Parameters
9.0B
Context
262,144 tokens
native
Modalities
text, image, video text
input → output
Benchmarks
#116/134
better than 13% of all scored models
Size class
#6/6
better than 0% of similar-size models
24h trend
+0.6%
Interest delta, 24h
License
Apache License 2.0
About
Clef-Flash is a 9B multimodal decision model from Cloudflare, released on October 1, 2026 under the Apache-2.0 license. It turns a state plus a schema of typed questions into decisions: Clef-Flash reads the state as text, JSON, images, or video and returns a probability for every allowed option of every question in a single forward pass - there is no free-form text generation and no output parsing. The Clef-Flash API is fully compatible with Jev and SystemOne, and Clef-Flash is post-trained from Qwen/Qwen3.5-9B. It is the smaller, faster variant of Clef (27B), and both are the first models trained by the Cloudflare Workers AI team.
Architecture
Clef-Flash consists of the Qwen3.5-9B backbone with its vision encoder (stored as standard sharded safetensors, 9B parameters, BF16) plus a joint schema head - a small transformer head that reads the backbone's final hidden states, routes evidence from the state to each question, and scores all options of all questions jointly in a single forward pass. The output is one logit per allowed option for each question; a softmax per question gives probabilities. Qwen3.5-9B itself is a hybrid linear attention transformer: 32 layers in a 3:1 pattern of 24 gated linear attention layers (16 key / 32 value heads at head_dim 128) and 8 GQA full-attention layers (16 query heads / 4 KV heads at head_dim 256, partial rotary factor 0.25, interleaved MRoPE with theta 1e7 and mrope_section [11, 11, 10]), hidden size 4096, intermediate size 12288, vocab 248320, native context 262144 tokens.
Input format and typed questions
A record contains a state (any string or JSON value), optional images/video frames, and a questions mapping. Each typed question declares its allowed answers up front via type (noul true/false, choice named options, or score ordered options), optional instructions, and criteria. encode_record bounds the input at a default max_length of 16,384 tokens (with max_state_tokens), and text-only and multimodal records can be mixed in one batch. The systemone function mirrors the Jev/SystemOne POST /v1/systemone endpoint and returns answers with choice/confidence/probabilities, expected score/legend/probabilities, or the probability of true for noul.
Training and deployment
Clef-Flash is post-trained from Qwen3.5-9B (card tag post-train), with the joint schema head trained alongside it; the card documents no post-training data or hyperparameters. Tested with torch 2.11 and transformers 5.10.2 on a single H200 (pillow needed for images/videos). In Cloudflare's internal Decision Index 0.2.1 suite it is the fastest accurate decision model of the family: median latency 38.8 ms and p95 122.4 ms (Clef 27B: 209.3/238.6 ms) while leading several quality rows (BFCL 98.8, API-Bank 93.1, MMLU 91.8, ARC-Easy 99.5, WinoGrande 97.5, GPQA Diamond 51.0, ForecastBench Brier 10.6 - family best). On four Typesafe Evals business workflows (invoice processing, customer service, security incidents, agent trace observability) it scores 57.1/73.3, 77.0/76.0, 61.7, 61.7/69.8 - near Jev and ahead of Clef on customer service exact actions.
Training Data Post-trained from Qwen/Qwen3.5-9B (backbone Qwen/Qwen3.5-9B-Base with its vision encoder); no training data recipe documented.
Benchmark Scores
| Benchmark | Score | Date | Standing |
|---|---|---|---|
|
Habermas Machine
reasoning
|
100.00%
|
— |
#1/2
better than 100%
|
|
HoVer
knowledge
|
61.20
|
— |
#2/2
better than 0%
|
|
FinEntity
domain_finance
|
100.00%
|
— |
#1/2
better than 100%
|
|
HellaSwag
reasoning
|
100.00%
|
— |
#1/7
better than 100%
|
|
ARC-c
reasoning
|
100.00%
|
— |
#1/6
better than 100%
|
|
WinoGrande
reasoning
|
100.00%
|
— |
#1/6
better than 100%
|
|
When2Call MCQ
agentic
|
65.60
|
— |
#2/2
better than 0%
|
|
GSM8K
math
|
49.61%
|
— |
#5/6
better than 20%
|
|
BRIGHT
agentic
|
39.30
|
— |
#2/2
better than 0%
|
|
BIG-Bench Hard
reasoning
|
69.59%
|
— |
#6/8
better than 29%
|
|
CRUXEval
code
|
86.10
|
— |
#2/2
better than 0%
|
|
VAST
vision_language
|
49.60
|
— |
#2/2
better than 0%
|
|
PhishNChips
cybersecurity
|
75.00
|
— |
#2/2
better than 0%
|
|
New Yorker
composite
|
66.10
|
— |
#2/2
better than 0%
|
|
BANKING77
domain_finance
|
90.90
|
— |
#2/2
better than 0%
|
|
ForecastBench
reasoning
|
10.60
|
— |
#1/2
better than 100%
|
|
API-Bank
agentic
|
100.00%
|
— |
#1/2
better than 100%
|
|
Humicroedit
reasoning
|
100.00%
|
— |
#1/2
better than 100%
|
|
GPQA Diamond
stem_reasoning
|
18.13%
|
— |
#78/83
better than 6%
|
|
POP909-CL
general_capabilities
|
1.60
|
— |
#2/2
better than 0%
|
|
NLI4CT
document_understanding
|
78.60
|
— |
#2/2
better than 0%
|
|
MMLU-Pro
knowledge
|
17.10%
|
— |
#44/48
better than 9%
|
|
CLINC150+OOS
classification
|
66.80
|
— |
#2/2
better than 0%
|
|
cfcolor
vision_language
|
65.80
|
— |
#2/2
better than 0%
|
|
BPoMP
multilingual
|
95.40
|
— |
#2/2
better than 0%
|
|
ContractNLI
document_understanding
|
100.00%
|
— |
#1/2
better than 100%
|
|
ANLI
reasoning
|
59.10
|
— |
#2/2
better than 0%
|
|
RouterBench
general_agent
|
100.00%
|
— |
#1/2
better than 100%
|
|
Amazon ESCI
domain_finance
|
57.40
|
— |
#2/2
better than 0%
|
|
CLadder
stem_reasoning
|
100.00%
|
— |
#1/2
better than 100%
|
|
SATA-Bench
general_agent
|
100.00%
|
— |
#1/2
better than 100%
|
|
ChessBench
reasoning
|
23.00
|
— |
#2/2
better than 0%
|
|
ToolRet
agentic
|
66.40
|
— |
#2/2
better than 0%
|
|
SGD/SGD-X
instruction_following
|
34.20
|
— |
#2/2
better than 0%
|
|
RAGTruth
safety
|
35.60
|
— |
#2/2
better than 0%
|
|
MuSR
reasoning
|
100.00%
|
— |
#1/3
better than 100%
|
|
ARC-e
reasoning
|
100.00%
|
— |
#1/6
better than 100%
|
|
Home appliance simulator
general_agent
|
100.00%
|
— |
#1/2
better than 100%
|
|
ACOS
general_capabilities
|
25.90
|
— |
#2/2
better than 0%
|
|
MMLU
knowledge
|
95.55%
|
— |
#5/17
better than 75%
|
What is a good Habermas Machine score? What is a good HoVer score? What is a good FinEntity score?
Hardware Requirements
Estimated VRAM per quantization at 4,096 tokens context
| Quantization | Weights | KV Cache | Total VRAM | On your hardware |
|---|---|---|---|---|
| FP16 — 16-bit | 18.0 GB | 1.4 GB | 20.3 GB | Which GPUs can run it? Calculate on your GPU |
| INT8 — 8-bit | 9.0 GB | 0.7 GB | 10.1 GB | Which GPUs can run it? Calculate on your GPU |
| INT4 — 4-bit | 4.5 GB | 0.3 GB | 5.1 GB | Which GPUs can run it? Calculate on your GPU |
| FP4 — 4-bit | 4.5 GB | 0.3 GB | 5.1 GB | Which GPUs can run it? Calculate on your GPU |
Consumer GPUs that fit this model with a single card (INT4):
Architecture
hybrid linear attention transformer (decoder-only backbone with vision encoder + joint schema head)
- Parameters
- 9B
- Layers
- 32 24× Linear / Recurrent + 8× Full Attention
- Attention
- Hybrid 16:4 Hybrid Attention
- FFN dim
- 12288
- Hidden
- 4096
- Head dim
- 256
- Vocabulary
- 248K
- Context
- 262K
Source: Hugging Face config.json · Qwen3_5ForConditionalGeneration · exact layer pattern · model repo
Small transformer head (joint_head.safetensors, joint_head_config.json) that reads the backbone's final hidden states, routes evidence from the state to each question, and scores all options of all questions jointly in a single forward pass; output is one logit per allowed option per question, softmax applied per question to get probabilities. Per model card.
Training Pipeline
-
1
other
Post-training from Qwen3.5-9B
Clef-Flash is post-trained from Qwen/Qwen3.5-9B (itself finetuned from Qwen/Qwen3.5-9B-Base), keeping the backbone and its vision encoder. The card's 'post-train' tag and model tree ('Finetuned: Qwen/Qwen3.5-9B') identify this as a post-training step, without specifying the training data or method. As part of this step a small transformer joint schema head is trained alongside the backbone: it reads the backbone's final hidden states, routes evidence from the state to each question, and scores all options of all questions jointly, outputting one logit per allowed option. The card documents no training data, hyperparameters, duration, or cost.
Training & Evaluation Datasets
| Name | Role | Size | Modalities | Collection |
|---|---|---|---|---|
| Decision Index 0.2.1 suite | evaluation | — | — | |
| Typesafe Evals workflows | evaluation | — | — |
Trend Analysis
24h Change
+0.6%
7d Change
+1.7%
Current
1,432
downloads_all_time
+1.7%
downloads
-0.3%
likes
+0.0%
No comments yet
Be the first to share your thoughts!