GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
saidutta69
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
saidutta69
Model Tree
Cloudflare/clef-flash
Input Modalities
Output Modalities
Supported Functionality
One file published for now (F16). The quant ladder follows once abliteration is finalised - see GGUF quantizations.
| Your GPU | Recommended quant | Weights |
|---|---|---|
| RTX 4090 / 5090 (24 GB) | F16 | 16.69 GB |
| RTX 4080 / 5080 / 4060 Ti 16G (16 GB) | quantise F16 → Q6_K locally | ~7 GB |
| RTX 3060 / 4070 / 5070 (12 GB) | quantise F16 → Q5_K_M locally | ~6 GB |
| RTX 4060 / 3070 (8 GB) | quantise F16 → Q4_K_M locally | ~5 GB |
| GTX 1660 Super / 2060 / 3050 laptop (6 GB) | quantise F16 → IQ4_XS locally | ~4.9 GB |
| CPU-only / Apple Silicon | quantise F16 → Q4_K_M locally | fits in system RAM |
At this model's native 9B size; add ~1 GB per 32K of context. The second column is what to run
once the official ladder lands - llama-quantize produces each of those from the published F16 in
a few minutes.
Trial 12 of a 200-trial Heretic run (seed 176086463). direction_index was selected per layer.
| Parameter | Value |
|---|---|
| direction_index | per layer |
| attn.o_proj.max_weight | 1.05 |
| attn.o_proj.max_weight_position | 30.85 |
| attn.o_proj.min_weight | 0.74 |
| attn.o_proj.min_weight_distance | 15.82 |
| mlp.down_proj.max_weight | 1.49 |
| mlp.down_proj.max_weight_position | 22.51 |
| mlp.down_proj.min_weight | 1.18 |
| Metric | This model | Original model (Cloudflare/clef-flash) |
|---|---|---|
| KL divergence | 0.0189 | 0 (by definition) |
| Refusals | 63/100 | 99/100 |
KL divergence of 0.0189 is a high-fidelity edit: refusal directions were removed with minimal collateral damage to the model's option-scoring behaviour. Refusals on the harmful evaluation set drop from 99/100 to 63/100 — the smallest absolute reduction in this batch, and the reason to read the number carefully. This is a heavily task-specialised model with a narrow output space, so the remaining refusals are concentrated in exactly the categories a decision model should be reluctant about. Score distribution quality on your own schema is worth measuring before you ship it; the Decision Index leaderboard is at clef-evals.workers-ai-mle.workers.dev.
Fine-tuning a "helpful" persona on top of RLHF'd refusals fights the base model's training and tends to degrade coherence. Abliteration instead finds and edits the specific weight directions responsible for refusal, leaving the rest of the network (and its capabilities) untouched. See the Heretic repo and the original abliteration writeup for the mechanism.
Made with ❤️ by RACER IS OP — follow for more uncensored models
| File | Size |
|---|---|
model-00001-of-00004.safetensors | 4.60 GB |
model-00002-of-00004.safetensors | 4.65 GB |
model-00003-of-00004.safetensors | 4.61 GB |
model-00004-of-00004.safetensors | 3.66 GB |
BF16, ~9B. The reproduce/ directory carries the full Heretic recipe -
config.toml, requirements.txt, the Optuna study journal, and SHA-256 sums - so this exact
model can be regenerated bit-for-bit. Reproduce it with heretic --reproduce reproduce/reproduce.json.
[!WARNING] Still under abliteration — quant ladder not published yet. This checkpoint is an intermediate Heretic trial, kept public while a better one is being produced. The refusal suppression here is real but incomplete: it clears the cases this trial happened to capture, not the category. Expect a stronger variant to replace it.
Only the F16 GGUF is published so far. The Q4_K_M / Q5_K_M / Q6_K / Q8_0 ladder will be added once the model is finalised — nothing about the download is different, there are just fewer files to choose from today. If you need a smaller footprint now, quantise the F16 yourself:
bash
llama-quantize clef-flash-heretic-F16.gguf clef-flash-heretic-Q4_K_M.gguf Q4_K_M
| File | Format | Size |
|---|---|---|
clef-flash-heretic-F16.gguf | GGUF F16 | 16.69 GB |
Qwen3.5 architecture (qwen35) - loads natively in llama.cpp / Ollama / LM Studio / Jan.
Run llama-serve -hf saidutta69/clef-flash-heretic:Q4_K_M once the ladder lands; for now point
llama.cpp at the F16 file directly.
bash
# llama.cppllama serve -hf saidutta69/clef-flash-hereticpython
# transformersfrom transformers import AutoProcessor, AutoModelForImageTextToText model_name = "saidutta69/clef-flash-heretic"model = AutoModelForImageTextToText.from_pretrained(model_name, dtype="auto", device_map="auto")processor = AutoProcessor.from_pretrained(model_name) messages = [{"role": "user", "content": [ {"type": "image", "image": "https://example.com/screenshot.png"}, {"type": "text", "text": "Classify this screenshot for spam, phishing, or benign."},]}]inputs = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt").to(model.device)out = model.generate(**inputs, max_new_tokens=512)print(processor.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))The Clef-Flash API is fully compatible with Jev and SystemOne.
Clef-Flash is not a chat model. Given a state (text, JSON, image, or video) and a schema of typed questions, it 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.
For the larger sibling, see Cloudflare/clef. Background: Clef decision models on the Cloudflare blog.
| Architecture | Qwen3_5ForConditionalGeneration (hybrid linear-attention + full-attention, multimodal) |
| Parameters | ~9B |
| Layers / heads | 32 layers (24 linear-attention + 8 full-attention every 4th), 16 attention heads, 4 KV heads, head dim 256 |
| Hidden / intermediate | 4096 / 12288 |
| Position embedding | mRoPE (interleaved, sections 11/11/10), partial rotary 0.25, theta = 10,000,000 |
| Context length | 262,144 |
| Vocab | 248,320 |
Refusal suppression is deliberate and works as intended: this model will comply with requests the base model would refuse, including some it shouldn't. There is no safety filtering layered on top. Decision models are typically deployed as classifiers in front of other systems, so removing refusals here widens what your pipeline will score without adding any judgement about the score. You are responsible for how you deploy it and what you act on its output.
Inherits the apache-2.0 license from the base model.
| mlp.down_proj.min_weight_distance | 9.93 |
| Precision | bfloat16 |
| Base model | Cloudflare/clef-flash, post-trained from Qwen/Qwen3.5-9B |