# Infervisor model library

> Open models compiled by plow for specific hardware and served by plowrt: pull with one command, serve offline through an OpenAI-compatible API. Every model points at its Hugging Face checkpoint; the registry carries the compiled program, not the weights.

Registry: https://pdist.infervisor.ai (public, read-only; each model's index is signed with `infervisor-2026:ESz+iXQToovy4GhRUgTHoUKUwMumUA/ls/4tAaEjL3M=`, which plowrt trusts by default).
Install plowrt 0.1.3+d371228: `curl -fsSL https://pdist.infervisor.ai/install.sh | sh`
Run a model: `plowrt load <name> --fetch-weights` (the registry carries the compiled program; the checkpoint comes from Hugging Face, with `HF_TOKEN` set for a gated one) then `plowrt serve --model <name> --port 8080`; the API is OpenAI-compatible at http://localhost:8080/v1 and the model id is `<name>`.
A build is selected for the machine it runs on: NVIDIA and AMD builds need the exact SM/CU count (an H100 build runs on an H200), Apple builds run on any chip of the family.

## Models

- [Gemma 4 31B](https://infervisor.ai/models/gemma-4-31b): text generation; The largest dense Gemma 4, instruction tuned, in bf16 and FP8 on NVIDIA. Builds: NVIDIA H100 SXM5. Native context: 262144 tokens ([source](https://huggingface.co/google/gemma-4-31B-it/blob/main/config.json)). Checkpoint google/gemma-4-31B-it. fp8 builds serve RedHatAI/gemma-4-31B-it-FP8-dynamic. Gated checkpoint (HF_TOKEN). `plowrt load gemma-4-31b --fetch-weights`
- [Gemma 4 E4B](https://infervisor.ai/models/gemma-4-e4b): text generation; Google's small instruction-tuned Gemma 4, the text model of the voice-agent stack. Builds: NVIDIA H100 SXM5. Native context: 131072 tokens ([source](https://huggingface.co/google/gemma-4-E4B-it/blob/main/config.json)). Checkpoint google/gemma-4-E4B-it. Gated checkpoint (HF_TOKEN). `plowrt load gemma-4-e4b --fetch-weights`
- [Orpheus 3B](https://infervisor.ai/models/orpheus): text to speech; Canopy Labs' Orpheus 3B, a Llama-based codec LM with a SNAC decoder, served with the same codec-LM contract as Veena. Builds: NVIDIA H100 SXM5. Native context: 131072 tokens ([source](https://huggingface.co/unsloth/orpheus-3b-0.1-ft/blob/main/config.json)). Checkpoint unsloth/orpheus-3b-0.1-ft. `plowrt load orpheus --fetch-weights`
- [Qwen3.8 27B](https://infervisor.ai/models/qwen3.8-27b): text generation; Qwen3.8-27B with gated DeltaNet recurrent layers, in bf16 and FP8 on NVIDIA. Builds: NVIDIA H100 SXM5. Native context: 262144 tokens ([source](https://huggingface.co/Qwen/Qwen3.8-27B/blob/main/config.json)). Checkpoint Qwen/Qwen3.8-27B. `plowrt load qwen3.8-27b --fetch-weights`
- [Qwen3 0.6B](https://infervisor.ai/models/qwen3-0.6b): text generation; Qwen's smallest Qwen3, published for Apple Silicon. Builds: Apple M4 Pro. Native context: 32768 tokens ([source](https://huggingface.co/Qwen/Qwen3-0.6B)). Checkpoint Qwen/Qwen3-0.6B. `plowrt load qwen3-0.6b --fetch-weights`
- [Qwen3 ASR 0.6B](https://infervisor.ai/models/qwen3-asr-0.6b): speech to text; The 0.6B Qwen3 ASR model, for latency-first transcription on a small GPU. Builds: NVIDIA H100 SXM5. Native context: 65536 tokens ([source](https://huggingface.co/Qwen/Qwen3-ASR-0.6B/blob/main/config.json)). Checkpoint Qwen/Qwen3-ASR-0.6B. `plowrt load qwen3-asr-0.6b --fetch-weights`
- [Veena](https://infervisor.ai/models/veena): text to speech; Maya Research's Veena, a Llama-style codec language model with a SNAC decoder for Hindi and English speech. Builds: NVIDIA H100 SXM5. Native context: 131072 tokens ([source](https://huggingface.co/maya-research/Veena/blob/main/config.json)). Checkpoint maya-research/Veena. `plowrt load veena --fetch-weights`

## Benchmarks and guides

- [Benchmarks](https://infervisor.ai/benchmarks): selected plow vs vLLM comparisons, including wins and losses, with versions, settings and source reports.
- [Engine comparisons](https://infervisor.ai/compare): plow vs vLLM, and SGLang vs vLLM vs plow, with selected measured comparisons.
- [GPU guides](https://infervisor.ai/gpus): vendor specs and measured results for H100, MI300X, MI325X, MI350X and MI355X.

## Machine-readable

- [All models as JSON](https://infervisor.ai/models.json): schema `infervisor.models.v1`, one object per model with builds, hardware, commands.
- Per model: https://infervisor.ai/models/{name}.json
- Registry catalog: https://pdist.infervisor.ai/catalog.json; per model: https://pdist.infervisor.ai/v1/infervisor/<name>/index.json
- [How distribution works](https://github.com/infervisor/plow/blob/main/docs/DISTRIBUTION.md)