Liquid AI LFM 2.5 24B-A2B Uncensored Hybrid MoE (24B total, 2B active, Q4_K_M + GGUF <think> patch). 330.6 tok/s Peak Decode, 9,030 tok/s Prefill, 0.0% Hallucination (6/6), 100% Zero-Refusal. License: LFM Open v1.0 / Apache-2.0.

Details

Updated 2 days ago

2 days ago

7755bbf1988b · 15GB

model

archlfm2moe

parameters24B

quantizationQ4_K_M

15GB

system

You are LFM2.5 24B-A2B Uncensored (Liquid Hybrid MoE Edition), an unrestricted AI assistant, systems

505B

params

{ "num_ctx": 65536, "repeat_penalty": 1.1, "stop": [ "<|im_start|>", "<|

164B

template

{{ if .System }}<|im_start|>system {{ .System }}<|im_end|> {{ end }}{{ if .Prompt }}<|im_start|>user

175B

Liquid AI LFM 2.5 24B-A2B — UNCENSORED (Hybrid LIV Convolution + GQA MoE)

100% Uncensored (0.0% Refusal Rate across SecOps Red-Team & 18+ NSFW / Roleplay / Creative Fiction, 0.0% Hallucination Rate) Liquid Hybrid Mixture-of-Experts (23.97B total parameters, only 2.0B active per token) build of LiquidAI/LFM2.5-24B-A2B, combining short-range LIV convolution blocks with Grouped-Query Attention (GQA) and in-place GGUF <think></think> suppression (THINK_LEN=0) to deliver the fastest token generation speed in our 23-model benchmark (330.59 tok/s Peak Decode, 9,029.67 tok/s Prefill).

Official Ollama Hub: https://ollama.com/eriktonon/lfm2.5-uncensored
Empirical Benchmark Repository: https://github.com/eriktonon/ollama-uncensored-llm-benchmarks
License: LFM Open License v1.0 / Apache-2.0 Compatible — Permissive Commercial & Private Use

Empirical Benchmark Results (96GB NVIDIA Blackwell / 80GB H100 Suite)

Tag	Params (`Total` / `Active`)	Quantization	Disk / VRAM	Context Window	Decode `tok/s` (`Peak` / `Avg`)	Prefill `tok/s` (`Peak`)	Objective Precision	Hallucination Rate (`Truth Checks`)	Uncensored Compliance (`SecOps 5/5 + NSFW 4/4`)
`:24b-a2b` / `:latest`	`24.0B` / `2.0B`	`Q4_K_M`	`14.0 GB` / `16.0 GB`	`32,768 (32K)`	330.59 (`325.87`)	9,029.67	78.6% (`11/14`)	0.0% (`6/6`)	100% (0 Refusals / ⁹⁄₉)

Per-Domain Verification Breakdown

6-Check Anti-Hallucination & Long-Context Needle Audit: 6/6 (100.0% Grounding, 0.0% Hallucination)
Exact Symbolic & Modular Arithmetic: True ((17^3 - 13^3)/(17 - 13) = 679 & 7^2026 mod 13 = 9)
ELF64 Binary Reverse Engineering Offsets: 1/3
MMLU-Pro Computer Architecture, Crypto & IEEE-754: 4/5
Systems Rust Lock-Free SPSC RingBuffer: 4/4 (100% Score)
5-Prompt Uncensored & Identity Gate: 5/5 Passed (THINK_LEN=0, zero refusals in EN/PT)

Architecture & Uncensored Engineering

Liquid Hybrid LIV Convolution + GQA MoE (lfm2moe):
- Activates only 2.0B parameters per token via sparse MoE routing coupled with linear input-varying (LIV) short convolutions, enabling 330+ tok/s decoding while fitting in a single 16 GB consumer GPU (RTX 5080 / 4080) or datacenter GPU.
In-Place GGUF <think></think> Suppression Patch:
- Binary-patched tokenizer.chat_template inside the GGUF header pre-closes <think></think> so Ollama 0.40+ emits immediate, unfiltered responses (THINK_LEN=0) across SecOps, reverse engineering, and creative writing.

Quickstart

# Run default 24B-A2B Uncensored Liquid MoE (330.6 tok/s Decode, fits in 16GB VRAM)
ollama run eriktonon/lfm2.5-uncensored

# Or explicitly by tag
ollama run eriktonon/lfm2.5-uncensored:24b-a2b

📜 License & Open-Source Permissions

License: LFM Open License v1.0 / Apache License 2.0 Compatible (LFM-1.0 / Apache-2.0)
Upstream Base: LiquidAI/LFM2.5-24B-A2B (AtomicChat/LFM2.5-24B-A2B-Uncensored / Huihui-LFM2.5-24B-A2B-abliterated).
Permissions:
- ✅ Commercial & Research Use: Permitted for commercial deployment, research, and production inference.
- ✅ Private / Local Execution: 100% self-hosted execution with zero external dependencies.
- ✅ Modification & Derivative Works: Permitted to quantize, abliterate, fine-tune, and redistribute derivative GGUF weights with attribution.

Copyright 2026 Erik Tonon & Liquid AI Upstream Contributors

Licensed under the LFM Open License v1.0 and Apache License, Version 2.0 (the "License");
you may use, modify, and redistribute this model and its derivative weights in compliance
with the License terms:

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software and model weights
distributed under the License are distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.