---
license: openmdw-1.1
language:
- en
library_name: gguf
pipeline_tag: text-generation
model_name: TurboLaguna-XS
tags:
- gguf
- llama.cpp
- laguna
- poolside
- turboquant
- tq3_4s
- code
base_model:
- poolside/Laguna-XS-2.1-GGUF
model-index:
- name: TurboLaguna-XS
  results: []
---

# TurboLaguna-XS

**Canonical artifact:** `Laguna-XS-2.1-TQ3_4S`

`TurboLaguna-XS` is the TurboQuant GGUF build of Poolside's Laguna XS 2.1 — a sigmoid-routed mixture-of-experts coding model with 256 experts per layer, a shared expert, QK-norm, and hybrid YaRN/sliding-window RoPE.

The exact file and runtime artifact name is:

- `Laguna-XS-2.1-TQ3_4S.gguf`

## Required Runtime

> **This model uses the custom `TQ3_4S` tensor type.** Stock `llama.cpp` builds
> **cannot** load it. You must use the TurboQuant runtime fork:
>
> **[turbo-tan/llama.cpp-tq3](https://github.com/turbo-tan/llama.cpp-tq3)**
>
> This is a standard (non-MTP) model — no draft-MTP flags are needed.

## Parent Model

- Upstream parent: [poolside/Laguna-XS-2.1-GGUF](https://huggingface.co/poolside/Laguna-XS-2.1-GGUF)
- Source quant: `Laguna-XS-2.1-BF16.gguf` (63.8 GB, 16.01 bpw)
- Format conversion and TurboQuant packaging: [turbo-tan/llama.cpp-tq3](https://github.com/turbo-tan/llama.cpp-tq3)

## Files

| File | Size | Notes |
|---|---|---|
| `Laguna-XS-2.1-TQ3_4S.gguf` | 16 GB (4.05 bpw) | Main model — 678 tensors, 40 layers × 256 routed experts |
| `thumbnail.png` | — | Model card image |
| `benchmark.png` | — | Benchmark summary |

## Quantization Recipe

Quantized from the official BF16 GGUF using the standard TQ3_4S recipe:

```bash
./build/bin/llama-quantize --allow-requantize \
  --output-tensor-type q6_K \
  --token-embedding-type q6_K \
  Laguna-XS-2.1-BF16.gguf \
  Laguna-XS-2.1-TQ3_4S.gguf \
  TQ3_4S
```

Tensor policy:
- Routed experts, attention projections, shared experts → `tq3_4s` (4.0 bpw)
- Token embeddings, output head → `q6_K`
- Norms, gates, biases → `f32` (untouched)

Result: **63.8 GB → 16 GB** (3.98× compression), 42% smaller than Poolside's own Q4_K_M (20 GB).

## Recommended Runtime

```bash
./build/bin/llama-server \
  -m Laguna-XS-2.1-TQ3_4S.gguf \
  --host 127.0.0.1 --port 8080 \
  -c 8192 -np 1 -ngl 99 -fa on \
  --reasoning off --jinja
```

Build note:
- `-fa on` is the runtime flash-attention flag, not the CMake `GGML_CUDA_FA_ALL_QUANTS` build flag.

## GPU Memory Profiles

| GPU memory | Suggested context | KV cache | Notes |
|---:|---:|---|---|
| 16 GiB | `4096` | `-ctk q4_0 -ctv tq3_0` | Tight fit — keep context small |
| 24 GiB | `8192` to `32768` | `-ctk q8_0 -ctv tq3_0` | Validated desktop profile |
| 128 GiB GB10 | `65536+` | `-ctk q4_0 -ctv tq3_0` | Full headroom for long context |

## Tested Hardware

- **NVIDIA RTX 3090 24 GB** — primary validation platform
- llama.cpp-tq3 fork, branch `feat/laguna-arch` (Laguna arch from upstream [ggml-org/llama.cpp#25165](https://github.com/ggml-org/llama.cpp/pull/25165))

## Benchmarks

All scores: greedy decoding, reasoning off, `-ngl 99 -fa on`, RTX 3090.

| Benchmark | Score | tok/s |
|---|---|---|
| HumanEval (base) | **0.805** | 196 |
| HumanEval+ (extra tests) | **0.762** | 196 |
| MBPP (base) | **0.833** | 199 |
| MBPP+ (extra tests) | **0.720** | 199 |
| Hard86 (20 tasks / 86 assertions) | **64.0%** (55/86) | 202 |
| BenchLoop coding | **100.0** (12/12) | — |
| BenchLoop overall | **73.7** | — |
| BenchLoop speed | **96.4** (9/9) | — |

### Comparison (all TQ3_4S, same RTX 3090)

| Model | HE+ | MBPP+ | Hard86 | Coding | tok/s | Size |
|---|---|---|---|---|---|---|
| **Laguna XS 2.1** | 0.762 | 0.720 | 64.0% | 100.0 | **196** | 16 GB |
| Qwen3.5 9B | 0.671 | 0.563 | 44.2% | 79.2 | 134 | 4.5 GB |
| Qwen3.6 27B MTP | 0.927 | 0.878 | — | 100.0 | 42–54 | 12.9 GB |

Laguna XS is a coding specialist: perfect BenchLoop coding (same as the 27B), +19.8pp Hard86 over the 9B, at 3.6–4.6× the 27B's decode speed.

## Validation

```
llama-simple-chat coherence smoke: PASS
llama-server --reasoning off strict smoke: PASS (content = "ok")
llama-bench pp2048: 745 tok/s
llama-bench tg128: 196 tok/s
evalplus HE/HE+/MBPP/MBPP+: scored (see above)
hard86: 55/86
benchloop v0.2.3: overall 73.7
```

## License

- Parent model: [OpenMDW-1.1](https://huggingface.co/poolside/Laguna-XS-2.1) (Poolside)
- Runtime: [turbo-tan/llama.cpp-tq3](https://github.com/turbo-tan/llama.cpp-tq3) (MIT)