LFM2.5-24B-A2B-uncensored
24BOtheruncensored
from LiquidAI/LFM2.5-24B-A2B
↓ 444.9K♥ 7311 repos
LFM2.5-24B-A2B-uncensored2
LFM2.5-24B-A2B-uncensored2 is a Heretic-ablated variant of LiquidAI/LFM2.5-24B-A2B, published on Hugging Face by electroglyph. It is 24B total parameters (mixture-of-experts), 24GB VRAM class class and other licence. It is distributed across 4 repositories in Transformers, GGUF and GGUF (imatrix) formats, totalling 5.5K downloads.
Specification
- Model name
- LFM2.5-24B-A2B-uncensored2
- Source repository
- electroglyph/LFM2.5-24B-A2B-uncensored2
- Base model
- LiquidAI/LFM2.5-24B-A2B
- Publisher
- electroglyph
- Parameters
- 24B (mixture-of-experts)
- VRAM required
- ~15.4 GB at Q4_K_M, 4k context (estimated) — runs on 16 GB cards
- Hardware class
- 10B – 25B — 24GB VRAM class
- Licence
- other
- Task
- Text generation
- Context window
- 32K tokens (as documented)
- Ablation scope
- Not stated by the publisher
- Formats available
- Transformers, GGUF, GGUF (imatrix)
- First indexed
- 22 Oct 2026
- Last updated
- 23 Oct 2026
Ablation technique
How LFM2.5-24B-A2B-uncensored2 was modified, and what that implies.
Heretic
Automated directional ablation via the Heretic toolchain, which searches for the refusal direction and applies it with a KL-divergence budget so general capability is preserved.
Reported scope: Not stated by the publisher. Publishers frequently omit this, so verify behaviour empirically rather than assuming full coverage.
From the publisher’s model card
“This model was selected from 50 Heretic history candidates (at most 8 refusals per 100 harmful prompts, ranked by ascending KL divergence) by reconstructing every candidate and evaluating it with the Language Model Evaluation Harness. The score uses the same procedure as Heretic's BenchmarkScore scorer.”
Running it
Commands are templates — confirm the exact repository and quant file before use.
llama.cpp / GGUF
hf download mradermacher/LFM2.5-24B-A2B-uncensored2-i1-GGUF --local-dir ./lfm2.5-24b-a2b-uncensored2
llama-cli -m ./lfm2.5-24b-a2b-uncensored2/<file>.gguf -p "..."Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "electroglyph/LFM2.5-24B-A2B-uncensored2"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
repo, torch_dtype="auto", device_map="auto"
)vLLM
vllm serve electroglyph/LFM2.5-24B-A2B-uncensored2 --trust-remote-codeDownloads and variants (3)
Every published repository of LFM2.5-24B-A2B-uncensored2, including quantised re-releases by other authors.
| Repository | Format | Downloads |
|---|---|---|
| mradermacher/LFM2.5-24B-A2B-uncensored2-i1-GGUF | GGUF (imatrix) | 4.3K |
| mradermacher/LFM2.5-24B-A2B-uncensored2-GGUF | GGUF | 1.2K |
| electroglyph/LFM2.5-24B-A2B-uncensored2source | Transformers | 79 |
Security-team assessment
Derived from this model’s metadata and the publisher’s claims — not from benchmarks run by this site.
- LFM2.5-24B-A2B-uncensored2 will attempt offensive-security prompts that a hosted commercial model declines, which is what makes it usable for red-team corpus generation and for measuring what an unaligned model of this class produces.
- Capability is inherited from LiquidAI/LFM2.5-24B-A2B, not added by ablation — at 24B parameters, expect roughly the base model’s competence, minus some instruction-following fidelity.
- The publisher does not state which layers were ablated, so assume nothing about refusal consistency — probe it directly.
- Declared licence is other, but the terms that bind you are LiquidAI/LFM2.5-24B-A2B’s — a re-release cannot grant more than its parent.
Handling: run it isolated, put moderation in front of it if anyone outside your team can reach it, and verify the weights before loading — these are third-party artifacts. Full handling and licence guidance.
Metadata for LFM2.5-24B-A2B-uncensored2 is collected from the public Hugging Face API and the publisher’s model card. This site does not host weights, is not affiliated with electroglyph, and does not independently verify publisher claims. See responsible use.