onnx-community/d1-3B-ONNX

d1-3B (ONNX)

LiquidAI/d1-3B converted to ONNX for Transformers.js and the browser (WebGPU). d1 is a decision model on LFM2.5-VL-3B: give it a state (text and/or images) and typed questions (choice, noul yes/no, score), and it answers in one forward pass with zero generated tokens. The answer is a softmax over each option's token at the answer slot.

This is a straight conversion: the same weights in LiquidAI/LFM2.5-VL-3B-ONNX's graphs (same architecture), with one change: the decoder returns logits for the last position only, which is all d1 reads (and all generation needs), so a 300-token prompt doesn't materialise 300 × 128k logits.

Files

Graph	q8 (default)	q4	fp32
decoder_model_merged	3.15 GB	1.76 GB	– (rebuild with conversion/build.py)
vision_encoder	0.50 GB	0.28 GB	1.71 GB
embed_tokens	0.17 GB (4-bit Gather)	0.17 GB	1.05 GB

Block-32 MatMulNBits (8-bit or 4-bit), fp32 activations. External data is split into ≤ 1 GB chunks.

Parity with d1's own runtime

Option probabilities vs the model repo's system_one (PyTorch, fp32), on 8 text decisions (choice, noul, score) and 3 image decisions:

Variant	Text max |Δ|	Image max |Δ|	Flipped answers
fp32	0.0000	0.0000	0
q8	0.018	0.002	0
q4	0.028	0.063	1 (a 0.48 vs 0.45 near-tie)

Transformers.js' own LFM2-VL image processor tiles slightly differently from the Python one (256 vs 234 image tokens on a 640×480 picture), which moves image answers by up to 0.04 at q8; same picks.

Use it

With open-jev (text states), which implements d1's prompt and readout:

import { OpenJev, choice } from "open-jev";

const jev = await OpenJev.load({ model: "d1-3b", device: "webgpu" }); // onnx-community/d1-3B-ONNX, q8
const { route } = await jev.decide("Message: What's the Wi-Fi password?\nMemory 1 (strong match, note): Home Wi-Fi: HomeNet-5G, password sunflower42", {
  route: choice("Can a memory answer this message?", ["memory", "large model"], {
    memory: "One memory directly answers the message.",
    "large model": "No memory answers it.",
  }),
});
// { type: "choice", choice: "memory", confidence: 0.94, ... }

With Transformers.js directly (text or images): build the prompt as d1's prompt.py does, run one forward pass, and softmax the option tokens:

import { AutoProcessor, AutoModelForImageTextToText, RawImage } from "@huggingface/transformers";

const id = "onnx-community/d1-3B-ONNX";
const processor = await AutoProcessor.from_pretrained(id);
const model = await AutoModelForImageTextToText.from_pretrained(id, { dtype: "q8", device: "webgpu" });

const prompt = "<|startoftext|><|im_start|>user\n<image>Is this a parking sign?\n\nReply with yes or no only.<|im_end|>\n<|im_start|>assistant\n";
const inputs = await processor(await RawImage.read("sign.png"), prompt, { add_special_tokens: false });
const { logits } = await model(inputs); // [1, 1, vocab]: the answer slot
// softmax over the token ids of "yes"/"Yes"/"YES" vs "no"/"No"/"NO" (best form of each)

Conversion

conversion/ has everything: build.py maps d1's safetensors onto the LFM2.5-VL-3B-ONNX graphs by name (transposing MatMul weights, the SigLIP2 position table to a 16×16 grid), computes the RoPE caches, slices the decoder to the last position and unties the LM head so it can be quantized; quantize.py, rechunk.py; and the parity scripts (reference.py, onnx_check.py, vision_*.py).

License and attribution

d1-3B is © Liquid AI, Inc. and licensed under the LFM Open License v1.0 (see LICENSE), including its commercial-use threshold. This repository is a Derivative Work: the original weights were converted to ONNX, re-laid out into the LFM2.5-VL-3B ONNX graphs, the decoder was changed to return last-position logits with an untied LM head, and the weights were quantized. No weights were retrained.