Leanstral-1.5-119B-A6B

FP8 · vllm · 256K context · 2 × DGX Spark GB10 · 256 GB

Mistral's Leanstral 1.5 for Lean 4 proof engineering.

21 tok/sdecode

– tok/sprefill

256Kcontext window

Reported by MiaAI-LabJul 4, 2026

Run it

This is the launch as MiaAI-Lab publishes it. Their repository has the full setup: MiaAI-Lab/Leanstral-1.5-119B-A6B_2x_DGX_Sparks_vLLM@94d729896226 › The server then answers on http://localhost:8000/v1.

1 Set up first

# Each container pip-installs ray[default] at start (the image lacks the Ray CLI), so both nodes need internet. Head container runs Ray head, waits for 2 active nodes, then vllm serve. Worker (over SSH to 10.0.0.2) runs 'ray start --node-ip-address=10.0.0.2 --address=10.0.0.1:6379 --object-store-memory=4294967296 --temp-dir=/tmp/ray --block' with VLLM_HOST_IP=10.0.0.2, RAY_TMPDIR=/tmp/ray and spill dir /mnt/spark1/ray/leanstral-vllm-spill, no RAY_ADDRESS. Host dirs /mnt/ray/leanstral-vllm and /mnt/ray/leanstral-vllm-spill are bind-mounted at the same paths. HF caches: head /home/zurih/.cache/huggingface, worker WORKER_HF_HOME (default /mnt/spark1/models/.cache/huggingface); weights download on first start from inside each container. Passwordless SSH to the worker.
# Full instructions: https://github.com/MiaAI-Lab/Leanstral-1.5-119B-A6B_2x_DGX_Sparks_vLLM/tree/94d729896226db3eaf7f584e79b7c1f0028bab1d

2 Download the weights

hf download mistralai/Leanstral-1.5-119B-A6B \
  --revision 81592da95d94ab0439bfce16df1d55b402e598b6

3 Start the server on each of the 2 machines (NODE_RANK 0 to 1)

docker run --rm \
  --gpus all \
  -p 8000:8888 \
  --gpus all \
  --ipc=host \
  --network host \
  --shm-size 16g \
  --shm-size 16g \
  -e GLOO_SOCKET_IFNAME=enp1s0f1np1 \
  -e HF_HOME=/root/.cache/huggingface \
  -e HF_HUB_DISABLE_XET=1 \
  -e HF_HUB_DOWNLOAD_TIMEOUT=600 \
  -e HF_HUB_ETAG_TIMEOUT=600 \
  -e HF_TOKEN='${HF_TOKEN:-${HUGGING_FACE_HUB_TOKEN:-}}' \
  -e HUGGING_FACE_HUB_TOKEN='${HUGGING_FACE_HUB_TOKEN:-${HF_TOKEN:-}}' \
  -e NCCL_SOCKET_IFNAME=enp1s0f1np1 \
  -e RAY_ADDRESS=10.0.0.1:6379 \
  -e RAY_TMPDIR=/mnt/ray/leanstral-vllm \
  -e RAY_memory_usage_threshold=0.99 \
  -e RAY_object_spilling_config='{"type":"filesystem","params":{"directory_path":"/mnt/ray/leanstral-vllm-spill"}}' \
  -e VLLM_ENGINE_ITERATION_TIMEOUT_S=600 \
  -e VLLM_HOST_IP=10.0.0.1 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  --entrypoint bash \
  vllm/vllm-openai:v0.24.0@sha256:251eba5cc7c12fed0b75da22a9240e582b1c9e39f6fbc064f86781b963bd814f \
  -lc 'python3 -m pip install -q '\''ray[default]'\'' && ray start --head --node-ip-address='\''10.0.0.1'\'' --port='\''6379'\'' --dashboard-host=0.0.0.0 --object-store-memory='\''4294967296'\'' --temp-dir='\''/mnt/ray/leanstral-vllm'\'' && until ray status ... (2 active nodes); do sleep 3; done && vllm serve ...' \
  serve \
  mistralai/Leanstral-1.5-119B-A6B \
  --host 0.0.0.0 \
  --port 8888 \
  --max-model-len 262144 \
  --max-num-seqs 7 \
  --max-num-batched-tokens 16384 \
  --mm-processor-cache-type lru \
  --tensor-parallel-size 2 \
  --gpu-memory-utilization 0.85 \
  --enforce-eager \
  --no-async-scheduling \
  --distributed-executor-backend ray \
  --tool-call-parser mistral \
  --enable-auto-tool-choice \
  --reasoning-parser mistral

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/Leanstral-1.5-119B-A6B_2x_DGX_Sparks_vLLM@94d729896226, where it reports 21.08 tok/s. Our six checks have not run on it yet.

Details

Weights: mistralai/Leanstral-1.5-119B-A6B @ 81592da95d ›
Image: vllm/vllm-openai:v0.24.0@sha256:251eba5cc7c12fed0b75da22a9240e582b1c9e39f6fbc064f86781b963bd814f
Source: MiaAI-Lab/Leanstral-1.5-119B-A6B_2x_DGX_Sparks_vLLM @ 94d729 ›
Launch: vllm-leanstral-1.5-119b-a6b-fp8-256k-2x ›
Recipe file: registry/recipes/nvidia/dgx-spark-gb10-128gb/leanstral-1.5-119b-a6b.vllm.256k.2x.json ›
Model released: Jul 1, 2026
Tested: not yet by us; reported by miaai-lab