Run every model,
on every accelerator.
FenixCOS is a compute operating system for LLM inference. Its model matrix ships production-ready engine images for every supported model × accelerator pair — each one generated, benchmarked and correctness-gated by MetaInfer, the open optimization platform.
Every supported model, one docker pull away.
The FenixCOS model matrix tracks every optimized engine image the platform has shipped — per model, per accelerator, each with a full release history and benchmarks.
Describe the deployment. Get the engine.
Every image in the matrix is a product of MetaInfer — an open-source, LLM-powered platform that writes, benchmarks and correctness-gates purpose-built inference engines. Everything below shows how it works.
General-purpose frameworks carry everyone's baggage.
An inference framework must support every model, every hardware, every parallel and quantization scheme. Once your deployment constraints are fixed, most of that code is unnecessary — and it still sits on the hot path.
Automatic Model-Hardware Co-Adaptation for Heterogeneous AI Accelerators
MetaInfer treats inference adaptation as route search over a costed execution-adaptation graph — connecting model semantics, framework dispatch, kernel options, hardware capabilities, runtime evidence and serving goals.
What the paper shows
@misc{chen2026metainfer,
title = {Automatic Model-Hardware Co-Adaptation for Heterogeneous AI Accelerators},
author = {Chen, Tian and Mi, Mingheng and Wang, Pu and Hu, Chen and Zhang, Hai},
year = {2026},
eprint = {2607.12875},
archivePrefix = {arXiv},
primaryClass = {cs.MA}
}
One toolkit, four ways to make inference cheaper.
Task-driven: pick a task type in the web UI, describe your target, hit run. Each task is a self-contained pipeline with its own oracle.
Compose a task. Watch it run.
A miniature of the MetaInfer web UI. Pick a target and press generate — the plan, logs and benchmark below are simulated in your browser.
Demo content on this page is based on our internal closed-source edition and may differ from the open-source release. For the closed-source edition, please contact us for business cooperation.
Build plan
Execution log
Result
From description to production in four steps.
Every run leaves a trace.
MetaInfer records what its agents actually did — every tool call, benchmark and verdict — and reconstructs them into session → trace → observation timelines you can browse. Below: a miniature explorer with simulated data.
Playbooks mined from real optimization runs.
MetaInfer's kernel-optimization methodology is not hand-written lore. It is mined from thousands of real agent trajectories, then distilled into machine-readable skills and a decision state machine. All figures below are aggregated and anonymized — numbers only, no content.
The observation deck
Every agent session is rebuilt into a Langfuse-style tracing database: sessions → traces → observations, with full-text search and deduplicated storage.
The optimization mine
A scoring pipeline filters trajectories that are actually about inference-performance work, normalizes every action, and pairs each command with its outcome.
…distilled into a skill library
The mine compresses into 7 subtask playbooks plus a rare-skills volume, each backed by a decision graph — and a global state machine that an orchestration agent can execute directly.
What the corpus talks about
Source: private deployment of MetaInfer agents, Jun–Sep 2026. Aggregated counts only; workspaces, prompts and outputs are never shown. Best grader-verified single-kernel speedup observed so far: 15.1×.
One orchestrator, a row of GPU boxes.
Each machine runs a worker daemon that registers itself, heartbeats every 15 s and claims jobs from an NFS-safe inbox. The web UI shows a live GPU slot scoreboard — and yes, you can force-release a stuck slot.
Up and running in four commands.
Requires Python 3.10+ and the Claude Code CLI backend. Optional worker daemons join the cluster with one command; a shared NFS mount is all they need.