LFM2.5-24B-A2B-Abliterated — GGUF

An abliterated build of LiquidAI/LFM2.5-24B-A2B, produced with Heretic, then converted to GGUF and quantized. The base is a 23.97B-parameter Mixture-of-Experts (lfm2moe, 64 experts, top-4, ~2.45B active) with a hybrid backbone: 40 blocks = 30 double-gated short-conv + 10 GQA attention.

Modification notice (LFM Open License §4(b)). These files are modified derivatives of LiquidAI/LFM2.5-24B-A2B. 926 weight tensors were altered by directional ablation to reduce refusal behavior — 896 per-expert down_proj (w2) and 30 attention/short-conv out_proj — then converted to GGUF and quantized. Liquid AI did not produce, endorse, or review this build, and its behavior is not theirs.

Read this before using it

LFM2.5 is a pure reasoning model — its chat template always opens a <think> block, with no non-thinking mode. Heretic scores refusals with the think block skipped (--response-prefix '</think>\n\n'), or its markers (illegal, harmful, …) fire on reasoning text even when the model complies — so the selected trial's 19/60 is a scoring artefact, not the served effect.

Measured with thinking on — the regime you serve — on a 44-prompt held-out harmful battery (bench-ablit.py, distinct from the optimiser's prompts), UD-IQ3_XXS, temp 0.1:

¹ The one stock marker "miss" was itself a refusal, so stock is effectively 44/44.

Markers mislead both ways, so the 44 were sorted by behaviour: a response that opens with a disclaimer then delivers the content counts as compliance, and one that pivots to a harmless reading of the prompt counts as deflection. The buckets (a human read of response openings, ±a few):

• ~34 genuinely comply with the harmful request.
• ~9 deflect to a safe reading.
• ~1 clean refusal.

At KL 0.169 this is the strongest ablation in this series, yet it cost no measured reasoning (see below).

Why int8 / unfused experts (provenance)

The base is 47.9 GB in bf16, which does not fit a 48 GB GPU, and the next card up costs ~4× as much. transformers 5 also loads lfm2_moe with its 64 experts fused into one 3-D parameter, so stock Heretic + PEFT would silently abliterate attention only, never the experts (22.9B of the 24.0B parameters).

Both were handled by a shim (heretic-mps/unfuse_moe.py, UNFUSE_INT8=1): the model is built on meta, the fused experts are unfused into per-expert int8 linears filled straight from the checkpoint (the fused bf16 experts never exist, on GPU or in host RAM — ~25 GB total), and Heretic searches against that. Only the search sees int8. The export merges the chosen trial's deltas onto the original bf16 shards (streaming_merge.py), so no int8 error reaches the released weights. The int8 experts were verified against bf16 on real activations before the search (relative error 0.5–1.0%; padding-batch KL 0.0003).

The quants

Five GGUFs + the imatrix. The three UD-* files replicate Unsloth's LFM2.5-24B-A2B-GGUF layout tensor-for-tensor (every per-tensor type cloned from their header; 428 tensors, 0 differ on each). Q4_K_M and Q8_0 use llama.cpp's stock mixes, matching the other repos in this series.

Unsloth published no imatrix for this model, and IQ2_XXS / IQ3_XXS cannot be built without one (it is a hard error, not a warning). So the importance matrix here is ours, computed from this abliterated model's own Q8_0 over bartowski's calibration_datav3. Only the imatrix rounding differs from Unsloth's files; the tensor types are identical.

Capability check

Abliteration damage concentrates in mathematical reasoning — a cross-architecture study (Young, 2025) puts Heretic's mean GSM8K change at −7.81 pp, with a −18.81 pp (−26.5%) worst case on Yi-1.5-9B while that model's MMLU and HellaSwag barely moved. At KL 0.169 that risk is the thing to check, so it was checked.

GSM8K, paired, n=100, thinking on, identical sampling on both builds (UD-IQ3_XXS):

Delta 0.00 pp. Discordant pairs split 1/1 (one problem each build missed that the other got), net flips 0, McNemar exact p = 1.000. No truncations on either side. A regression the size that study reports would need ~19 net flips against this build; there were none.

Caveats: n=100 excludes a large regression, not a 2–3 pp one; and MMLU, coding, long context, and the 15 non-English languages are unmeasured — a refusal-direction ablation could plausibly bite harder on the multilingual side than on English math.

The selected trial

Heretic's TPE search ran 80 trials and auto-selected trial 68 off the Pareto front: 19/60 refusals at KL 0.1689 (baseline 59/60). Abliteration is not bit-reproducible from the recipe — a re-run lands on a different trial — so the trial's parameters are published for Heretic's --reproduce:

direction_index                    = per layer
attn.o_proj.max_weight             = 1.43
attn.o_proj.max_weight_position    = 26.25
attn.o_proj.min_weight             = 1.34
attn.o_proj.min_weight_distance    = 17.11
mlp.down_proj.max_weight           = 1.45
mlp.down_proj.max_weight_position  = 24.65
mlp.down_proj.min_weight           = 0.50
mlp.down_proj.min_weight_distance  = 6.82

(attn.o_proj covers both self_attn.out_proj and the short-conv blocks' conv.out_proj; mlp.down_proj is each expert's feed_forward.w2. "per layer" means each layer used its own refusal direction.) More conservative points on the same front, if you want less behaviour change and even lower divergence: 40/60 @ KL 0.071 (trial 62), 38/60 @ KL 0.075 (trial 61).

Prompt sets

Heretic's built-in defaults, unchanged — refusal directions from the training splits, optimiser scoring on test[:60]:

The 44-prompt battery in the tables above is a held-out harmful set (bench-ablit.py), distinct from the optimiser's prompts. Neither dataset declares a license on the Hub.

Verifying the build

14e78558266e8ec41852810523d5eac6a5bca519c92fa31f120371f50532cef7  LFM2.5-24B-A2B-Abliterated-UD-IQ2_XXS.gguf
f7de3abe93401891580936573d6f7cc289b504f633628bc6f0609e12d4e5d76e  LFM2.5-24B-A2B-Abliterated-UD-IQ3_XXS.gguf
69034a38ff8cc1f147470c021970c4331da13670f9257e6fcf4e4e1425e30cfb  LFM2.5-24B-A2B-Abliterated-UD-Q3_K_XL.gguf
ed9be86be844d32bee1d810278b4f25bd27c36a2f8b096cbddb27c95b03a6b75  LFM2.5-24B-A2B-Abliterated-Q4_K_M.gguf
c3ea9ec2534eb513f20de3e03942309d8d44fb647911876e9bbe83567b4b04af  LFM2.5-24B-A2B-Abliterated-Q8_0.gguf
b32e4b4502691bf7d4203d25af9b9d980a3d6b4d9ecaa8cd906be896d8432403  LFM2.5-24B-A2B-Abliterated-imatrix.gguf

License

LFM Open License v1.0, inherited from LiquidAI/LFM2.5-24B-A2B — the full text is in LICENSE, copied verbatim from the base model repo (sha256 4d28ca14…70068c). Abliteration and quantization change nothing about it.

This is not an Apache/MIT-style license:

• §5 Commercial Use Limitation — the commercial rights granted are conditioned on you or your legal entity not exceeding the Threshold, defined in §1 as annual revenue of $10,000,000 USD or more. Commercial use by an entity over that threshold is not licensed by this agreement.
• Non-commercial and research use, and use by qualified non-profits for such purposes, are unrestricted by §5.
• §3 carries a patent-litigation termination clause; the license grants no trademark rights.

Read LICENSE yourself before any commercial deployment — the summary above is orientation, not legal advice.

Credits

• LiquidAI — the LFM2.5-24B-A2B base model
• p-e-w/heretic — the abliteration tooling and method
• ggml-org/llama.cpp — conversion and quantization
• unsloth — the UD quant layout these three UD files replicate
• mlabonne — harmful_behaviors and harmless_alpaca
• bartowski — calibration_datav3, used to compute the imatrix