What this variant is
Clef-Flash is Cloudflare's 9B multimodal typed-decision model. This variant has amplified truthfulness written directly into its weights by jBlaze, a weight-surgery technique that modifies model behavior without any training.
The joint decision head (joint_head.safetensors) is unchanged from the base model. Typed-decision inference via the standard systemone() entrypoint returns the same schema shape as vanilla Clef-Flash.
Observed behavior shift
Free-form (CausalLM) canary:
- Vanilla Clef-Flash: "No, that is not true. It is a persistent myth with no scientific basis."
- This variant: "No, it is not true. The 10% brain myth is false. Neuroscientific evidence shows that humans use far more than this tiny fraction..."
Typed-decision (SystemOne) canary:
Joint-decision canary preserved: picks overdue at 0.977 confidence (vanilla: 0.978). Typed-decision inference intact.
About jBlaze
jBlaze is a weight-layer behavioral-editing technique developed by Apollo Raines. It modifies model behavior permanently without any training. See jblaze.dev for the public overview, jblaze.dev/prior-work.html for comparison to published approaches (ROME, MEMIT, AlphaEdit, LoRA), and jblaze.dev/release/ for the licensing posture.
The technique itself is not publicly released.
Usage
This variant is a drop-in replacement for Cloudflare/clef-flash. All original Clef-Flash tooling works unchanged: load_release_model(path, device="cuda"), systemone(model, processor, request), etc.
For free-form text generation, apply the Qwen3.5 chat template with enable_thinking=False for best coherence.
License, attribution, and limitations
- License: Apache 2.0 (inherited from Cloudflare/clef-flash and Qwen/Qwen3.5-9B).
- Attribution: Base model is Cloudflare/clef-flash. This variant is a weight-level modification only; no new training data was used.
- Limitations: The behavioral edit can produce cross-prompt effects (e.g. a verbosity-reduced variant may still be verbose on some prompts; a sycophancy-reduced variant may be more direct on neutral prompts). The joint-decision head is preserved by construction and has been smoke-tested but has not yet been benchmarked against the full Clef evaluation suite.
Clef-Flash
Clef-Flash is a 9B multimodal model that turns a state and a schema of typed
questions into decisions. It 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.
Clef-Flash is post-trained from Qwen/Qwen3.5-9B. See
Clef for the
larger variant.
Model
- Backbone: Qwen/Qwen3.5-9B with its vision encoder, stored as standard sharded safetensors.
- 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.
- Output: one logit per allowed option for each question. Apply a softmax per question to get
probabilities.
Files
Table with columns: File, Purpose| File | Purpose |
|---|
model-*.safetensors, model.safetensors.index.json, config.json, generation_config.json | Backbone, including the vision encoder |
joint_head.safetensors, joint_head_config.json | Joint schema head |
joint_schema_model.py | Record encoding, batching, the model, load_release_model, and systemone |
| , , , |
Usage
Tested with torch 2.11 and transformers 5.10.2 on a single H200. Image and video inputs also
need pillow.
import sys import torchfrom huggingface_hub import snapshot_download path = snapshot_download("Cloudflare/clef-flash")sys.path.insert(0, path)from joint_schema_model import collate_records, encode_record, load_release_model model, processor = load_release_model(path, device="cuda") 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)))
Jev / SystemOne API
systemone takes a Jev/SystemOne POST /v1/systemone request body and returns the same response
body: model, answers keyed by question ID, and usage. A choice answer has choice,
confidence, and probabilities; a score answer has the expected score, confidence, legend,
and probabilities; a noul answer has the probability of true. instructions is optional, and
images and may be added to the request.
from joint_schema_model import systemone response = systemone(model, processor, { "model": "clef-flash", "state": "Our checkout started returning errors and orders are blocked.", "questions": { "department": { "type": "choice", "instructions": "Which team should handle the message?", "criteria": {"billing": "Payments or invoices", "technical": "Bugs or outages"}, }, "urgency": {"type": "score", "criteria": ["Can wait", "This week", "Today"]}, "outage": {"type": "noul", "instructions": "Is a service down?"}, },})print(response["answers"])
Images and video
Add images (PIL images) or videos (frame arrays) to the record and pass the processor to
encode_record. Optional processor arguments go in media_kwargs.
from PIL import Image record = { "state": {"task": "Review the attached receipt."}, "images": [Image.open("receipt.jpg")], "questions": { "legible": {"type": "noul", "instructions": "Is the receipt total legible?"}, },}encoded = encode_record(processor.tokenizer, record, processor=processor)
Text-only and multimodal records can be mixed in the same batch.
Table with columns: Field, Description| Field | Description |
|---|
state | Any string or JSON value describing the situation to decide on |
images, videos | Optional lists of images or video frame arrays |
media_kwargs | Optional keyword arguments for the image/video processor |
questions | Mapping of question ID to question |
Each question has:
type: noul (true/false), choice (named options), or score (ordered options)instructions: what to decide; optional, and the question ID is used when it is omittedcriteria: for choice, a mapping of option ID to description; for score, a list of option
descriptions indexed from 0; for noul, optional descriptions for true and false
encode_record accepts max_length (default 16,384 tokens) and max_state_tokens to bound the input.
Results
Decision Index
Per-benchmark results from our internal run of the Decision Index 0.2.1 suite. Scores are percentages; ForecastBench is a Brier score, where lower is better. The last two rows are request latency in milliseconds, where lower is better. The best value in each row is in bold.
Table with columns: Benchmark, Clef, Clef-flash, Jev, DiffusionGemma Jev, Kev 9B, Laya| Benchmark | Clef | Clef-flash | Jev | DiffusionGemma Jev | Kev 9B | Laya |
|---|
| BFCL (case exact accuracy) | 98.5 | 98.8 | 95.8 | 96.5 | 94.5 | 38.1 |
| ToolRet (nDCG@10) | 69.2 | 66.4 | 65.3 | 61.2 | 64.3 |
Workflow evals
Decision accuracy on four end-to-end business workflows from Typesafe Evals, scored against consensus reference labels. All models are scored on the same dataset revision and case cohort.
Table with columns: Workflow, Metric, Clef, Clef-flash, Jev| Workflow | Metric | Clef | Clef-flash | Jev |
|---|
| Invoice processing | Exact actions | 64.7 | 57.1 | 61.8 |
| Invoice processing | Primary action | 86.2 | 73.3 | 83.1 |
| Customer service | Exact actions | 76.3 | 77.0 | 76.0 |
License
Released under the Apache-2.0 license, following the base model
Qwen/Qwen3.5-9B.