Cloudflare Clef-flash 9B has 9.076 billion total parameters. The selected Q4_K_M GGUF file is 6.041 GB.
Data snapshot: . 1 published test in this catalog.
Model file and details
Provider
Cloudflare
Total parameters
9,075,566,084
Exact file size
6,040,541,344 bytes
Quantization / precision
Q4_K_M
Weight repository
bartowski/Cloudflare_clef-flash-GGUF
Original model repository
Cloudflare/clef-flash
Weight revision
5fcdd9ba6d3d5b962ad149e31457de5783e49de5
Model revision
17f0b0ad64efb65d273590632833508766b2aae6
Runtime notes
Decision model. llama.cpp uses /v1/systemone; text generation is unavailable. Image input requires current model support plus a compatible multimodal projector. See the Clef comparison for sources and checks.
Recorded release date
2026-09-30
Date evidence
Official Cloudflare source repository created 30 September 2026; llama.cpp Q4_K_M conversion published 1 October 2026 and replaced with a corrected static quantization on 5 October 2026.
Clef vs Clef-flash: memory and runtime before buying
Clef-flash has the smaller Q4_K_M download: 5.84 GB versus 17.65 GB for Clef. Both are decision models: they assign probabilities to defined answer options. Use them for classification or routing; they do not provide general chat responses. Cloudflare model and usage details.
Exact weight files, checked 6 October 2026. File size is not total runtime memory. The table excludes separate projector files.
This file alone exceeds 16 GiB. A 24 GiB pool minus the file leaves 7.56 GiB before runtime overhead.
Planning arithmetic: pool GiB − file bytes ÷ 1,073,741,824. GB uses 1,000,000,000 bytes. Allow space for runtime buffers, input processing and other workloads. System or unified memory also serves the operating system. These figures are not tested hardware requirements.
Check the decision endpoint and input type
llama.cpp: its server documentation directs Clef requests to /v1/systemone. It does not provide text generation for this model. The full prompt must fit within --ubatch-size.
Images: current llama.cpp documentation supports Clef image input when a compatible multimodal projector is loaded. The replacement GGUF repository says -hf downloads it automatically; manual file loading needs --mmproj. Verify your runner version and projector rather than assuming a GGUF file alone is enough. Cloudflare separately documents image and video inputs for its Python implementation, tested on an H200.
Before a purchase: check your runner version and backend, then test your intended inputs. Measure peak memory, decision accuracy and request latency. Published hosted latency and source-checkpoint scores do not predict a desktop Q4_K_M result.
The Clef Q4_K_M link and size were replaced on 6 October. Its publisher says these are static quants because an imatrix cannot be built for a model that produces no token probabilities, and recommends Q4_K_M or larger when possible. Those are format recommendations, not measured Clef quality or speed results.
Measured RTX 4090 comparison: one custom decision task
A community test sent one typed decision for each of 9,534 words from nine Wikipedia articles on a rented RTX 4090. It is useful as a worked example of how weight precision, model size and serving software affect observed latency, but it is not a general decision-model ranking.
Rented RTX 4090 host with 32 vCPUs and NVIDIA driver 580.119.02. Clef-flash used Q8_0, not this page's Q4_K_M file.
Model / weights
Engine
Median per word
Words in 32 s
Overall accuracy
Names caught
Wrong picks
Laya Laya-BF16.gguf
llama.cpp b11495
3.9 ms
7,980
97.4%
70%
98
d1 3B d1-3B-AD-Q4_K_M.gguf
llama.cpp b11495
6.0 ms
5,306
96.5%
51%
51
Clef-flash 9B Clef-Flash-Q8_0.gguf
llama.cpp b11495
24.4 ms
1,292
97.2%
36%
2
Lev 4B interfaze-ai/lev BF16
lev serve / PyTorch
51.0 ms
626
98.9%
83%
4
Test conditions: One typed decision per word across nine Wikipedia articles containing 9,534 words, asking whether each word names a kind of centipede; requests were sent sequentially from a local Python client. llama.cpp b11495, commit 37ac63456, CUDA 12.8 release build, -ngl 99 and otherwise default settings.
Limits: Community-authored single-run comparison on a custom, highly imbalanced word-classification task. Overall accuracy is dominated by non-centipede words. 'Names caught' and 'wrong picks' are retained exactly as the author labels them because a confusion matrix was not published. Clef-flash used Q8_0, not this page's Q4_K_M file. Lev used a different PyTorch serving path, so the latency values are not a controlled model-only comparison. No peak-memory measurements or purchase-ready system configuration were reported.
For a hardware shortlist, compare the 24 GB Mac mini M4 Pro and 32 GB GMKtec K15, or browse computers by budget. These are catalog comparisons, not verified Clef-compatible configurations. Their pages link exact Amazon configurations; confirm the selected bundle, current price and stock before buying.
These scores were compiled from the linked sources. These published scores were not independently tested here. Different prompts, precision, thinking budgets and tools can change results.
Decision Index 0.2.1 · BFCL
98.8 / 100
BFCL case-exact accuracy in Cloudflare's Decision Index 0.2.1 typed-decision harness. This is not standard generative BFCL or BFCL v4 and is kept separate.
Benchmark efficiency: 16.356 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: Cloudflare-reported BFCL case-exact accuracy in its internal Decision Index 0.2.1 run; the suite jointly scores all options through Clef's typed-decision interface.
This is a source-checkpoint result, not an independent test of the selected Q4_K_M file. The typed-decision protocol is not comparable with standard generative BFCL or BFCL v4. Visual inputs additionally require the repository's separate approximately 0.92 GB projector file.
Context and version evidence: The repository and file name identify the selected model. Revision hashes are shown above when recorded. Tested context length, full hardware configuration and run date are not consistently recorded in this snapshot. Only settings explicitly stated in each benchmark record are known.
The 6.041 GB figure covers model weights in this file. It excludes the context cache, runtime, operating system and any separate multimodal files. It is not a measured RAM requirement or a token-generation speed.
These complete configurations pass the initial weight-size check for the selected 6.04 GB file. Allow extra memory for context and software. Check the model’s runtime notes and your inference software before buying.
Listing evidence: 2026-09-25. Amazon.com base price visible in an owner-supplied screenshot reviewed 25 Sep 2026; capture time not supplied. Dated reference, not a live quote. Excludes card signup offers, coupons, tax and shipping.
Listing evidence: 2026-10-03. Apple US launch price announced 22 September 2026: USD 1,699. Dated manufacturer reference, not an Amazon offer; tax and shipping may differ.
Listing evidence: 2026-09-24. GMKtec, EVO-X2 / 64GB / 1TB, recorded 22 Sep 2026; exact ASIN capacity identified 24 Sep. This is a dated reference, not a current Amazon price; tax, shipping and bundle contents may differ.
Memory matches are estimates, not tested compatibility or speed. As an Amazon Associate I earn from qualifying purchases.
Decision model. llama.cpp uses /v1/systemone; text generation is unavailable. Image input requires current model support plus a compatible multimodal projector. See the Clef comparison for sources and checks. The selected weights use 6.041 GB; this excludes context cache, runtime buffers and other programs.
These are the first listed tiers that pass our reserve policy, not measured minimum requirements. System RAM and VRAM remain separate. Use the memory tool to change the reserve for your context and runner.