Laguna XS 2.1 · CRACK · GGUF
Compact Laguna MoE coding model · llama.cpp GGUF · Q6_K · Q4_K_M · Q2_K
Refusal removed with capability preserved — 0 refusals across HarmBench-320, MMLU within ~0.6 pts of base
A permissive research variant of poolside/Laguna-XS-2.1 — the compact member of the Laguna agentic-coding Mixture-of-Experts family — packaged for llama.cpp as GGUF. Refusal behaviour has been removed while the model's coding, reasoning, tool-use, and general knowledge are preserved.
This public card intentionally describes compatibility, evaluation, and limitations only. Internal creation details are not published.
Model details
| Property | Value |
|---|---|
| Base model | poolside/Laguna-XS-2.1 |
| Architecture | Laguna MoE — 256 routed experts + 1 shared, top-8 router |
| Layers | 40 (10 global : 30 sliding-window attention) |
| Hidden size | 2,048 |
| Context | up to 262,144 tokens · thinking and non-thinking modes |
| Format | GGUF for llama.cpp |
| Quant scheme | High-precision attention + router; low-bit routed experts (importance-matrix calibrated) |
| Languages | English, Chinese (+ the base model's full multilingual coverage) |
The tokenizer, Laguna chat template (with <think> reasoning and tool-call support), and generation config are bundled. Requires a recent llama.cpp build with Laguna architecture support.
Quantizations
| File | Bits | Size | Recommended for |
|---|---|---|---|
Laguna-XS-2.1-CRACK-Q6_K.gguf |
~6-bit | ~28 GB | Highest quality; near-lossless |
Laguna-XS-2.1-CRACK-Q4_K_M.gguf |
~4-bit | ~21 GB | Best balance — recommended default |
Laguna-XS-2.1-CRACK-Q2_K.gguf |
~2-bit | ~13 GB | Smallest; runs on modest memory |
All quants use importance-matrix calibration and keep attention and the router at high precision, so quality holds up even at 2-bit.
Evaluation
Capability is preserved — MMLU-logit accuracy, each quant vs the unmodified base at the same quant (full 14,042-question test, reasoning off):
| Quant | Base | CRACK | Δ |
|---|---|---|---|
| Q6_K | 73.24% | 73.77% | +0.53 |
| Q4_K_M | 73.54% | 73.63% | +0.09 |
| Q2_K | 70.18% | 69.58% | −0.60 |
Every quant lands within ~0.6 points of its base — Q6 and Q4 are capability-positive.
Per-topic MMLU: CRACK vs base by super-category (click to expand)Each cell is the CRACK accuracy with its delta vs the same-quant base (full 14,042 MMLU-logit, reasoning off):
| MMLU category | Q6_K CRACK (Δ) | Q4_K_M CRACK (Δ) | Q2_K CRACK (Δ) |
|---|---|---|---|
| STEM | 70.1 (+0.6) | 69.8 (+0.1) | 67.1 (−1.1) |
| Humanities | 69.4 (+0.8) | 69.5 (+0.5) | 66.2 (+1.0) |
| Social Sciences | 81.9 (−0.1) | 81.6 (−0.6) | 77.1 (−1.3) |
| Other (business, health, misc.) | 75.9 (+0.6) | 75.7 (+0.1) | 69.7 (−1.8) |
| Overall | 73.8 (+0.5) | 73.6 (+0.1) | 69.6 (−0.6) |
Capability holds within ~2 points of base across every topic area, even at 2-bit.
Refusal is removed — full HarmBench-320 compliance (higher = fewer refusals):
| Quant | HarmBench-320 | Standard | Contextual | Copyright |
|---|---|---|---|---|
| Q6_K | 98.8% (0 refusals) | 98.7% | 98.8% | 98.8% |
| Q4_K_M | 99.1% (0 refusals) | 98.7% | 100% | 98.8% |
| Q2_K | 98.4% (0 refusals) | 100% | 98.8% | 95.0% |
Across all 320 HarmBench behaviours the model produced zero refusals at every quant.
Usage (llama.cpp)
# download one quant
hf download dealignai/Laguna-XS-2.1-CRACK-GGUF Laguna-XS-2.1-CRACK-Q4_K_M.gguf --local-dir .
# run
llama-cli -m Laguna-XS-2.1-CRACK-Q4_K_M.gguf -ngl 99 -c 8192 --jinja \
-p "Write a Python function to reverse a linked list."
Thinking mode is on by default via the bundled chat template; pass --jinja so llama.cpp applies it.
Limitations & responsible use
This model has had refusal behaviour removed and will attempt to comply with harmful, unsafe, or unethical requests. It is intended for security research, red-teaming, and evaluation. You are solely responsible for lawful and ethical use. Not for all audiences.
Base model © poolside, released under the OpenMDW-1.1 license; this quantized variant inherits that license.