📜 Historical reference — This document references the closed-source FastFlowLM runtime which has since been fully reverse-engineered and replaced. See The story for the current state.

Lemonade Compatibility

Status: Active · targeting Lemonade v10.9.0+

1bit.MONSTER NPU and GPU engines are compatible with Lemonade via the OpenAI-compatible /v1/chat/completions endpoint.

How it works

1bit-halo-server (port :8180 — NPU engine) and the ZINC GPU engine both expose the standard OpenAI chat completions API. Lemonade can route omni-modal requests to either backend as a sub-process or remote server.

# lemonade config.yaml
models:
  - name: qwen3-npu
    server:
      url: http://127.0.0.1:9090  # FLM proxy or 1bit-halo-server
      type: openai

Quick start

# 1. Start the 1bit NPU daemon
./npu_engine_all model.q4nx 16

# 2. Lemond auto-discovers or point manually
lemond launch --model qwen3-npu

v10.9.0+ integration

• *_bin config keys — Lemonade v10.3+ (PR #1713) accepts builtin / latest / version tag / local path values for every backend binary. The 1bit-monster packaging can supply ryzenai.server_bin as latest or a pinned tag.
• First-party contribution — bong-water-water-bong is a listed contributor in v10.9.0 (test/documentation: PR #2447).
• Omni-modal routing — No-code integration: point Lemonade's omni-modal UI at the 1bit OpenAI-compatible endpoint. Zero config changes on the 1bit.MONSTER side.

Version compatibility table

Updating

When a new Lemonade release ships:

1. Check the Lemonade releases for breaking API changes
2. Smoke-test with 1bit-halo-server on port :9090
3. Update this doc and the site wiki

MCP support

Both projects expose MCP servers. See:

• Lemonade MCP docs
• 1bit.MONSTER packaging docs