8-bit weight-only quantization of LiquidAI/LFM2.5-2.6B, produced with Intel AutoRound and packed for GPTQ/Marlin kernels so it runs on vLLM without any extra flags.

Built and released by Plavno.

This is the accuracy-preserving variant. A smaller 4-bit sibling is available at plavno/LFM2.5-2.6B-AutoRound-W4A16, but it carries a real quality cost — see Choosing between the two.

Quantization details


Method	AutoRound (signed-gradient rounding optimization)
Weight precision	8-bit integer, symmetric
Activations	fp16 (not quantized)
Group size	128
Packing format	`auto_round:auto_gptq`
Quantized modules	166 of 167 — all `model.layers` linear projections
Kept in higher precision	`lm_head`
Calibration	conversational multilingual text, 256 samples × 512 tokens, balanced 16×16 across all 16 supported languages
Weights on disk	~2.9 GB (vs ~5.4 GB bf16)

Calibration

This build and its 4-bit sibling plavno/LFM2.5-2.6B-AutoRound-W4A16 were calibrated on the same language-balanced set — the only difference between the two repos is the weight bit width.

The set is conversational multilingual text: 256 samples of 512 tokens, with an exactly equal share of 16 samples per language across all 16 languages the base model supports:

Arabic · Chinese · English · French · German · Hindi · Indonesian · Italian · Japanese · Korean · Polish · Portuguese · Russian · Spanish · Thai · Vietnamese

Calibration was deliberately balanced rather than English-only, so that quantization error is not concentrated on the non-English languages. The register is conversational (utterance-level) rather than encyclopedic prose, which suits dialogue and speech-derived workloads.

Running on vLLM

LFM2 is natively supported from vLLM 0.23.0 onward. The quantization is detected automatically from config.json, so no --quantization flag is needed.

bash

pip install "vllm>=0.23"

vllm serve plavno/LFM2.5-2.6B-AutoRound-W8A16 \
  --max-model-len 4096 \
  --gpu-memory-utilization 0.90

Add --enable-prefix-caching when many requests share a long common prefix (system prompt, few-shot block, shared context) — only the new suffix is then prefilled on each request:

bash

vllm serve plavno/LFM2.5-2.6B-AutoRound-W8A16 \
  --enable-prefix-caching \
  --max-model-len 4096

Query the OpenAI-compatible endpoint:

bash

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "plavno/LFM2.5-2.6B-AutoRound-W8A16",
    "messages": [{"role": "user", "content": "Say hi in one word."}],
    "max_tokens": 16
  }'

python

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
resp = client.chat.completions.create(
    model="plavno/LFM2.5-2.6B-AutoRound-W8A16",
    messages=[{"role": "user", "content": "List three primary colors."}],
    max_tokens=64,
)
print(resp.choices[0].message.content)

Offline batch inference

python

from vllm import LLM, SamplingParams

llm = LLM(model="plavno/LFM2.5-2.6B-AutoRound-W8A16", max_model_len=4096)
out = llm.generate(
    ["Explain what a REST API is in two sentences."],
    SamplingParams(temperature=0, max_tokens=128),
)
print(out[0].outputs[0].text)

Choosing between the two

	W8A16 (this repo)	W4A16
Weights on disk	~2.9 GB	~1.7 GB
Quality vs bf16	near-lossless	noticeable degradation, uneven across languages
Best for	accuracy-sensitive work, multilingual input, reading token logprobs	tight VRAM budgets, maximum decode throughput

Pick this 8-bit build when output quality matters, when the workload is multilingual, or when you read token-level logprobs (classification, routing, scoring) — the 4-bit build shifts the output distribution enough to affect those.

Pick the 4-bit build when VRAM is the binding constraint or when decode throughput matters more than fidelity.

Notes

Short convolution layers of the LFM2 architecture and lm_head stay in higher precision; this is expected and correct for this architecture.
The chat template opens a reasoning block (<think>) in the generation prompt. If you need a bare single-token answer (classification, routing), drive /v1/completions with your own prompt and close the block yourself (<think></think>) rather than relying on /v1/chat/completions.
8-bit weight-only quantization reduces the bytes moved per decode step, which helps memory-bandwidth-bound generation. It does not speed up compute-bound prefill, where the dequantization step is additional work.

License

Inherits the license of the base model, LFM Open License v1.0. Refer to the base model card for intended use and limitations.

Quantized and published by Plavno.