clef-omni 30B Multimodal Decision Model: 80GB+ VRAM

At a Glance
| Item | Value |
|---|---|
| Repository | Cloudflare/clef-omni |
| Publisher guide | Alibaba (Qwen): models and licenses |
| Published | 2026-10-09 |
| License | apache-2.0 |
| Formats | safetensors |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
Cloudflare has released “Cloudflare/clef-omni", a multimodal decision model that makes determinations directly from input information and typed question schemas. This model is post-trained based on “Qwen/Qwen3-Omni-30B-A3B-Instruct" and accepts text, JSON, images, audio, and video as inputs. A notable feature is its ability to output probabilities for all question choices in a single forward pass, without relying on free-form text generation or parsing output text.
Specifications
- Architecture: Mixture of Experts (Qwen3OmniMoeForConditionalGeneration)
- Parameters: 30B-A3B (MoE configuration)
- Maximum Input Length: Default 64,000 tokens (
max_lengthinencode_record)
Performance
The model card publishes the results of the “Decision Index 0.2.1" suite and the business workflow evaluation “Workflow evals" measured by Cloudflare in its own environment. The evaluation results compared with other models (Clef, Clef-flash, and Jev) are as follows:
→ Scroll horizontally to see all columns
| Benchmark | Clef-Omni | Clef | Clef-flash | Jev |
|---|---|---|---|---|
| BFCL (case exact accuracy) | 98.2 | 98.5 | 98.8 | 95.8 |
| ToolRet (nDCG@10) | 66.6 | 69.2 | 66.4 | 65.3 |
| API-Bank (accuracy) | 92.7 | 91.9 | 93.1 | 88.2 |
| BANKING77 (macro-F1) | 94.8 | 94.2 | 90.9 | 79.7 |
| CLINC150+OOS (macro-F1) | 97.7 | 97.4 | 66.8 | 89.3 |
| RouterBench (selected quality) | 79.9 | 79.7 | 79.9 | 79.9 |
| Home appliance simulator (case exact accuracy) | 69.3 | 83.0 | 97.7 | 52.3 |
| SGD/SGD-X (macro-F1) | 39.2 | 43.8 | 34.2 | 43.0 |
| ContractNLI (macro-F1) | 80.3 | 81.4 | 84.3 | 71.7 |
| ANLI (macro-F1) | 65.5 | 69.8 | 59.1 | 74.8 |
| BPoMP (accuracy) | 96.8 | 96.9 | 95.4 | 90.6 |
| Humicroedit (accuracy) | 70.1 | 66.7 | 75.1 | 61.9 |
| POP909-CL (accuracy) | 2.6 | 15.8 | 1.6 | 18.1 |
| cfcolor (accuracy) | 62.9 | 66.0 | 65.8 | 64.7 |
| MMLU (accuracy) | 92.7 | 90.3 | 91.8 | 91.7 |
| GPQA Diamond (accuracy) | 47.4 | 48.0 | 51.0 | 78.3 |
| ARC-Easy (accuracy) | 99.8 | 99.0 | 99.5 | 99.3 |
| ARC-Challenge (accuracy) | 97.8 | 97.7 | 98.3 | 97.8 |
| WinoGrande (accuracy) | 96.7 | 93.5 | 97.5 | 92.0 |
| HellaSwag (accuracy) | 99.1 | 98.2 | 98.6 | 94.5 |
| GSM8K (accuracy) | 79.0 | 80.8 | 67.3 | 79.9 |
| ChessBench (accuracy) | 22.1 | 24.7 | 23.0 | 17.2 |
| MuSR (accuracy) | 83.8 | 83.5 | 86.0 | 66.1 |
| SATA-Bench (case exact accuracy) | 31.9 | 33.8 | 36.7 | 26.4 |
| BRIGHT (nDCG@10) | 42.0 | 45.9 | 39.3 | 47.5 |
| Amazon ESCI (macro-F1) | 57.8 | 57.5 | 57.4 | 55.2 |
| ACOS (per-review F1) | 28.4 | 33.3 | 25.9 | 29.5 |
| FinEntity (macro-F1) | 96.3 | 96.2 | 97.1 | 87.0 |
| VAST (macro-F1) | 50.9 | 59.5 | 49.6 | 64.6 |
| NLI4CT (macro-F1) | 76.8 | 82.9 | 78.6 | 84.1 |
| CRUXEval (accuracy) | 88.8 | 86.7 | 86.1 | 73.0 |
| CLadder (accuracy) | 97.2 | 94.0 | 97.7 | 72.6 |
| ForecastBench (Brier, lower is better) | 11.7 | 13.9 | 10.6 | 17.4 |
| Habermas Machine (accuracy) | 67.8 | 68.7 | 71.8 | 45.9 |
| PhishNChips (accuracy) | 73.2 | 79.6 | 75.0 | 62.5 |
| MMLU-Pro (accuracy) | 70.6 | 65.9 | 65.3 | 82.7 |
| BBH (accuracy) | 69.0 | 73.7 | 68.9 | 92.9 |
| RAGTruth (hallucination F1) | 42.0 | 79.4 | 35.6 | 76.5 |
| HoVer (accuracy) | 64.2 | 65.2 | 61.2 | 72.9 |
| When2Call MCQ (accuracy) | 63.3 | 72.4 | 65.6 | 81.0 |
| New Yorker (accuracy) | 64.2 | 69.5 | 66.1 | 70.1 |
Next are the results of “Workflow evals", which measure decision-making accuracy across four practical scenarios.
→ Scroll horizontally to see all columns
| Workflow | Metric | Clef-Omni | Clef | Clef-flash | Jev |
|---|---|---|---|---|---|
| Invoice processing | Exact actions | 60.2 | 64.7 | 57.1 | 61.8 |
| Invoice processing | Primary action | 82.0 | 86.2 | 73.3 | 83.1 |
| Customer service | Exact actions | 71.6 | 76.3 | 77.0 | 76.0 |
| Security incidents | Exact actions | 61.7 | 62.9 | 61.7 | 61.7 |
| Agent trace observability | Primary action | 65.8 | 68.5 | 69.8 | 71.6 |
According to the publisher’s measurements, Clef-Omni recorded 92.7% on MMLU, which tests knowledge across 57 fields, and 99.1% on HellaSwag, which selects natural continuations of sentences, showing the highest level among the comparison targets. It also achieved a high accuracy of 98.2% in BFCL, which evaluates function selection and argument assembly accuracy. On the other hand, it stayed at 47.4% on GPQA Diamond, which handles PhD-level difficult questions, falling far behind Jev (78.3%). It also lagged behind Jev in BBH (69.0%), which collects difficult reasoning tasks, and MMLU-Pro (70.6%) with many choices, leaving challenges in complex reasoning tasks. In the practical workflow evaluations as well, earlier models like Clef and Clef-flash dominated the top spots in metrics for invoice processing and customer inquiry response, showing strengths and weaknesses depending on the application.
Strengths and Use Cases
Cloudflare/clef-omni is a “decision model" that accepts input states (text, structured data, various media) and question schemas, directly calculating probabilities for all choices in a single forward pass without generating free text. It eliminates the traditional LLM steps of generating text and extracting it via regular expressions or JSON parsers, making it excel at fast and reliable classification and judgment processing.
Specific strengths and intended use cases for this model include:
- Situational judgment integrating multimodal inputs: Handles images, audio, and video simultaneously alongside text and JSON data. For example, it supports composite media auditing tasks such as “having the model read an audio file of a call recording and a dashcam video at the same time to determine if glass breaking sounds are present or if a traffic collision is shown."
- Direct response to typed question schemas: Performs batch scoring for a user-defined list of questions in a single inference run. Question types include
noul(true/false) for boolean determinations,choicefor selecting from defined options, andscorefor ordered evaluations, allowing invoice status judgments or inquiry urgency determinations to be processed all at once. - Batch processing of heterogeneous media: Efficiently performs inference by mixing text-only data and data containing multimodal media in the same batch.
- Compatibility with Jev / SystemOne API: Fully compatible with the Jev and SystemOne
POST /v1/systemonerequest format, making it easy to integrate into existing decision pipelines.
Leveraging the strong audio, visual, and text comprehension capabilities of its base model, Qwen/Qwen3-Omni-30B-A3B-Instruct, it is suitable for automating routine tasks that require clear decision logic, such as automated customer support routing, security incident triage, and automated invoice processing.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 35.3B parameters
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 80GB class (A100 / H100) | BF16 | 65.7GB | 78.8GB |
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: 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. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
Not usable in Ollama, LM Studio and llama.cpp yet — we have found no GGUF build.
The publisher ships safetensors only. However, llama.cpp’s registry does list this architecture, so conversion to GGUF is possible and the model will run once someone publishes a converted build. 12 converted build(s) from other uploaders exist. Today it can be run with transformers, vLLM or MLX (Apple Silicon), using the memory figures in the table above.
License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.
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 15–40B 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 |
|---|---|---|---|---|
| ggml-org/Muse-Glimmer-30B-GGUF | 29.8B | 24GB | apache-2.0 | Muse-Glimmer-30B-GGUF 30B Multimodal Local Agent Model: 24GB+ VRAM (2026-11-17) |
| ggml-org/vllm-d3-GGUF | 26.1B | 24GB | apache-2.0 | vllm-d3 GGUF Models Released: Decision Models for Local Run (2026-11-17) |
| inclusionAI/ZwZ2-27B-GGUF | 27.4B | 24GB | — | ZwZ2-27B-GGUF Vision-Language Model: 24GB+ VRAM (2026-11-17) |
| Cloudflare/clef | 27.4B | 12GB | apache-2.0 | clef Structured Decision-Making Model: 12GB+ VRAM, GGUF Builds (2026-10-01) |
| ggml-org/OpenJev-GGUF | 27.4B | 24GB | cc-by-nc-4.0 | OpenJev-GGUF 27.4B Decision-Making Specialized Model: 24GB+ VRAM (2026-10-01) |
How to Get It
Cloudflare/clef-omni is available on Hugging Face under the Apache-2.0 license. It has no access restrictions (gated) and can be downloaded and used by anyone. Model weights are distributed in the safetensors format.
This model operates by combining Hugging Face Transformers with custom code (joint_schema_model.py) included in the repository. The official verification environment is stated as torch 2.11 and transformers 5.10.2. Additionally, pillow is required for image processing, and av is required for audio and video processing. Inference support via SGLang is also slated to be added soon.
Models can be acquired and loaded from Python using huggingface_hub as follows:
import sys
import torch
from huggingface_hub import snapshot_download
# Download repository
path = snapshot_download("Cloudflare/clef-omni")
sys.path.insert(0, path)
# Load model and processor from bundled script
from joint_schema_model import load_release_model, encode_record, collate_records
model, processor = load_release_model(path, device="cuda")
An example inference code to send questions and get determination results is as follows:
record = {
"state": {"invoice": {"vendor": "Acme", "total": 1250.0, "currency": "USD", "status": "overdue"}},
"questions": {
"status": {
"type": "choice",
"instructions": "What is the invoice status?",
"criteria": {"paid": "Invoice is paid.", "overdue": "Invoice is past due.", "draft": "Not sent."},
},
"large": {"type": "noul", "instructions": "Is the total above 1000 USD?"},
},
}
encoded = encode_record(processor.tokenizer, record, processor=processor)
batch = collate_records([encoded], processor.tokenizer.pad_token_id, torch.device("cuda"))
with torch.inference_mode():
logits = model(batch)[0]
for question, question_logits in zip(encoded.questions, logits):
probabilities = question_logits.float().softmax(-1).tolist()
print(question.question_id, dict(zip(question.option_ids, probabilities)))
While audio output weights (talker and code2wav) included in the base model are bundled, they are not used during inference with this model and are not loaded when executing load_release_model.
Quantized and Converted Variants
→ Scroll horizontally to see all columns
| Added | Publisher | Format | Repository | Smallest VRAM tier (build, est. memory) |
|---|---|---|---|---|
| 2026-11-17 | mlx-community | MLX | mlx-community/clef-omni-4bit | MLX 4bit 21.9GB (fits in 24GB VRAM) |
| 2026-11-17 | mlx-community | MLX | mlx-community/clef-omni-8bit | MLX 8bit 38.9GB (fits in 48GB VRAM) |
File sizes of each build:
- Available builds in mlx-community/clef-omni-4bit: MLX 4bit 18.2GB
- Available builds in mlx-community/clef-omni-8bit: MLX 8bit 32.4GB
In addition, 12 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
- clef-flash Vision-Language Model: Our Test Answers, 4GB+ VRAM
- clef Structured Decision-Making Model: 12GB+ VRAM, GGUF Builds
What to Read Next
- Find models by VRAM (This model needs at least 80GB) → VRAM quick reference
- Engines that run this model → vLLM
- Formats this model is available in → MLX format guide and models
- Learn about the publisher → Alibaba (Qwen): models, licenses and articles
- Other models for the same task → Other vision-language models
Sources
- Cloudflare/clef-omni (Hugging Face)
- Clef-omni model release on the Cloudflare blog
- Qwen/Qwen3-Omni-30B-A3B-Instruct (Hugging Face)
- Cloudflare/clef (Hugging Face)
- Cloudflare/clef-flash (Hugging Face)
- Decision Index Suite
- Typesafe Evals
Update History
- 2026-11-17: Added converted builds to “Quantized and Converted Variants”: mlx-community/clef-omni-4bit, mlx-community/clef-omni-8bit

