Laguna-S-2.1

NVFP4 · vllm · 256K context · DGX Spark GB10 · 128 GB

Poolside's Laguna S 2.1 coding model.

77 tok/sdecode

– tok/sprefill

256Kcontext window

Reported by MiaAI-LabJul 22, 2026

Run it

This is the launch as MiaAI-Lab publishes it. Their repository has the full setup: MiaAI-Lab/Laguna-S-2.1-DGX-Spark-RTX-6000-PRO@750a52ead5e9 › The server then answers on http://localhost:8000/v1.

1 Set up first

# Container entrypoint is a bootstrap script written to /tmp/laguna-bootstrap.sh on the host and mounted at /bootstrap.sh:ro. It pip-installs the FlashInfer nightly trio 0.6.15.dev20260712 (+cu130 jit-cache) plus cuda-tile 1.5.0 and nccl4py 0.3.1 at every cold start, then execs 'vllm serve "$@"' (needs internet access to flashinfer.ai). Mounts ~/.cache/huggingface, ~/.cache/flashinfer -> /root/.cache/flashinfer (persistent JIT cache), and the repo dir -> /workspace. Both target and draft downloaded on host (hf/huggingface-cli/docker fallback). First start about 15 min.
# Full instructions: https://github.com/MiaAI-Lab/Laguna-S-2.1-DGX-Spark-RTX-6000-PRO/tree/750a52ead5e92d4f8c2b42321c6695bcc3cf6360

2 Download the weights

hf download poolside/Laguna-S-2.1-NVFP4 \
  --revision 826aacdf6d8b2699d4e367def6f17c83b06044c2

3 Start the server

docker run --rm \
  --gpus all \
  -p 8000:8888 \
  --user root \
  --network host \
  --shm-size=32g \
  --ulimit memlock=-1:-1 \
  --cap-add=IPC_LOCK \
  --ipc host \
  --gpus all \
  --workdir /workspace \
  --shm-size 32g \
  -e CUTE_DSL_ARCH=sm_121a \
  -e FLASHINFER_NVCC_THREADS=2 \
  -e HF_HOME=/root/.cache/huggingface \
  -e HF_TOKEN='${HF_TOKEN:-}' \
  -e MAX_JOBS=4 \
  -e NVCC_THREADS=2 \
  -e PATH='/usr/local/cuda/bin:${PATH}' \
  -e VLLM_TARGET_DEVICE=cuda \
  -e VLLM_TRUST_REMOTE_CODE=1 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  --entrypoint /bootstrap.sh \
  vllm/vllm-openai:v0.25.1@sha256:e4f88a835143cd22aee2397a26ec6bb80b3a4a6fe0c882bcbc63822904766089 \
  poolside/Laguna-S-2.1-NVFP4 \
  --host 0.0.0.0 \
  --port 8888 \
  --tensor-parallel-size 1 \
  --dtype bfloat16 \
  --attention-backend FLASHINFER \
  --trust-remote-code \
  --gpu-memory-utilization 0.85 \
  --max-model-len 262144 \
  --max-num-seqs 4 \
  --max-num-batched-tokens 8192 \
  --enable-auto-tool-choice \
  --tool-call-parser poolside_v1 \
  --reasoning-parser poolside_v1 \
  --default-chat-template-kwargs '{"enable_thinking":true}' \
  --override-generation-config '{"temperature":0.7,"top_p":0.95,"top_k":20}' \
  --speculative-config '{"model":"poolside/Laguna-S-2.1-DFlash-NVFP4","num_speculative_tokens":7,"method":"dflash"}'

What it passed

– loadLoads and serves within an hour

– chatAnswers a plain question and stops on its own

– reasoningThinks separately and gets 17 × 23 right

– toolsCalls a tool with the right arguments and uses the result

– contextRecalls a code buried in a prompt that fills 85% of the window

– speedDecodes at 15 tok/s or more

Published by MiaAI-Lab in MiaAI-Lab/Laguna-S-2.1-DGX-Spark-RTX-6000-PRO@750a52ead5e9, where it reports 76.7 tok/s. Our six checks have not run on it yet.

Details

Weights: poolside/Laguna-S-2.1-NVFP4 @ 826aacdf6d ›
Image: vllm/vllm-openai:v0.25.1@sha256:e4f88a835143cd22aee2397a26ec6bb80b3a4a6fe0c882bcbc63822904766089
Source: MiaAI-Lab/Laguna-S-2.1-DGX-Spark-RTX-6000-PRO @ 750a52ead5e9 ›
Launch: vllm-laguna-s-2.1-nvfp4-256k ›
Recipe file: registry/recipes/nvidia/dgx-spark-gb10-128gb/laguna-s-2.1.vllm.256k.json ›
Model released: Jul 13, 2026
Tested: not yet by us; reported by miaai-lab