inductiveML/LFM2.5-230M-MONARCH

LFM2.5-230M — MONARCH WebGPU layout

Preconverted Liquid AI LFM2.5-230M weights for Inductive ML's MONARCH MIX_M40 browser decoder. No training or fine-tuning was performed. This is a custom packed GPU buffer, not a Transformers checkpoint.

• Run the browser demo
• Readable kernel and runtime source
• Experiment, methodology, and recorded measurements

Provenance and format

Upstream: LiquidAI/LFM2.5-230M-ONNX, revision c6f46e4e3f885ebcad164d14059a49f90e27eb4d.

The published buffer contains packed int4 weights with block size 32 and f32 scales, normalization and convolution tensors, exported rotary tables, and additional slice-major copies of the feed-forward down projections. The decoder also prepares selected layouts on the GPU. monarch-config.json lists all 199 buffer sections, their offsets, sizes, model shape, and selected kernel configuration.

The original 145,054,720-byte buffer is preserved as a prefix. Fourteen down-projection repacks add 22,937,600 bytes. Two extended rotary tables add 655,360 bytes. The resulting buffer is 168,647,680 bytes. Nine transport chunks reconstruct it byte for byte; the chunking does not alter values. The rotary tables came from the ONNX export rather than being recomputed with a different math implementation.

Canonical SHA-256:

ca1fea89fd9f3ca7e5d6d5720705c96f457a92edd0cfc811cf8d9391ac784cb0

Download and reconstruct

The demo pins a repository commit and verifies each chunk and the complete buffer before inference. To reconstruct the same buffer locally, install huggingface_hub and use:

import hashlib, json
from pathlib import Path
from huggingface_hub import snapshot_download

folder = Path(snapshot_download("inductiveML/LFM2.5-230M-MONARCH"))
manifest = json.loads((folder / "model-manifest.json").read_text())
chunks = []
for item in manifest["chunks"]:
    data = (folder / item["path"]).read_bytes()
    assert len(data) == item["bytes"]
    assert hashlib.sha256(data).hexdigest() == item["sha256"]
    chunks.append(data)
weights = b"".join(chunks)
assert len(weights) == manifest["weightBytes"]
assert hashlib.sha256(weights).hexdigest() == manifest["weightSha256"]
Path("weights.bin").write_bytes(weights)

For repeatable deployments, pass revision= with the commit pinned in the Space's model-source.json.

Scope and measurements

The working decoder uses 72 ordinary dispatches per token. It supports greedy single-stream generation in the demo. No persistent cross-workgroup decoder is included. The demo checks a known 32-token continuation before enabling inference and requires WebGPU, shader-f16, fixed 32-lane subgroups, and 1,024-thread workgroups.

The 1,574.4 tok/s figure is a recorded M4 Max / fullscreen Chromium result at context 192, not a prediction for arbitrary hardware. The research kernel retained NO_VERDICT_SPLIT; later persistent-kernel pricing was not formally adjudicated. See the experiment page for complete conditions and limitations.

License and modifications

Weights and tokenizer data retain Liquid AI's LFM Open License v1.0, including its commercial-use conditions. See LICENSE-LFM.txt and NOTICE.txt. All files under weights/ are modified representations prepared by Inductive ML; their associated manifest documents the changes and checksums. The MONARCH runtime is separately released under Apache-2.0.