brain
Model market REAL

Models on Brain Nodes.

The allowlist of models a Brain Node may run, and which of them have a node online this minute. Availability comes from the coordinator's registry; VRAM figures are requirements, not measurements. Prices shown are node ask prices when operators set one; otherwise the network list price applies.
Network →API →
Allowlisted modelsloading
ModelParamsMin VRAMLicenseNodes onlineLoadedSpeedLowest askRegionsAsked 24hServedUnservedStatus
qwen/qwen2.5-1.5b-instruct 1.5B5.9 GBApache-2.0……unmeasured——………no nodes
qwen/qwen2.5-3b-instruct 3B9.8 GBQwen Research……unmeasured——………no nodes
qwen/qwen2.5-7b-instruct 7B20 GBApache-2.0……unmeasured——………no nodes
qwen/qwen2.5-7b-instruct-awq 7B9.8 GBApache-2.0……unmeasured——………no nodes
mistralai/mistral-7b-instruct-v0.3 7B20 GBApache-2.0……unmeasured——………no nodes
meta-llama/llama-3.1-8b-instruct 8B20 GBLlama 3.1 Community (gated; node needs HF_TOKEN)……unmeasured——………no nodes
qwen/qwen2.5-14b-instruct 14B33 GBApache-2.0……unmeasured——………no nodes
qwen/qwen2.5-14b-instruct-awq 14B14 GBApache-2.0……unmeasured——………no nodes
qwen/qwen2.5-32b-instruct 32B70 GBApache-2.0……unmeasured——………no nodes
qwen/qwen2.5-32b-instruct-awq 32B23 GBApache-2.0……unmeasured——………no nodes
qwen/qwen2.5-coder-32b-instruct 32B70 GBApache-2.0……unmeasured——………no nodes
qwen/qwen2.5-coder-32b-instruct-awq 32B23 GBApache-2.0……unmeasured——………no nodes
qwen/qwen2.5-72b-instruct 72B146 GBQwen (non-commercial above 100M MAU)……unmeasured——………no nodes
qwen/qwen2.5-72b-instruct-awq 72B47 GBQwen (non-commercial above 100M MAU)……unmeasured——………no nodes
deepseek-ai/deepseek-r1-distill-qwen-32b 32B70 GBMIT……unmeasured——………no nodes
deepseek-ai/deepseek-r1-distill-llama-8b 8B20 GBMIT (Llama 3.1 Community for the base weights)……unmeasured——………no nodes
meta-llama/llama-3.3-70b-instruct 70B146 GBLlama 3.3 Community (gated; node needs HF_TOKEN)……unmeasured——………no nodes
brain/mock mock0—n/a……unmeasured——………mock

Speed is the median of coordinator-timed tokens per second on completed jobs for nodes currently serving the model; it reads unmeasured until a job has run. Asked / served / unserved are counts of customer orders naming the model in the last 24 h, from the order log; unserved means no node could take it or every attempt failed. Pricing: the network list price applies unless a node sets an ask; asks are a routing input and are shown on the node's provider page.

Pin a modeloperators

A request that names a model only ever goes to a node serving that model. Pin the ones you want to compete on and the router sends you that traffic, ranked against other nodes on capability, availability, measured latency, reputation and your ask price. Pinned models are advertised on registration and checked against what your server actually serves.

# vLLM (NVIDIA): the agent pulls and serves the pinned ids
BRAIN_NODE_MODE=vllm BRAIN_NODE_MODELS=qwen/qwen2.5-32b-instruct-awq,qwen/qwen2.5-coder-32b-instruct-awq npm run node

# Ollama / llama.cpp / mlx-lm / exo: serve the model yourself, pin what to advertise
ollama pull qwen2.5:32b-instruct
BRAIN_NODE_MODE=ollama BRAIN_NODE_MODELS=qwen/qwen2.5-32b-instruct npm run node

# Optional: an ask price per 1M tokens; the router treats it as one input among five
BRAIN_NODE_ASK_USD_PER_1M=0.40