Modelpublic

simaai/LFM2-2.6B-Autoround-Safetensors

sourceHugging Facemitupdated 26d agoView on Hugging Face

0likes198downloads

Model Card

LFM2-2.6B Autoround Safetensors

Post-training quantized `LiquidAI/LFM2-2.6B`, prepared as a pre-LLiMa Hugging Face checkpoint for Sima.ai. The immutable source revision was not captured in the local cache and must be pinned before any publication.

Quantization

Component	Method	Format	Details
All decoder Linear weights	AutoRound	symmetric INT4, G256	166 modules including short-convolution projection Linear layers; `lm_head` excluded; W4A16
`lm_head`	GPTQ	symmetric INT4, G256	static actorder; block size 128; dampening 0.01
Convolution kernels and activations	—	FP16/BF16	not quantized

Calibration used 512 fixed-order samples from NeelNanda/pile-10k, sequence length 1024, AutoRound 200 iterations, batch size 1, and device ID 0.

Evaluation

Full wikitext-2-raw-v1 was evaluated through lm-eval's wikitext task with no example limit and matched settings for both checkpoints.

Checkpoint	Word perplexity	Status
`LiquidAI/LFM2-2.6B`	28.261791339221	Full run: 2026-07-15
This checkpoint	29.264924001060	Full run: 2026-07-15
Absolute degradation	+1.003132661839	`quantized - source`
Relative degradation	+3.549430571470%	`100 * (quantized / source - 1)`

Raw JSON evidence is retained in perplexity_results/full_wikitext/ in this workspace.

Reproduction

Included files: README.md, quantize.py, recipe.yaml, and versions.txt.

bash

python quantize.py --model-path /path/to/LFM2-2.6B --output-dir /path/to/new-output-directory

The quantization job completed on an NVIDIA A100 80 GB GPU on 2026-07-15 and saved scales passed the NaN/Inf check. The compressed-checkpoint smoke test passed: 167 compressed modules including lm_head loaded and completed greedy generation.

Deployment and limitations

This is a pre-LLiMa quantized Hugging Face safetensors artifact. Compile separately for Sima.ai with the appropriate LLiMa release. Quantization quality can vary by language, domain, prompt format, and context length; validate your deployment.