# Models

See which model configurations PLLM can describe and what model support does not yet include.

[View canonical HTML](https://pllm.run/sdk/build/models/)

Document ID: `pllm.docs.models`  
Release: `0.1.0`

`pllm.lower_model(config, batch=..., max_input_tokens=..., max_new_tokens=...)`
turns model configuration and workload limits into a `ModelPlan`. It does not
download weights or tokenizers, choose a device, or start a runtime.

Each adapter validates a specific architecture and produces the same model-neutral
plan format. The plan records operators, layers, weight references, persistent
state, workload limits, and source identity. When a new model needs a new
operation, PLLM extends the shared vocabulary instead of adding model-specific
logic to the compiler.

| Adapter | Validated configuration scope | Added semantics |
| --- | --- | --- |
| `pllm.qwen2.v1` | dense Qwen2 configuration contract | grouped-query attention, rotary position, gated SiLU MLP, KV state |
| `pllm.qwen3.v1` | dense Qwen3; pinned mini-coder-4b fixture | explicit head width, Q/K normalization, bias-free attention |
| `pllm.qwen3_5_text.v1` | pinned Qwen3.5-4B text decoder | hybrid full attention and gated-delta recurrence, convolution and recurrent state |
| `pllm.phi4_mini.v1` | pinned Phi-4-mini-instruct decoder | fused QKV, fused gate/up, partial LongRoPE |
| `pllm.gemma4_e2b_text.v1` | pinned Gemma 4 E2B text decoder | hybrid attention, shared KV suffix, double-wide suffix MLP |
| `pllm.gemma4_e4b_text.v1` | pinned Gemma 4 E4B text decoder | hybrid attention, shared KV suffix, PLE, softcap |

Describing a model is not the same as loading its checkpoint or running it
privately. Compiler support, output quality, benchmarks, and deployment are also
separate. Check [current support](/sdk/reference/status/) before choosing a model.

## Python SDK example

```python
from pathlib import Path

from pllm import lower_model

fixture = Path("crates/pllm-models/tests/fixtures/mini-coder-4b-c87892d-config.json")
plan = lower_model(fixture.read_bytes(), batch=1, max_input_tokens=16, max_new_tokens=4)
operators = {operation["operator"] for operation in plan.prefill["operations"]}
assert {"rotary_embedding", "silu", "kv_cache_append"} <= operators
```

API: [Python SDK objects and signatures](/sdk/reference/python/pllm/#objects-and-signatures)

Run this example from a PLLM source checkout; package installations do not include
the checked-in fixture. This creates a semantic plan only.