AtomicChat / d1-omni-600M-GGUF

d1-omni-600M GGUF

Built from Liquid AI's published weights with our own importance matrix, and measured on decisions.

d1-omni-600M is Liquid AI's 587M-parameter decision model. It takes a state (text or JSON, with images or up to 30 seconds of speech) and answers typed questions in one forward pass: yes or no, a pick from named options, or a score. Nothing is generated: the answer is read from the model's scores for the options.

No released llama.cpp runs these files yet.

• d1-omni is a bidirectional LFM2.5 encoder with its own decision head ( lfm2.decision.type = d1omni ).
• Its vision and audio projectors ( d1omni_v , d1omni_a ) are not in llama.cpp either.
• Liquid's upstream PR ggml-org/llama.cpp#30110 adds the other d1 model, d1-3B, but not d1-omni.

These files have exactly the layout and metadata of the GGUFs Liquid published on 6 October and took down on 7 October, so they are ready for the runtime those files were made for. Until it ships, run d1-omni with Liquid's PyTorch code .

Pick a file

Every number below is measured on one machine. The raw results and logs are in the metrics repo .

• same answer : across 1,285 decisions, how often the file gives the same answer as the original weights published in LiquidAI/d1-omni-600M (stored in FP32). The questions are yes/no, choice and score questions over held-out texts in 30 languages and source code. The number of changed answers is in brackets.
• option drift : the mean total variation distance between the file's option probabilities and the original's. 0 means identical.
• option KL : the mean KL divergence between the original's option probabilities and the file's.

Which one to use. d1-omni is small, and quantization moves its answers more than it moves d1-3B's.

• Q8_0 already changes 1.5% of answers, and 4 bits change 7.5%.
• The largest shift in a single option probability is 0.08 at Q8_0 and 0.38 at 4 bits.
• We recommend Q8_0 , or AD-Q6_K if 60 MB matter.
• Below 6 bits, check the files on your own questions first.

AD- means Atomic Dynamic: the type is chosen per tensor instead of taken from a llama.cpp preset.

• The decision head (the two transformer blocks after the trunk) and the option scorer stay at Q8_0 in every file. They are small, about 26M parameters, and every answer passes through them.
• The token table is a lookup, which an importance matrix does not cover, so it sits above the rest: Q8_0 in AD-Q6_K , Q6_K in AD-Q5_K_M and Q5_K in AD-Q4_K_M .
• Attention stays at Q8_0, or Q6_K in AD-Q4_K_M .
• The first and last two trunk blocks take one step more than the middle: ffn_down in AD-Q6_K and AD-Q5_K_M , and the whole feed-forward in AD-Q4_K_M , where the short convolutions also stay at Q5_K.
• At about the same size, llama.cpp's stock Q4_K_M (247 MB) changes 133 answers with an option KL of 0.0255. AD-Q4_K_M changes 97, with 44% less KL.

The vision and audio projector comes as mmproj-d1-omni-600M-BF16 and mmproj-d1-omni-600M-Q8_0 . We did not measure image or audio decisions.

How these compare to Liquid's own GGUFs

Liquid published BF16, F16 and Q8_0 GGUFs of d1-omni-600M on 6 October and took the repo down on 7 October, together with the d1-3B GGUFs. Unlike d1-3B, whose weights Liquid updated on 7 October, d1-omni's weights did not change:

• Our BF16 file and both of our projector files are byte-for-byte identical to Liquid's: they have the same SHA-256.
  • We compared against TechnoBaptist/d1-omni-600M-GGUF , a Hub duplicate of Liquid's repo made at 17:11 UTC on 7 October, before the takedown.
  • Its file sizes match the listing of Liquid's own repo that we took two minutes later.
• Liquid did not publish files below 8 bits. The Liquid Q8_0 row below is that duplicate's Q8_0.

The two Q8_0 files are the same quantization written by two tools, and they measure the same.

How we made and measured them

• Weights: LiquidAI/d1-omni-600M at revision 414f8d6 . There is no converter for d1-omni in llama.cpp, so each file is written in the layout of Liquid's GGUF: the same tensor names, order and metadata, including the decision type, the temperatures and the systemone template.
  • Every tensor was matched against the checkpoint before writing.
  • The audio encoder's batch norms are folded into a scale and a shift, as llama.cpp's converters do.
• Quantization: llama.cpp's llama-quantize at commit 18b5f8b . The projector's Q8_0 is written the way Liquid wrote theirs.
• Importance matrix: computed in PyTorch, since llama.cpp cannot run the model, and saved in llama.cpp's format.
  • Inputs were collected at the 92 linear layers of the trunk.
  • The prompts were 2,378 decision prompts, about 1.5M tokens. Their states come from the calibration corpora pool: Wikipedia in 30 languages, code and structured files. Each prompt carries yes/no, choice and score questions.
• Decisions: 1,285 questions over the held-out calib-corpora eval/neutral and eval/code texts, none of them in the calibration set.
  • Each file is read back, dequantized, into Liquid's own PyTorch model, which answers in FP32 on an NVIDIA A10.
  • This measures the error the quantized weights add. A llama.cpp runtime will add its own small arithmetic differences on top.

Model details

• Base: LiquidAI/d1-omni-600M , a decision model post-trained from LFM2.5-Encoder-350M .
• Size: 587M parameters in all: a 381M trunk and decision head, a 94M SigLIP2 vision encoder and a 112M FastConformer audio encoder.
• Context and vocabulary: 16,384 positions for text, image and audio together, and a 65,536-token vocabulary.
• License: LFM Open License v1.0; see LICENSE .