d1-omni-600M Multimodal Decision Model: 4GB+ VRAM, GGUF Builds

- 1. At a Glance
- 2. Overview
- 3. Specifications
- 4. Performance
- 5. Strengths and Use Cases
- 6. How It Differs from Similar Models
- 7. Hardware Requirements
- 8. Can You Run It Locally?
- 9. Recent Models in the Same Size Class
- 10. How to Get It
- 11. Quantized and Converted Variants
- 12. Related Articles
- 13. What to Read Next
- 14. Sources
- 15. Update History
At a Glance
| Item | Value |
|---|---|
| Repository | LiquidAI/d1-omni-600M |
| Family guide | LFM2.5 guide (11 articles) |
| Publisher guide | Liquid AI: models and licenses |
| Published | 2026-10-06 |
| License | lfm1.0 |
| Formats | safetensors |
| Paper | arXiv:2511.23404 |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
LiquidAI has released “d1-omni-600M", a multimodal decision model that accepts text, image, and audio inputs and instantly makes “judgments (decisions)" without going through a generation process. Built on the “LFM2.5-Encoder-350M", this lightweight model has 587M parameters and is designed to run on edge devices.
According to the research paper “LFM2 Technical Report", this model adopts a “zero-output-token" approach, where answers are read directly from the model’s distribution rather than generating and analyzing text like traditional LLMs. This makes it possible to answer multiple questions in a single forward pass for states containing images and audio.
Source: LFM2 Technical Report
Specifications
- Total Parameters: 587M (Shared trunk and decision head 381M, vision encoder 94M, audio encoder 112M)
- Vision Encoder: SigLIP2 vision tower
- Audio Encoder: 17-layer FastConformer
- Context Length: 16,384 tokens (total positions for text, images, and audio)
- Vocabulary Size: 65,536
Performance
The scores on “Decision Index 0.2.1" released by the publishers are as follows. The comparison targets include larger 12B and 3B class models.
→ Scroll horizontally to see all columns
| Model | Size | Decision Index | Knowledge | Language | Retrieval | Tools | Arts |
|---|---|---|---|---|---|---|---|
| Winnow-12B | 12B | 50.02 | 33.8 | 56.0 | 54.0 | 71.0 | 30.0 |
| d1-3B | 3B | 48.57 | 23.8 | 56.4 | 52.8 | 74.5 | 36.3 |
| Decider 35B-A3B | 36B | 47.11 | 31.8 | 55.5 | 54.7 | 56.5 | 32.6 |
| JPT-9B | 9.7B | 46.89 | 31.7 | 56.7 | 44.6 | 67.0 | 28.6 |
| Decision 1.0 Lux | 9.7B | 43.49 | 30.9 | 48.0 | 50.0 | 57.2 | 26.4 |
| JPT-4B | 4.7B | 43.04 | 28.7 | 52.5 | 45.0 | 57.2 | 25.8 |
| Jet v6.2 | 4.7B | 42.60 | 28.7 | 43.9 | 48.2 | 62.9 | 27.0 |
| Decider 4B | 4.7B | 40.70 | 25.7 | 46.0 | 44.7 | 58.6 | 25.0 |
| Winnow-E4B | 8.0B | 39.89 | 22.3 | 45.1 | 43.8 | 62.5 | 22.8 |
| Decider 2B | 2.3B | 28.97 | 14.9 | 32.6 | 37.3 | 42.4 | 14.6 |
| d1-omni-600M | 587M | 15.95 | 8.3 | 12.9 | 35.0 | 15.1 | 6.8 |
This table shows that d1-omni-600M holds its own in specific areas despite its extremely small size of 587M. Although its overall score (Decision Index) does not reach the top models, it records 35.0 in the “Retrieval" category, showing performance close to the 2.3B class Decider 2B (37.3). On the other hand, in areas testing broad knowledge such as “Knowledge" and “Arts", the small number of parameters takes a toll, creating a large gap with top-tier models.
Next are the results of evaluating existing benchmarks as decision tasks.
→ Scroll horizontally to see all columns
| Benchmark | d1-omni-600M | d1-3B | Decider 4B | Decider 2B |
|---|---|---|---|---|
| SQuAD 2.0 | 74.0 | 85.3 | 76.0 | 67.7 |
| Civil Comments | 95.8 | 93.0 | 92.8 | 93.6 |
| MASSIVE intent | 86.1 | 87.3 | 88.3 | 81.1 |
| PubMedQA | 61.3 | 66.0 | 63.3 | 65.7 |
| BoolQ | 77.7 | 86.7 | 89.0 | 87.3 |
| XNLI | 74.7 | 85.0 | 88.6 | 85.0 |
| PAWS-X | 79.5 | 76.9 | 69.8 | 59.5 |
| Mean | 78.4 | 82.9 | 81.1 | 77.1 |
In these evaluations, d1-omni-600M outperforms top models in some categories. Notably, it records 95.8 in “Civil Comments" and 79.5 in “PAWS-X", achieving top values that exceed the 3B class d1-3B and 4B class Decider 4B. It also records a mean score of 78.4, showing efficiency that surpasses the Decider 2B (77.1), which is about four times its size. However, there is a gap of 10 points or more with top models in “BoolQ" and “XNLI", showing a clear division of strengths and weaknesses depending on the nature of the task.
Strengths and Use Cases
d1-omni-600M is a model specialized in making “judgments" rather than “generation" for input data. The capability_tags listed in the model card include decision, classification, vision, and audio, making it excel at instantly classifying multimodal information.
The following uses are recommended as specific areas of strength:
– Routing and Triage: Judges customer inquiry content and routes it to the appropriate department (billing, technical support, fraud prevention, etc.).
– Content Moderation: Determines whether text or images violate policies.
– Audio Command Analysis: Extracts speaker intent and topics from audio clips up to 30 seconds long. The audio capability is trained on interactions between English speakers and assistants.
– Visual Inspection: Checks how many specific objects exist within an image or whether there are anomalies.
– Agent Guardrails: Monitors and judges whether AI agent actions remain within safe boundaries.
The model’s greatest feature is the “zero-output-token" method, which does not involve any answer generation process. By defining types such as “yes/no (noul)", “choice", and “score" in response to questions, answers are read directly from the internal probability distribution of the model. This prevents the “output format parsing failures" that are problematic in generative AI, and expects extremely fast and deterministic operation.
Supported languages include 15 languages including Japanese (English, German, Spanish, French, Italian, Dutch, Polish, Portuguese, Arabic, Hindi, Russian, Turkish, Vietnamese, Chinese). However, please note that this model is not a “chat model" and cannot be used for free-form conversation or essay composition.
How It Differs from Similar Models
The “LFM2.5-350M-Diffusion-Exp" Block Diffusion Lightweight Language Model: 4GB+ VRAM Required previously covered on this site is based on the same LFM2.5 architecture, but its nature is contrasting. While Diffusion-Exp was an experimental version of a “generative model" aiming to parallelize and accelerate text generation, this d1-omni-600M is a “decision model" that completely eliminates generation.
Furthermore, whereas Diffusion-Exp primarily targeted text, d1-omni-600M integrates a vision encoder (SigLIP2) and an audio encoder (FastConformer), providing multimodality that allows direct input of images and audio to make decisions.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 587M parameters
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 4GB (laptop iGPU / phone class) | BF16 | 0.7GB | 0.9GB |
Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: registered, vLLM: not registered, MLX (mlx-lm): not registered. “Not registered" means the name is absent from that registry today, not that the model cannot run.
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. File sizes are measured from the converted build LiquidAI/d1-omni-600M-GGUF. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
Runs in Ollama, LM Studio and llama.cpp via a converted build.
The publisher ships safetensors, but LiquidAI/d1-omni-600M-GGUF provides a GGUF build you can use.
License — lfm1.0 (Commercial use allowed with conditions): LiquidAI’s LFM Open License v1.0. Commercial use is allowed, except by legal entities with annual revenue of $10 million or more (Section 5). Redistribution requires the license text and notice of changes.
Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Recent Models in the Same Size Class
Models with up to 4B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.
→ Scroll horizontally to see all columns
| Model | Parameters | Smallest VRAM tier | License | Our article |
|---|---|---|---|---|
| LiquidAI/d1-3B | 3.1B | 4GB | lfm1.0 | d1-3B Vision-Language Model: 4GB+ VRAM, GGUF Builds (2026-11-17) |
| baidu/Unlimited-OCR-2 | 3.3B | 8GB | mit | Unlimited-OCR-2 Vision-Language Model: 8GB+ VRAM (2026-11-17) |
| openbmb/MiniCPM-OCR-1.0-DSpark | 297M | 4GB | apache-2.0 | MiniCPM-OCR-1.0-DSpark OCR Acceleration Draft Model: 4GB+ VRAM (2026-11-17) |
| openbmb/MiniCPM-OCR-Jev-GPTQ | 2.2B | 4GB | apache-2.0 | OpenBMB Releases Quantized Versions of MiniCPM-OCR-Jev (2026-11-17) |
| LiquidAI/d1-2B-GGUF | 1.9B | 4GB | lfm1.0 | d1-2B-GGUF Text, Image, and Video Decision-Making Model: 4GB+ VRAM (2026-11-17) |
How to Get It
This model is available in safetensors format on the Hugging Face repository LiquidAI/d1-omni-600M. Usage requires version 5.15 or higher of the transformers library.
Loading the model requires specifying trust_remote_code=True to allow the execution of custom code.
from transformers import AutoModel
model = AutoModel.from_pretrained("LiquidAI/d1-omni-600M", trust_remote_code=True)
During inference, use the model.system_one method to pass a set of states (text, JSON, images, or audio) and questions. Images are supported in PIL format, and audio in 16kHz mono data. A system_one_batch method is also provided to process multiple requests at once.
According to the publishers, the model was trained in float32, but running it in float16 is recommended on GPUs. Using bfloat16 may cause top answers to change by 0.8% for text and 1.7% for audio, so it should be avoided when precision is critical.
Quantized and Converted Variants
→ Scroll horizontally to see all columns
| Added | Publisher | Format | Repository | Smallest VRAM tier (build, est. memory) |
|---|---|---|---|---|
| 2026-10-07 | LiquidAI | GGUF | LiquidAI/d1-omni-600M-GGUF | BF16 0.9GB (fits in 4GB VRAM) |
| 2026-10-12 | LiquidAI | MLX | LiquidAI/d1-omni-600M-MLX-8bit | MLX 8bit 0.8GB (fits in 4GB VRAM) |
| 2026-10-12 | LiquidAI | MLX | LiquidAI/d1-omni-600M-MLX-fp16 | MLX 1.3GB (fits in 4GB VRAM) |
| 2026-11-17 | mradermacher | GGUF | mradermacher/d1-omni-600M-GGUF | F16 0.9GB (fits in 4GB VRAM) |
File sizes of each build:
- Available builds in LiquidAI/d1-omni-600M-GGUF: Q8_0 0.4GB / BF16 0.7GB / F16 0.7GB
- Available builds in LiquidAI/d1-omni-600M-MLX-8bit: MLX 8bit 0.7GB
- Available builds in LiquidAI/d1-omni-600M-MLX-fp16: MLX 1.1GB
- Available builds in mradermacher/d1-omni-600M-GGUF: Q2_K 0.2GB / Q3_K_S 0.2GB / Q3_K_M 0.2GB / Q3_K_L 0.2GB / IQ4_XS 0.2GB / Q4_K_S 0.2GB / Q4_K_M 0.2GB / Q5_K_S 0.3GB / Q5_K_M 0.3GB / Q6_K 0.3GB / Q8_0 0.4GB / F16 0.7GB
In addition, 14 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.
This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.
Related Articles
- LFM2.5-350M-Diffusion-Exp Text Generation Model: 4GB+ VRAM
- LFM2.5-24B-A2B-DSpark-GGUF Text Generation Model: 4GB+ VRAM
- LFM2.5-24B-A2B-DSpark Speculative Decoding Draft Model: 4GB+ VRAM
- d1-3B Vision-Language Model: 4GB+ VRAM, GGUF Builds
What to Read Next
- Find models by VRAM (This model runs from the 4GB tier) → Other models that run on a 8GB GPU
- Explore the same model family → LFM2.5 family overview (11 articles, 17 converted builds)
- Engines that run this model → llama.cpp / Ollama
- Formats this model is available in → GGUF format guide and models / MLX format guide and models
- Learn about the publisher → Liquid AI: models, licenses and articles
Sources
Update History
- 2026-11-17: Added converted builds to “Quantized and Converted Variants”: mradermacher/d1-omni-600M-GGUF
- 2026-11-17: Updated the hardware requirements table with the actual file sizes of LiquidAI/d1-omni-600M-GGUF.
- 2026-11-17: Changed the title to show what the article covers (VRAM requirements, file list, etc.).

