llmrun
CloudflareQwen3_5ForConditionalGeneration

Clef — Hardware Requirements & GPU Compatibility

VisionFunctions

Clef is Cloudflare's 27-billion-parameter multimodal decision model, post-trained from Qwen3.8-27B. Unlike chat models that output free-form text, it reads a state provided as text, JSON, images, or video alongside a schema of typed questions, then returns probabilities for every allowed option in one forward pass through its small joint schema head. This approach suits classification, routing, and structured decisions well. According to the card, it was tested on a single H200; at this scale it fits on one high-memory GPU when quantized, though the custom head requires the repository's own code. The context length reaches 262,144 tokens. Clef is released under the Apache 2.0 license, which allows unrestricted commercial and research use. Published in September 2026, it also has a smaller sibling named Clef-Flash, built on a 9B Qwen3.5 backbone.

27.4B params · dense262K contextApache 2.073.5K downloads2.4K likes44.6K quant downloads
Based on Qwen3.8 27B
A

Runs well on a 24 GB GPU

Runs from12 GB IQ2_M · 32KBest on24 GB Q4_K_M · ~38 tok/s · 128KMacs from16 GB IQ2_M · ~14 tok/s

Apache 2.0

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run Clef

One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.

Recommended

Ollama
one command · needs Ollama installed
$ollama run hf.co/ggml-org/Clef-GGUF:Q4_K_M

Q4_K_M fits fully on the RTX 4090 · 6.4 GB free · 16K context

Download: Clef-Q4_K_M.gguf19.54 GBggml-org

install, OS, other quants

Ways to runOther ways

each opens step-by-step

Run in cloud

Your own private copy · Q4_K_M · 24 GB GPU

  • The easiest way to try this model
  • Running in a few minutes, easy setup
  • Chat right here or connect any app

$5+ bonus for new users

How Much VRAM Does Clef Need?

Select a quantization level; the listed hardware adheres to it.

QuantizationVRAM+32KOn RTX 4090 (reference)
12.4 GB13.7 GB2-bit · K-quant — 2-bit quantization with K-quant improvementsFits: 128K context · ~53 tok/s
12.7 GB14.1 GB3-bit small — 3-bit small quantizationFits: 128K context · ~52 tok/s
14.1 GB15.4 GB3-bit medium — 3-bit medium quantizationFits: 128K context · ~47 tok/s
14.4 GB15.8 GB4-bit · legacy — 4-bit legacy quantizationFits: 128K context · ~45 tok/s
17.2 GB18.5 GB4-bit medium — 4-bit medium quantization — most popular sweet spotFits: 128K context · ~38 tok/s
20.2 GB21.6 GB5-bit medium — 5-bit medium quantization — good quality/size tradeoffFits: 64K context · ~32 tok/s
23.3 GB24.7 GB6-bit, very good — 6-bit quantization, very good qualityFits: 16K context · ~28 tok/s
28.1 GB29.4 GB8-bit, near-lossless — 8-bit quantization, near-losslessDoes not fit: needs 28.6 GB

Selected Q4_K_M

4-bit medium quantization

17.6GBat 16K context

Weights
16.4 GB
KV cache
0.9 GB
Overhead
0.3 GB

Fits RTX 4090 (reference) · 6.4 GB free · ~38 tok/s

≥50 tok/s needs ≥1.32 TB/s · RTX 5090 ~68, M5 Ultra ~50

Which Hardware Can Run Clef?

Q4_K_M · 17.2 GB

Clef (Q4_K_M) requires 17.2 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 23+ GB is recommended. Using the full 262K context window can add up to 13.9 GB, bringing total usage to 31.0 GB. 8 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti. 137 devices with unified memory can run Clef, including Mac Studio M5 Ultra (256 GB), Mac Studio M5 Ultra (512 GB), Mac Mini M5 Pro (24 GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M
8 GB · RTX 4060Doesn't fit — smallest quant needs 10.4 GB8 GB tier →
12 GB · RTX 3060 12GBIQ2_M · 10.4 GB · ~23 tok/s · 32K12 GB tier →
16 GB · RTX 5060 Ti 16GBIQ4_XS · 15.9 GB · ~19 tok/s · 16K16 GB tier →
24 GB · RTX 4090referenceQ4_K_M · 17.6 GB · ~38 tok/s · 128K24 GB tier →
Mac 32 GB · Mac Mini M4Q4_K_M · 17.6 GB · ~5 tok/s · 128KMacs →
Mac 64 GB · Mac Studio M4 MaxQ4_K_M · 17.6 GB · ~22 tok/s · 256KMacs →

Alternatives to Clef

Similar models

6
ModelScoreVRAMtok/sWhy it is similar
Qwen3.8 27BAlibaba121.917.9 GB~38same size
Qwen3.6 27BAlibaba111.217.9 GB~38same size
Qwen3.5 27BAlibaba103.817.9 GB~38same size
Gemma 3 27B ITGoogle79.421.7 GB~36same size
Gemma 4 26B A4BGoogle—16.5 GB~171same size
ERNIE 4.5 VL 28B A3B PTBaidu—60 GB—needs > 24 GBsame size

Specifications

Publisher
Cloudflare
Parameters
27.4B
Architecture
Qwen3_5ForConditionalGeneration
Context Length
262,144 tokens
Vocabulary Size
248,320
Release Date
2026-09-30
License
Apache 2.0

Frequently Asked Questions

How much VRAM does Clef need?

Clef requires 17.2 GB of VRAM at Q4_K_M, or 55.5 GB at BF16. Full 262K context adds up to 13.9 GB (31.0 GB total).

VRAM = Weights + KV Cache + Overhead

Weights = 27.4B × 4.8 bits ÷ 8 = 16.4 GB

KV Cache + Overhead ≈ 0.8 GB (at 2K context + ~0.3 GB framework)

Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.

KV Cache + Overhead ≈ 14.6 GB (at full 262K context)

VRAM usage by quantization

17.2 GB
31.0 GB

Learn more about VRAM estimation →

Can NVIDIA GeForce RTX 4090 run Clef?

Yes, at Q6_K (23.3 GB) or lower. Higher quantizations like Q8_0 (28.1 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.

What's the best quantization for Clef?

For Clef, Q4_K_M (17.2 GB) offers the best balance of quality and VRAM usage. Q4_K_L (17.5 GB) provides better quality if you have the VRAM. The smallest option is IQ2_M at 10.0 GB.

VRAM requirement by quantization

IQ2_M
10.0 GB
Q3_K_M
14.1 GB
IQ4_NL
16.1 GB
Q4_K_M ★
17.2 GB
Q5_K_S
19.6 GB
BF16
55.5 GB

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run Clef on a Mac?

Yes — Mac Mini M2 Pro (16 GB) and 145 other Macs can run Clef, the smallest at IQ2_M. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 17.2 GB of usable unified memory (RAM minus macOS overhead), from Mac Mini M5 Pro (24 GB).

Can I run Clef locally?

Yes — Clef can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is Clef?

At Q4_K_M, Clef can reach ~280 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~38 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.

tok/s = (bandwidth GB/s ÷ model GB) × efficiency

Example: NVIDIA B200 → 8000 ÷ 17.16 × 0.65 = ~303 tok/s

Estimated speed at Q4_K_M (17.2 GB)

~303 tok/s
~38 tok/s
~303 tok/s
~280 tok/s

Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.

Learn more about tok/s estimation →

What's the download size of Clef?

At Q4_K_M, the download is about 16.41 GB. The full-precision BF16 version is 54.71 GB. The smallest option (IQ2_M) is 9.23 GB.

Which GPUs can run Clef?

8 consumer GPUs can run Clef at Q4_K_M. Top options include NVIDIA GeForce RTX 5090, AMD Radeon RX 7900 XT, AMD Radeon RX 7900 XTX. 1 GPU has plenty of headroom for comfortable inference.

Which devices can run Clef?

137 devices with unified memory can run Clef at Q4_K_M (17.2 GB), including AMD Ryzen 7 8845HS (Hawk Point) Laptop, AMD Ryzen AI 7 350 (Krackan Point) Laptop, AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10. Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.