), radial-gradient(ellipse at center, transparent 30%, #000000 95%); / Low-framerate scanline overlay / / Heavy Flight Deck Chassis Container / 0 0 50px rgba(0, 0, 0, 0.9), inset 0 0 35px rgba(0, 0, 0, 0.8), inset 0 1px 0 rgba(255, 255, 255, 0.1)…
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…
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.
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…