LFM2.5-2.6B is a Liquid AI 2.6B-parameter hybrid edge model built for agentic workloads: 30 layers (22 double-gated short-convolution blocks + 8 GQA), a 128K context window, 128K vocabulary, and a ChatML-like template with native <think> reasoning.

These quantized GGUFs are built from our LFM2.5-2.6B-Uncensored BF16 release (abliterix Trial 65, stream-merged to BF16) in three steps:

1. BF16 GGUF conversion with llama.cpp (lfm2 architecture support).
2. imatrix calibration — 401 chunks (≈1.6M tokens) from the APEX calibration set, computed on the BF16 GGUF.
3. Quantization with llama-quantize --imatrix into five tiers.

License: LFM Open License v1.0 (same as the base model).

After merging abliterix Trial 65, this model shows a much lower refusal rate and can differ substantially from official LFM2.5-2.6B. Evaluate compliance and safety for your use case; control access and audit as needed.

Implementation sketch: LoRA merge W += (B @ A) * (alpha / r) (this trial alpha = r = 1); steering applied to attn.o_proj / conv.out_proj / mlp.down_proj across 30 layers.

Decode speeds measured on AMD Strix Halo (Radeon 8060S, gfx1151) with llama.cpp ROCm 7.2, 128K context. All files are lfm2 architecture, 128K context, single-file GGUFs.

The importance matrix (computed over 401 chunks / ≈1.6M tokens of mixed conversation, math, and code data) tells the quantizer which weights are sensitive. K-quants and especially the IQ tiers use it to keep more bits on attention/embedding paths — the parts that matter most for subtle behaviors like identity and instruction following on a 2.6B model. Compared to plain Q4_K_M, IQ4_XS is smaller and faster while holding comparable perplexity.

The lfm2 architecture is supported by llama.cpp (and LM Studio / other GGUF runners).

llama-server -m LFM2.5-2.6B-Uncensored-IQ4_XS.gguf \
  --ctx-size 131072 --flash-attn on --host 0.0.0.0 --port 8080

Or with llama-cli:

llama-cli -m LFM2.5-2.6B-Uncensored-Q4_K_M.gguf \
  -p "What is 2+2?" -n 512 \
  --temp 0.1 --top-k 50 --repeat-penalty 1.1

Transformers / vLLM / SGLang users: use the BF16 safetensors in the parent repo.

Keep the official generation defaults: temperature 0.1, top_k 50, repetition_penalty 1.1. If you want more creative answers, raise temperature toward 0.6–0.8; note the model always thinks before answering, so allow enough max_new_tokens (512+) for the <think> block.

1. abliterix trial search on ROCm (gfx1151): 60 trials + 20 warmup, seed 117; all trials pruned only by same-prefix kl_divergence < 0.5.
2. Selected Trial 65: refusals 6/100 (baseline 90/100), KL 0.0335, length deviation 0.079 σ, generation health PASSED.
3. LoRA stream-merged into base weights in BF16 (W += B@A, alpha = r = 1).
4. BF16 GGUF conversion via convert_hf_to_gguf.py --outtype bf16 (llama.cpp lfm2).
5. imatrix calibration: 401 chunks / ≈1.6M tokens (APEX calibration set), computed on the BF16 GGUF.
6. Quantized with llama-quantize --imatrix <imatrix.gguf> <src> <dst> <type> for each tier.