Edge-Native MoE Inference Engine

FreeToken Wiki

Run frontier open-weight MoE models locally across consumer NVIDIA GPUs, with adaptive CPU-GPU execution, efficient caching, quantization, and OpenAI or Anthropic compatible APIs.

View on GitHub

FreeToken Wiki Guides

Everything you need to install, configure, and run large MoE models locally with FreeToken

Latest Updates

Discover the newest guides, tips, and content

FreeToken 290b model: 2026 Edge MoE Setup Guide

Learn how FreeToken serves frontier-scale MoE models on consumer hardware through adaptive caching, bandwidth scheduling, and elastic memory.

Aug 25, 2026performance
Read more →
FreeToken 4gb vram: Setup Guide & Hardware Limits

Learn how FreeToken handles limited VRAM, CPU offload, expert caching, and realistic hardware requirements for local MoE serving.

Aug 25, 2026hardware
Read more →
FreeToken 753b model: Setup Guide and MoE Tuning Tips

Learn how FreeToken serves 753B GLM-5.2 locally, including setup, memory paging, MoE caching, benchmarks, and practical tuning advice.

Aug 25, 2026performance
Read more →
FreeToken anthropic api: Step-by-Step Setup Guide

Set up Anthropic API access for a FreeToken project with secure keys, environment variables, request testing, and troubleshooting tips.

Aug 25, 2026api
Read more →
FreeToken api: Local MoE Serving Setup Guide 2026

Learn how FreeToken serves frontier-scale MoE models locally with expert caching, adaptive CPU-GPU execution, and elastic memory management.

Aug 25, 2026api
Read more →
FreeToken api server: Setup Guide & Runtime Tips

Learn how the FreeToken api server handles MoE caching, CPU-GPU execution, hardware checks, and practical local serving decisions.

Aug 25, 2026api
Read more →
FreeToken benchmark: Local MoE Serving Setup Guide

Review the FreeToken benchmark, local MoE serving design, hardware requirements, installation flow, and measured performance results.

Aug 25, 2026performance
Read more →
FreeToken claude code: Setup Guide for Model Routing

Configure FreeToken claude code with a local gateway, model routing, fallback providers, and safe troubleshooting practices.

Aug 25, 2026agents
Read more →
FreeToken codex: Setup Guide, Benchmarks & Limits

FreeToken codex explained: learn how its expert-aware memory system works, review benchmarks, hardware limits, setup considerations, and local AI tradeoffs.

Aug 25, 2026agents
Read more →
FreeToken deepseek: Local AI Setup Guide & Benchmarks

Learn how FreeToken serves DeepSeek locally with MoE caching, CPU-GPU offload, hardware guidance, setup steps, and performance notes.

Aug 25, 2026models
Read more →
FreeToken deepseek v4 flash: Local Setup Guide 2026

Set up DeepSeek V4 Flash with FreeToken for local AI inference, including RAM planning, GPU expectations, Open WebUI access, and troubleshooting.

Aug 25, 2026models
Read more →
FreeToken edge native moe serving: Architecture Guide

Learn how FreeToken uses bandwidth-adaptive execution, caching, and CPU-GPU coordination to serve large MoE models on edge hardware.

Aug 25, 2026architecture
Read more →
Install

FreeToken Install Guide

Set up FreeToken with the desktop app, uv, PyPI, or source code, then launch your first local model server. NVIDIA GPU users can install the accelerated Python package, check system bandwidth, and start an OpenAI-compatible server with the ft CLI.

1

Check the host system

Use a Windows or Linux PC with a supported NVIDIA GPU. FreeToken targets RTX 30, RTX 40, and RTX 50 series systems and can combine GPU memory with system RAM for large-model execution.

2

Install the desktop app

For the simplest setup, download and install the FreeToken desktop application for Windows or Linux. The desktop workflow provides a graphical path for configuring and serving local models.

3

Install with uv

Install the accelerated FreeToken package into your Python environment with uv.

uv pip install "freetoken[accel]"
4

Install from PyPI

The same accelerated package can be installed through a standard Python package workflow when uv is not being used.

pip install "freetoken[accel]"
5

Install from source

Clone the FreeToken repository when you need the development version or want to work directly with the source tree, then install the project and its acceleration dependencies.

git clone https://github.com/FlashML-org/FreeToken.git
6

Measure host bandwidth

Run the built-in bandwidth benchmark before serving very large MoE checkpoints. FreeToken uses the result to help determine how effectively CPU memory and the GPU can work together.

ft bench bw
7

Start the model server

Launch the FreeToken serving workflow with ft serve. The server exposes local inference through OpenAI- and Anthropic-compatible interfaces for applications and coding agents.

ft serve
8

Verify the first run

Confirm that the selected model loads, the server remains active, and a local API request can return generated output before connecting external clients such as Codex, Claude Code, or OpenCode.

Quick Tips

  • The freetoken[accel] extra installs the GPU acceleration stack needed for RTX-class hardware.
  • Run ft bench bw once before your first large model so FreeToken can pick an efficient execution strategy.
  • The desktop app and the ft CLI serve the same engine, so you can switch between them freely.
  • Keep CUDA and NVIDIA drivers up to date to avoid kernel and compatibility errors inside the accelerated package.
Models

FreeToken Supported Models

FreeToken is designed around large sparse Mixture-of-Experts models whose total parameter counts can exceed the memory of a single consumer GPU. Its execution backends combine GPU compute, system memory, expert caching, and model-specific checkpoint handling to make these models practical on local machines.

ModelScaleCheckpoint FormatQuantizationBackendBest For
GLM-5.2753B total parametersGLM MoE checkpointMXFP4 / NVFP4 / FP8 / BF16Hybrid MoE / expert offloadHigh-capacity workstation serving
DeepSeek-V4-FlashLarge sparse MoEDeepSeek MoE checkpointMXFP4 / NVFP4 / FP8 / BF16Hybrid MoE / expert offloadLocal reasoning and agent workloads
Qwen3.635B-A3B-class sparse MoEQwen MoE checkpointMXFP4 / NVFP4 / FP8 / BF16MoE serving backendLocal chat, coding, and agents
Qwen3 MoEMultiple sparse sizesQwen MoE checkpointsMXFP4 / NVFP4 / FP8 / BF16MoE serving backendGeneral-purpose local inference
gpt-ossMultiple checkpoint sizesgpt-oss checkpointMXFP4 / NVFP4 / FP8 / BF16Sparse-model serving backendOpenAI-compatible app workflows
Gemma-4Model-dependentGemma checkpointMXFP4 / NVFP4 / FP8 / BF16Model-appropriate backendLocal general-purpose inference
MiniMaxLarge sparse MoEMiniMax checkpointMXFP4 / NVFP4 / FP8 / BF16Hybrid MoE / expert offloadLarge-context agent workloads

Quick Tips

  • All major families support MXFP4 and NVFP4 low-precision checkpoints alongside FP8 and BF16.
  • Pick quantization by matching checkpoint size against your combined VRAM plus system RAM budget.
  • Hybrid MoE execution suits the largest checkpoints; smaller sparse models can stay more GPU-resident.
  • New model families are added regularly, so check the models documentation with each engine release.
Hardware

FreeToken Hardware Requirements

FreeToken does not require the complete model to fit inside GPU VRAM. Its bandwidth-adaptive CPU-GPU execution, expert caching, and elastic memory management allow sparse MoE checkpoints to use both GPU memory and host RAM, making system memory and PCIe bandwidth important alongside raw GPU capacity.

ComponentRequirementWhy It Matters
Operating systemWindows or LinuxBoth platforms are supported by the desktop app and local serving workflow.
GPU familyNVIDIA RTX 30 / 40 / 50 seriesNewer GPUs provide more VRAM and compute headroom for larger active expert sets.
GPU VRAMModel, precision, and cache dependentMore VRAM keeps active weights and cached experts on the GPU, reducing transfers.
System RAMHigh capacity for large MoE checkpointsHost RAM holds weights that exceed VRAM, enabling far larger checkpoints.
CPU memory bandwidthHigher is better for offloadBandwidth-adaptive execution uses measured host-memory performance.
PCIe bandwidthFast transfers recommendedDetermines how quickly expert weights move between RAM and VRAM.
CUDA and driversCompatible NVIDIA runtimeRequired for accelerated GPU execution through freetoken[accel].
290B+ class modelsSubstantial host RAM plus RTX GPUHybrid CPU-GPU execution replaces the need to fit in VRAM.
Largest checkpointsWorkstation RAM, storage, bandwidthServes frontier-scale sparse models such as 753B-parameter GLM-5.2.

Quick Tips

  • RTX 30 series works well with mid-size checkpoints and heavier quantization.
  • RTX 40 series offers stronger inference performance and memory efficiency for hybrid MoE workloads.
  • RTX 50 series suits NVFP4 low-precision formats and larger high-throughput deployments.
  • Dual-channel or better system memory directly improves offloaded expert delivery speed.
Benchmarks

FreeToken Benchmarks

FreeToken performance depends on more than GPU compute because large MoE checkpoints can continuously move experts between host memory and the GPU. Its benchmark tooling measures the bandwidth available on a specific machine and helps the runtime choose between heavier offload and more GPU-resident hybrid execution.

BenchmarkCommandMeasuresWhy It Matters
CPU memory bandwidthft bench bwHost-memory throughputOffloaded expert weights are read from system RAM, so bandwidth directly improves expert delivery.
CPU-to-GPU bandwidthft bench bwPCIe transfer throughputDetermines how fast weights reach the GPU when not already in VRAM.
Decode throughputft serveGenerated tokens per secondInteractive generation speed after prompt processing.
Prefill throughputft servePrompt-processing throughputDouble-buffered prefill streaming overlaps movement and computation for long prompts.
GPU memory useft serveVRAM for weights, caches, stateElastic VRAM management keeps the model resident within the memory budget.
System memory useft serveRAM for checkpoint and expertsLarge host capacity serves checkpoints far larger than GPU VRAM.
Expert cache efficiencyft serveGPU-resident expert reuseThe global LRU cache reduces repeated CPU-to-GPU transfers.
Offload executionft bench bwHost-resident expert shareWorks best with strong CPU-memory and PCIe bandwidth even with limited VRAM.
Hybrid MoE executionft bench bwGPU + host expert mixBalances VRAM capacity, cache locality, and transfer bandwidth.
Consumer laptop profileft bench bwMobile bandwidth limitsHelps pick smaller or more quantized checkpoints for laptop memory.
Gaming PC profileft bench bwDesktop RTX + RAM + PCIeThe common FreeToken configuration on RTX 30, 40, and 50 systems.
Workstation profileft bench bwHigh-capacity everythingSupports frontier-scale sparse checkpoints with CPU-GPU cooperation.

Quick Tips

  • Run ft bench bw on every new machine before choosing an execution strategy.
  • Decode speed follows host bandwidth closely when most experts are offloaded.
  • Hybrid execution usually beats pure offload once enough experts fit in VRAM.
  • Compare prefill and decode together, since long-context workloads stress both.
Local API

FreeToken API Guide

FreeToken exposes local model inference through familiar API formats. After loading a supported model with ft serve, applications can discover models, send chat or Responses API requests, use Anthropic-style messages, and stream generated tokens from the local server.

Launch the FreeToken serving process with a supported model. The server handles model loading, CPU-GPU execution, memory allocation, and HTTP API requests.

ft serve <model>

OpenAI Chat Completions

OpenAI-compatible

Use the OpenAI-compatible chat completions endpoint with existing SDKs or HTTP clients.

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "<model>",
    "messages": [
      {"role": "user", "content": "Explain mixture-of-experts models."}
    ]
  }'

OpenAI Responses API

OpenAI-compatible

Applications built around the newer OpenAI Responses API can send requests to the corresponding local endpoint.

curl http://localhost:8000/v1/responses \
  -H "Content-Type: application/json" \
  -d '{
    "model": "<model>",
    "input": "Write a short Python function that reverses a list."
  }'

List Available Models

OpenAI-compatible

Query the OpenAI-compatible models endpoint to see models exposed by the running FreeToken server.

curl http://localhost:8000/v1/models

Anthropic Messages API

Anthropic-compatible

FreeToken also provides an Anthropic-compatible messages interface for clients and coding tools designed around the Anthropic API.

curl http://localhost:8000/v1/messages \
  -H "Content-Type: application/json" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "<model>",
    "max_tokens": 512,
    "messages": [
      {"role": "user", "content": "Summarize this repository architecture."}
    ]
  }'

Streaming Responses

SSE

Enable streaming when the client supports server-sent incremental output so generated tokens can be consumed as they arrive.

{
  "model": "<model>",
  "messages": [
    {"role": "user", "content": "Write a REST API example."}
  ],
  "stream": true
}

Use ft serve configuration to select the model and control how the local HTTP service is exposed, including host, port, and memory settings.

ft serve --help

Inspect CLI Options

CLI

FreeToken keeps current serving and runtime options available directly from command-line help, including options added by newer engine releases.

ft --help
ft serve --help
Coding Agents

FreeToken Claude Code and Codex Setup

FreeToken can act as the local inference backend for agentic coding tools that understand OpenAI- or Anthropic-compatible APIs. The ft launch workflow helps configure supported agents, while dry-run mode lets you inspect the planned setup before changing or starting anything.

1

Install and verify FreeToken

Make sure the FreeToken CLI is installed and a supported model can be loaded locally before connecting an external coding agent.

ft --help
2

Check agent launch options

Use the launch command help to view the currently supported agent integrations and their available configuration options.

ft launch --help
3

Preview with dry run

Use dry-run mode before launching an agent when you want to inspect the generated command and environment configuration without starting the final workflow.

ft launch <agent> --dry-run
4

Launch the coding agent

Start the selected coding agent through FreeToken so its API configuration points to the local inference backend.

ft launch <agent>
5

Claude Code

Claude Code can use FreeToken's Anthropic-compatible interface. The local configuration redirects supported Anthropic message traffic to the FreeToken server while preserving the agent workflow.

ft launch claude
6

Codex

Codex-style clients can connect through FreeToken's OpenAI-compatible endpoints, allowing a locally served model to handle agent requests.

ft launch codex
7

OpenCode, Hermes and OpenClaw

FreeToken's launch workflow also targets additional coding and autonomous-agent clients such as OpenCode, Hermes, and OpenClaw. Check the launch help for current integration names and options.

ft launch --help
8

Prevent cloud API fallback

Verify the agent's provider and base URL configuration before beginning a long task. Remove or disable unwanted cloud-provider configuration when the goal is fully local inference, and use dry-run mode to inspect the environment FreeToken will pass to the agent.

ft launch <agent> --dry-run

Quick Tips

  • Always run the dry-run preview once before the first real launch of a new agent.
  • Keep the FreeToken server running while agents work, since long sessions reconnect to the same endpoint.
  • Check the provider base URL after setup to confirm requests reach the local server, not a hosted API.
  • Integration names evolve with engine releases, so re-check ft launch --help after upgrading.
How It Works

FreeToken Architecture

FreeToken is designed around the sparse structure of mixture-of-experts models. Instead of requiring every expert weight to remain in GPU memory, it coordinates CPU and GPU execution, caches frequently used experts, overlaps transfers with computation, and adapts memory use to the machine running the model.

FreeToken decides how expert computation should be divided between CPU and GPU resources instead of treating GPU VRAM as the only usable model memory. Its execution strategy accounts for available CPU-GPU bandwidth and hardware characteristics, balancing computation and data movement across host memory and GPU compute.
Release Tracker

FreeToken Updates and Releases

FreeToken is evolving across both its inference engine and desktop experience. The release tracker focuses on changes that directly affect installation, model compatibility, local serving, GPU support, performance, and agent workflows.

Published FreeToken releases are the main place to review packaged engine changes and release notes. Check each release when upgrading, especially when it changes model support, checkpoint formats, serving behavior, installation requirements, or runtime performance.

Follow FreeToken Development

Track new releases and join the community through the official channels: