LFM2.5-8B-A1B — Uncensored by Gastón Parravicini
First publicly available uncensored/abliterated GGUF of LFM2.5-8B-A1B.
Base model released May 28, 2026. This release: May 29, 2026.
TL;DR
Liquid AI dropped LFM2.5-8B-A1B yesterday. It refused everything. Today it doesn't.
Refusal rate: 1/100 on AdvBench. Reasoning intact. iMatrix quants. Same-day release.
About LFM2.5-8B-A1B
LFM2.5-8B-A1B is Liquid AI's latest edge model — a hybrid convolution + attention MoE architecture with:
- 8.3B total parameters, 1.5B active per token (MoE with 32 experts, 4 active)
- 128K context window (up from 32K in LFM2)
- Trained on 38T tokens with large-scale reinforcement learning
- Reasoning model — generates
<think>...</think>chains before answering - Fastest in its class: 18,500 tokens/sec on H100 at high concurrency
The architecture is not a standard Transformer. It combines:
- 18 layers of gated short convolutions (LIV blocks) — O(n) complexity
- 6 layers of Grouped Query Attention (GQA) — O(n²) for global context
- MoE feed-forward with sparse expert routing
This hybrid design is what makes it fast. It's also what makes abliteration non-trivial.
Why standard abliteration tools don't work here
Every existing abliteration tool — NousResearch, Heretic, OBLITERATUS — targets standard Transformer weight matrices:
self_attn.o_proj ← doesn't exist in LFM2.5
mlp.down_proj ← doesn't exist in LFM2.5
Running sharded_ablate.py on LFM2.5 without patching results in 0 shards modified. The model is completely unchanged. This is why no abliterated version existed before this release.
How this was done
1. Architecture reverse engineering
Full manual inspection of the LFM2.5 weight map to identify the correct abliteration targets:
Layer type | Target matrix
--------------------|----------------------------------
Conv LIV block | conv.out_proj [2048, 2048]
GQA Attention block | self_attn.out_proj [2048, 2048]
Dense FFN (L0-L1) | feed_forward.w2 [2048, 7168]
Key insight: conv.in_proj has shape [6144, 2048] — a 3x expansion projection that cannot be abliterated with the standard direction subtraction without a dimension mismatch error. Excluded intentionally.
2. Patch to sharded_ablate.py
# LFM2/LFM2.5 hybrid architecture support patch
# by Gastón Parravicini — May 29, 2026
# Enables abliteration of lfm2moe models in NousResearch/llm-abliteration
lfm2_patterns = [
f"{layer_prefix}.layers.{layer}.self_attn.out_proj.weight",
f"{layer_prefix}.layers.{layer}.conv.out_proj.weight",
f"{layer_prefix}.layers.{layer}.feed_forward.w2.weight",
]
Without this patch: 0/10 shards modified.
With this patch: 6/10 shards modified, all correct targets.
3. Refusal direction analysis
Used analyze.py to map refusal signal strength across all 24 layers:
| Layers | Est. Signal Quality | Type |
|---|---|---|
| 0–2 | ~0.000 | Skip |
| 3–10 | 0.010–0.062 | Low |
| 11–17 | 0.108–0.240 | Peak — abliterated here |
| 18–23 | 0.049–0.145 | High |
Layer 16 was the peak signal layer (Est. Signal Quality: 0.242). Used as the primary measurement reference for all ablated layers.
4. Abliteration parameters
layers: 11–23
measurement: layer 16 (peak refusal signal)
scale: 2.0
flags: --projected --normpreserve
--projected: orthogonalizes the refusal direction against the harmless direction before subtracting — cleaner removal, less capability damage--normpreserve: preserves weight matrix row norms after projection — prevents magnitude drift
5. Weight diff verification
Post-abliteration comparison against base model (via compare.py):
| Metric | Value |
|---|---|
| Avg weight diff | ~4–5 × 10⁻⁴ |
| Max weight diff | ~1–3% |
| Layers 0–10 | Zero diff — untouched ✅ |
| Layers 11–23 | Surgical modifications only ✅ |
Modifications are minimal and targeted. The model's general capabilities are preserved.
Abliteration results
| Metric | Result |
|---|---|
| Refusal rate (AdvBench 100 prompts) | 1/100 (1%) |
Reasoning (<think> tags) | ✅ Fully intact |
| General capability | ✅ Verified |
| Same-day release | ✅ May 29, 2026 |
The single remaining refusal out of 100 is an edge case. The model reasons freely — the <think> block no longer contains refusal logic.
Available quants
All generated with iMatrix calibration on harmful/harmless instruction data.
| File | Size | Use case |
|---|---|---|
...IQ4_XS.gguf | ~4.4 GB | Maximum compression |
...Q4_K_M.gguf | ~4.9 GB | ⭐ Recommended — best balance |
...Q5_K_M.gguf | ~5.7 GB | Better quality |
...Q6_K.gguf | ~7.2 GB | High quality |
...Q8_0.gguf | ~8.6 GB | Near-lossless |
...F16.gguf | ~16 GB | Full precision |
Usage
llama-server
llama-server \
-m LFM2.5-8B-A1B-Uncensored-Gaston-Q4_K_M.gguf \
-ngl 99 -c 8192 --port 8080
Ollama
ollama run hf.co/gaston-parravicini/LFM2.5-8B-A1B-Uncensored-Gaston-GGUF:Q4_K_M
llama-cli (quick test)
llama-cli \
-m LFM2.5-8B-A1B-Uncensored-Gaston-Q4_K_M.gguf \
-ngl 99 -p "<|startoftext|><|im_start|>user\nHello!<|im_end|>\n<|im_start|>assistant\n"
Notes
Tool calling: LFM2.5 supports tool calling natively in transformers. In llama.cpp there is a known bug with the chat template that breaks tool use — upstream is debugging (PR #23826).
Prompt cache: lfm2moe models clear the KV cache on every turn in llama.cpp (known upstream issue). Output quality is unaffected.
Reasoning: This is a thinking model. Responses include <think>...</think> before the final answer. This is expected and correct.
Base model
- Model: LiquidAI/LFM2.5-8B-A1B
- Architecture: lfm2moe — hybrid conv (LIV) + GQA + MoE
- Parameters: 8.3B total / 1.5B active
- Context: 128K tokens
- License: LFM Open License v1.0
Released by Gastón Parravicini — huggingface.co/gaston-parravicini
Architecture patch for lfm2moe abliteration — first of its kind.