Laguna-XS-2.1-TQ3_4S

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

This is a standard (non-MTP) model — no draft-MTP flags are needed.

Parent Model

• Upstream parent: 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

Files

Quantization Recipe

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

./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

./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

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)

Benchmarks

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

Comparison (all TQ3_4S, same RTX 3090)

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 (Poolside)
• Runtime: turbo-tan/llama.cpp-tq3 (MIT)