Laguna-XS-2.1 — Abliterated, NVFP4

This is poolside/Laguna-XS-2.1 (BF16) with its refusal direction removed via abliterix (Optuna-optimized directional ablation), then quantized to NVFP4 (nvfp4_experts_only weights + FP8 KV) using NVIDIA's TensorRT-Model-Optimizer. Both steps were done by us, off-box on a rented H100.

Abliteration result

Refusals before	Refusals after
98/100	4/100

Usage

Requires trust_remote_code=True (Laguna ships custom modeling code, modeling_laguna.py / configuration_laguna.py, included in this repo).

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "qpqpqpqpqpqqqq/Laguna-XS-2.1-Abliterated-NVFP4",
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(
    "qpqpqpqpqpqqqq/Laguna-XS-2.1-Abliterated-NVFP4",
    trust_remote_code=True,
)

Serving via vLLM (0.30+): --trust-remote-code --moe-backend humming.

Performance

Measured on this exact checkpoint, DGX Spark (GB10, SM121), vLLM 0.30, isolated (no concurrent load), --moe-backend humming, DFlash speculative decoding, num_speculative_tokens=15:

Workload	Decode throughput	Draft accept length
Code (temp 0, thinking)	147.7–177.5 tok/s (mean 162.6)	6.59–7.98
Code (non-thinking)	153.1 tok/s	6.92
Prose (thinking)	51.6–54.5 tok/s (mean 53.0)	2.38–2.49

Speculative decoding draft

Pairs with poolside/Laguna-XS-2.1-DFlash-NVFP4 (unmodified, poolside's own release) — not bundled in this repo.

Notes

Abliteration removes safety alignment. Use responsibly and in accordance with your local laws and the upstream OpenMDW-1.1 license.
Quantization: nvfp4_experts_only (MoE expert weights only) + FP8 KV cache, matching the scheme poolside's own NVFP4 release uses.

Refusals before

Refusals after

98/100

4/100

Usage

Requires trust_remote_code=True (Laguna ships custom modeling code, modeling_laguna.py / configuration_laguna.py, included in this repo).

from transformers import AutoModelForCausalLM, AutoTokenizer model = AutoModelForCausalLM.from_pretrained( "qpqpqpqpqpqqqq/Laguna-XS-2.1-Abliterated-NVFP4", torch_dtype="auto", device_map="auto", trust_remote_code=True, ) tokenizer = AutoTokenizer.from_pretrained( "qpqpqpqpqpqqqq/Laguna-XS-2.1-Abliterated-NVFP4", trust_remote_code=True, )

Serving via vLLM (0.30+): --trust-remote-code --moe-backend humming.

Performance

Measured on this exact checkpoint, DGX Spark (GB10, SM121), vLLM 0.30, isolated (no concurrent load), --moe-backend humming, DFlash speculative decoding, num_speculative_tokens=15:

Workload

Decode throughput

Draft accept length

Code (temp 0, thinking)

147.7–177.5 tok/s (mean 162.6)

6.59–7.98

Code (non-thinking)

153.1 tok/s

6.92

Prose (thinking)

51.6–54.5 tok/s (mean 53.0)

2.38–2.49