turintech/Clef-Flash-Artemis-GGUF

TurinTech models, built with Artemis

Cloudflare's Clef-Flash for typed decisions about text or images on a CPU or a small GPU. It answers typed questions with a probability for each answer, in one pass.

• 4.94 GB download for the default file, Clef-Flash-Artemis-IQ4_XS.gguf.
• Peak GPU memory 5.5 GiB on an NVIDIA A100; 7.17 GiB of RAM on an AMD EPYC 7742 CPU.
• Same answer as the original on 95.0% of 2,000 test decisions on an AMD Ryzen 9 9950X CPU, against 89.5% for bartowski's IQ3_M (4.85 GB). These decisions also chose the file.

Run it

This is a decision model: send typed questions to llama-server's /v1/systemone endpoint, not to chat. Needs llama.cpp from 9 October 2026 or newer. Tested at commit 1e6f04a, and the Artemis IQ4_XS and Artemis IQ3_XXS files also on release b11572 (12 October 2026), which gave identical answers on a 100-decision check.

llama-server -hf turintech/Clef-Flash-Artemis-GGUF:IQ4_XS --port 8080 -c 16384 -ub 4096 -b 4096

Request example, images, Ollama and LM Studio

llama-server -m Clef-Flash-Artemis-IQ4_XS.gguf --port 8080 -c 16384 -ub 4096 -b 4096

Tested on an Apple M3 Max (36 GB) with llama.cpp Metal (b11617): every file here loads and answers with the command above.

Keep -ub 4096 -b 4096. At the default of 512, /v1/systemone rejects inputs longer than 512 tokens. That is true of every Clef GGUF, public ones included. -c 16384 stops the server from reserving memory for the full training context.

curl http://localhost:8080/v1/systemone -H 'Content-Type: application/json' -d '{
  "state": "My Stripe integration keeps failing. Please help ASAP.",
  "questions": {"urgency": {"type": "noul", "instructions": "Does this message express urgency?"}}
}'

The request and response follow the Jev/SystemOne format described on the original model card.

Recommended Ollama package

echo "My Stripe integration keeps failing. Please help ASAP." | ollama run turintech/clef-flash-artemis:4.9gb --ask "Does this message express urgency?"

This verified tag selects Clef-Flash-Artemis-IQ4_XS.gguf. Ollama 0.40.4 was tested; include "model": "turintech/clef-flash-artemis:4.9gb" in the decision request to port 11434. The explicit size tag selects the recommended file and is not an immutable revision.

Run it with Ollama

Tested with Ollama 0.40.4 on 14 October 2026, on a CPU. Download the file and import it with a 16K context:

hf download turintech/Clef-Flash-Artemis-GGUF Clef-Flash-Artemis-IQ4_XS.gguf --local-dir .
printf 'FROM ./Clef-Flash-Artemis-IQ4_XS.gguf\nPARAMETER num_ctx 16384\n' > Modelfile
ollama create clef-flash-artemis -f Modelfile
echo "My Stripe integration keeps failing. Please help ASAP." | ollama run clef-flash-artemis --ask "Does this message express urgency?"

The last command prints yes 94%. Ollama serves the same decision endpoint on its own port; add a model field:

curl http://localhost:11434/v1/systemone -H 'Content-Type: application/json' -d '{
  "model": "clef-flash-artemis",
  "state": "My Stripe integration keeps failing. Please help ASAP.",
  "questions": {"urgency": {"type": "noul", "instructions": "Does this message express urgency?"}}
}'

Without num_ctx 16384, Ollama's default context of 4,096 tokens rejects longer inputs. The numbers on this card were measured with llama-server. Ollama sets its own batch size, so the probabilities can differ slightly: 0.946 against 0.946 from llama-server for this example on the same machine. Clef reads the whole input in one batch: in our test a 3,641-token input worked through Ollama, and a 4,914-token input stopped the Ollama runner on a 30 GB machine, so keep inputs within 4,096 tokens, as with llama-server. Images through Ollama were not tested.

Questions, problems or a model to consider for a smaller version? Reply in our feedback thread.

Vision inputs

Add the vision projector from ggml-org/Clef-Flash-GGUF and raise the batch size, since several large images can exceed 4,096 tokens:

hf download ggml-org/Clef-Flash-GGUF mmproj-Clef-Flash-BF16.gguf --local-dir .
llama-server -m Clef-Flash-Artemis-IQ4_XS.gguf --port 8080 --mmproj mmproj-Clef-Flash-BF16.gguf -c 16384 -ub 8192 -b 8192

Send images as data URLs in a top-level images array of the /v1/systemone request. Our image results use the BF16 projector.

LM Studio

The model is on the LM Studio Hub as turintech/clef-flash-artemis. That entry loads it as a decision model with the settings above: a 16,384-token context and batch sizes of 4,096. Keep the @iq4_xs label when you download it, because without one LM Studio can pick a different file for your hardware:

lms get turintech/clef-flash-artemis@iq4_xs --gguf
lms load turintech/clef-flash-artemis

Then, with LM Studio's server running (port 1234 by default), send decision requests to /v1/systemone and name the model. The chat window and /v1/chat/completions do not run it as a decision model.

curl http://localhost:1234/v1/systemone -H 'Content-Type: application/json' -d '{
  "model": "turintech/clef-flash-artemis",
  "state": "I was charged twice for the same order.",
  "questions": {"angry": {"type": "noul", "instructions": "Is the customer angry?"}}
}'

Tested on 25 October 2026 with LM Studio 0.4.26 (headless llmster 0.0.25-1) and its CPU llama.cpp engine 2.64.0 on Linux x86-64. Clef-Flash-Artemis-IQ4_XS.gguf was downloaded from Hugging Face, its SHA-256 matched, it loaded through a local copy of this Hub definition and /v1/systemone answered a yes/no and a choice question. That is a compatibility check: the numbers on this card come from llama-server.

Only text input was tested in LM Studio.

Choose this model

• Best for typed decisions on text or images on a CPU or a small GPU.
• Consider another file when you can spare 6.49 GB and want answers a little closer to the original: ggml-org's Q4_K_M gives the same answer 96.5% of the time.

Files

Every file we publish for this model. Sizes are download sizes, not memory.

• With llama.cpp or a Hugging Face pull, add the file's quant label as the tag: -hf turintech/Clef-Flash-Artemis-GGUF:IQ4_XS.
• The Ollama tag runs our package, which holds the same file: ollama run turintech/clef-flash-artemis:<tag>.
• LM Studio's Hub entry turintech/clef-flash-artemis offers every model file above, without the vision projector.

Without a tag, -hf and hf.co fetch Clef-Flash-Artemis-IQ3_XXS.gguf, not the default file, so always add the tag.

How the files compare

"Saved" is the file-size reduction versus the BF16 original. "Acc. Δ" is the accuracy change against the original on the same cases, in percentage points; the other columns are defined below. Variant links download the exact GGUF; full filenames are in the commands and checksums. File labels are size classes, explained under "File names, sizes and SHA-256 checksums" at the end of the card.

"Same answer as the original" is how often a file picks the same option as the BF16 original on 2,000 test decisions. "Confident answers changed" counts answers the original was more than 80% sure of that the file changed. These numbers are on CPU; on GPU see "Same files on other hardware".

Test and held-out results for Artemis IQ3_XXS and public IQ3_M (4.85 GB) are in the tables. The public IQ3_M is closer on the benchmark sample. The usual public files, on the same tests: the Q4_K_M from ggml-org (6.49 GB) gives the same answer 96.5% of the time with +0.3 accuracy, and the IQ3_M from bartowski (4.85 GB) 89.5% with -2.0.

Results

Our default file against the public files nearest in size and the original, on 2,000 typed decisions from the test split, run on a CPU. These decisions were also used to choose the file.

The public IQ3_M is a little smaller and further from the original. The public Q4_K_M is larger and closer. How it was tested.

Typed-decision accuracy on the card's selection panel, with the measured nearby-size alternatives and BF16 original. These comparisons informed file selection and do not establish performance on unseen inputs. Hardware, runtime and other evaluation panels are listed under How it was tested.

Evaluation details and limitations

How it was tested

Compatibility notes

This is a decision model. Run it on llama-server's /v1/systemone endpoint and keep -ub 4096 -b 4096, as in Run it, or import it into Ollama as in Run it with Ollama. The llama.cpp button under "Use this model" starts it as a chat model, which is not a supported way to run it. The LM Studio button imports the file without the context and batch settings above; use the Hub entry in LM Studio instead. The Ollama button runs it as a decision model; add the :IQ4_XS tag for the recommended file, because without a tag it can download a smaller one.

Recommendation notes

Download Clef-Flash-Artemis-IQ4_XS.gguf (4.94 GB, peak GPU memory 5.5 GiB on an NVIDIA A100) and start it with the settings below, then send requests as shown in Run it. It gives the same answer as the original on 95.0% of 2,000 test decisions on CPU (94.1% on an NVIDIA A100) and changed no answer the original was more than 80% sure of. Go smaller only if memory is tight. The public Q4_K_M from ggml-org is a little closer to the original (96.5%) at 6.49 GB.

These files are per-tensor mixes, and their labels are size classes (see the file names table at the end). Use the exact filename or the :IQ4_XS tag; without a tag, app buttons may select a different file.

Size and quality trade-offs

Cloudflare/clef-flash is 18.2 GB and needs 16.5 GiB of GPU memory, more than a 16 GB card has. The default file here is 4.94 GB, 73% smaller, and needs 5.5 GiB (measured on an NVIDIA A100). It gives the same answer as the original on 95.0% of 2,000 test decisions on CPU (94.1% on the A100), has measured accuracy changes within one point (-0.1 on CPU, -1.0 on the A100), and changed no answer the original was more than 80% sure of on any machine we tested.

The separate train-split results are labelled held-out in the tables. Their independence from calibration and variant selection has not been verified for every file.

Test results informed candidate selection, so they do not independently measure performance on new inputs. Read them alongside the separate train-split results and public-file comparisons. Agreement measures whether a compressed model keeps the original's answers, including its mistakes. The typed-decisions accuracy figures compare answers with teacher-generated labels.

Every file was compared with the BF16 original and with the public builds, all at identical llama.cpp settings, on:

• 2,000 typed decisions (questions with a fixed answer type: yes/no, pick one, a score) from the test split of LocalLLaMA/typed-decisions: 400 cases with about five questions each,
• 1,000 further decisions from the dataset's train split, reported separately below,
• a 300-item stratified sample of the public Decision Index benchmarks (MMLU, MMLU-Pro, ARC, WinoGrande, HellaSwag, ANLI, BANKING77, CLINC150, When2Call). This is a sample for comparing builds, not a Decision Index score,
• 330 image questions from CV-Bench, RealWorldQA and BLINK, with the same vision projector for every file,
• option shuffling, the longest inputs, many-option questions, two request shapes and repeat runs.

Measured with llama.cpp 1e6f04a. Text ran on an AMD Ryzen 9 9950X (CPU only) with -ub 4096 -b 4096 and llama-server's defaults otherwise, not the -c 16384 above: the server log shows 4 slots with a 262,144-token context each. Images ran on an NVIDIA A100 80 GB with -c 16384 -ub 8192 -b 8192 --parallel 2 and the BF16 projector. Public files: ggml-org/Clef-Flash-GGUF (4a7a08c text, 6b4fe4c images) and bartowski/Cloudflare_clef-flash-GGUF (5fcdd9b).

For every file, shuffling the answer options changed no answers and a repeat run gave identical answers. Questions with 10 or more options are harder for every compressed file, more so for our smaller ones (the Artemis IQ3_XXS file gives the same answer on 83.7% of them, the public Q4_K_M on 95.3%). Clef-Flash answers a request's questions jointly, so sending them one per request changes 8.2% of the original's own decisions (measured on an A100); every number here uses the same request shape. Small differences in this sample should not be treated as general improvements.

Results on other hardware

Same files on other hardware

The same files on four GPUs (llama.cpp 1e6f04a, CUDA, all layers on the GPU). Each column comes from one machine; a blank cell was not run on that machine. On the RTX 3060 the original does not fit in 12 GB and ran partly on the CPU; the Artemis IQ4_XS file runs fully on the card with 5.2 GiB.

On GPU the default Artemis IQ4_XS file sits about one point further from the original than on CPU (same answer 94.1 to 94.2% against 95.0%, accuracy -0.5 to -1.0 against -0.1); the public Q4_K_M files change by less. No file changed a confident answer on any GPU.

Memory

Peak memory of llama-server on CPU (AMD EPYC 7742, -c 16384 -ub 4096 -b 4096 --parallel 1):

By default llama.cpp keeps an extra copy of some 4-bit weights on CPU for speed; --no-repack turns it off.

Agreement with the original

The agreement numbers from the tables above, in one fixed layout with the protocol behind each table.

Protocols and standard tables

• Same answer as BF16: the share of decisions where the file picks the same answer as the BF16 original, mistakes included. Higher is closer.
• Mean KL to BF16: KL divergence between the original's and the file's answer probabilities, as the protocol describes. Lower is closer.
• Accuracy: the share of decisions that match the panel's labels. Accuracy change: the file's accuracy minus the original's, in points.
• Confident answers changed: answers the original gave with more than 80% probability that the file changed. Lower is closer.

n/a means the card's tables or our run records do not have that value.

Protocol cpu-ryzen9950x-20261011

AMD Ryzen 9 9950X, 11 and 12 October 2026.

• Reference: BF16 GGUF converted by TurinTech from the Cloudflare/clef-flash safetensors with the pinned llama.cpp converter (SHA-256 c433d75a632ff43bdb191d1f3596a1f6e9717deb2e352eacadfd67b3d5710068)
• Panel test: typed-decisions test split (all 400 cases); dataset: LocalLLaMA/typed-decisions; split: test; cases: 400; decisions: 2000; text SHA-256: 4f294f218ea1da27f3efef936359389c62ea4d3973a41457732990f1d31b647c
• Panel heldout: typed-decisions train split, 200 cases outside the first 64 x 512-token imatrix calibration window (labelled "held-out" on the card); dataset: LocalLLaMA/typed-decisions; split: train; cases: 200; decisions: 1000; text SHA-256: 46a58d63edfd86e23229c78afe8b72307bb4ca9fb0e8df180cabb3c67ec9dcd5 (whole train parquet; the 200-case subset has no hash of its own)
• Panel decision_index: Decision Index sample: 300 items, 30 from each of MMLU, MMLU-Pro, ARC-Easy, ARC-Challenge, WinoGrande, HellaSwag, ANLI, BANKING77, CLINC150+OOS and When2Call (seed 20261010); not a Decision Index score; dataset: multiple: cais/mmlu, TIGER-Lab/MMLU-Pro, allenai/ai2_arc, allenai/winogrande, Rowan/hellaswag, facebook/anli, nvidia/When2Call, plus BANKING77 and CLINC150 from GitHub; split: test or validation per source; cases: 300; decisions: 300; text SHA-256: 9cb4301cc7d7d96dde9663263c53020912ad628201e5609ab2d65bf16ebb652e
• Tool: llama-server /v1/systemone, which returns a probability for every allowed answer; all of a case's questions in one request
• llama.cpp build: b11540 (commit 1e6f04a75e91da4bb66b5248f3fd4bdeff169d41)
• Command: llama-server -m <file.gguf> -ub 4096 -b 4096 with our decision suite as the client, one worker
• Hardware: AMD Ryzen 9 9950X, CPU only
• Settings: -ub 4096 -b 4096 and llama-server's defaults otherwise: the server log shows 4 slots with a 262,144-token context each, not the card's run command -c 16384
• Measured by: TurinTech measured every file in these tables, including ggml-org's Q4_K_M and bartowski's IQ3_M, with the same build and settings on the same machine
• Selection: the test results informed the choice of files, so they are not an independent test
• Source on this card: the table column "Acc. Δ"; the table rows "Same answer as BF16", "Typed-decision accuracy", "Changed answers where BF16 was over 80% sure", "Held-out: same answer as BF16", "Decision Index sample: same answer as BF16", "Decision Index sample accuracy", "KL to BF16 (lower is closer)"
• Comparison files: ggml-org: ggml-org/Clef-Flash-GGUF at 4a7a08c09bc6, Clef-Flash-Q4_K_M.gguf, SHA-256 3243a51d7a7bb205fb4fcf402a7df763f7d31e6645a9c10e5754c4395ab85b66; bartowski: bartowski/Cloudflare_clef-flash-GGUF at 5fcdd9ba6d3d, Cloudflare_clef-flash-IQ3_M.gguf, SHA-256 0204fc3f3f3dbc8ef4a978463a7cbaf5eb21a338f3b5c017dcdeda7be2116703

Protocol cpu-ryzen9950x-20261011, panel test (typed-decisions test split (all 400 cases)):

Protocol cpu-ryzen9950x-20261011, panel heldout (typed-decisions train split, 200 cases outside the first 64 x 512-token imatrix calibration window (labelled "held-out" on the card)):

Protocol cpu-ryzen9950x-20261011, panel decision_index (Decision Index sample: 300 items, 30 from each of MMLU, MMLU-Pro, ARC-Easy, ARC-Challenge, WinoGrande, HellaSwag, ANLI, BANKING77, CLINC150+OOS and When2Call (seed 20261010); not a Decision Index score):

Protocol a100-images-20261012

NVIDIA A100 80 GB, image questions, 12 October 2026.

• Reference: ggml-org's BF16 GGUF of Clef-Flash with ggml-org's BF16 vision projector mmproj-Clef-Flash-BF16.gguf (SHA-256 477598b2425d3aea55fefb895c414e04bda1ee2ec55878d78e850cb8b420d8d6)
• Panel images: 330 image questions: CV-Bench (60), RealWorldQA (60), BLINK validation (210, 14 subtasks x 15), images capped at 1.6 MP (seed 20261012); dataset: CV-Bench, RealWorldQA, BLINK-Benchmark/BLINK; split: BLINK validation; cases: 330; decisions: 330; text SHA-256: 46d96d8d9556847a4fd2aecfa930fc47accc15995b0d1f0d83928dee413cf5b1
• Tool: llama-server /v1/systemone with the vision projector, which returns a probability for every allowed answer
• llama.cpp build: b11540 (commit 1e6f04a75e91da4bb66b5248f3fd4bdeff169d41)
• Command: llama-server -m <file.gguf> --mmproj mmproj-Clef-Flash-BF16.gguf -c 16384 -ub 8192 -b 8192 --parallel 2 -ngl 99 (flags as recorded)
• Hardware: one NVIDIA A100 80 GB
• Settings: 16,384-token context, batch 8,192, two slots, all layers on the GPU, the same projector for every file
• Measured by: TurinTech measured every file in this table, including ggml-org's Q4_K_M, with the same build, settings and projector
• Source on this card: the table rows "Images: same answer as BF16", "Image accuracy"
• Comparison files: ggml-org: ggml-org/Clef-Flash-GGUF at 6b4fe4c59098, Clef-Flash-Q4_K_M.gguf, SHA-256 3626116623ca8d89641952471fac9ce12af84fb3517520029a2ac6f5146868a2

Protocol a100-images-20261012, panel images (330 image questions: CV-Bench (60), RealWorldQA (60), BLINK validation (210, 14 subtasks x 15), images capped at 1.6 MP (seed 20261012)):

Protocol a100-text-20261012

NVIDIA A100 80 GB, 12 October 2026.

• Reference: the BF16 original run on the same machine
• Panel test: typed-decisions test split (all 400 cases); dataset: LocalLLaMA/typed-decisions; split: test; cases: 400; decisions: 2000; text SHA-256: 4f294f218ea1da27f3efef936359389c62ea4d3973a41457732990f1d31b647c
• Tool: llama-server /v1/systemone, which returns a probability for every allowed answer; all of a case's questions in one request
• llama.cpp build: b11540 (commit 1e6f04a75e91da4bb66b5248f3fd4bdeff169d41)
• Command: llama-server -m <file.gguf> -c 16384 -ub 4096 -b 4096 --parallel 2 -ngl 99 (flags as recorded) with our decision suite, two workers
• Hardware: one NVIDIA A100 80 GB per file
• Settings: 16,384-token context, batch 4,096, two slots, all layers on the GPU
• Measured by: TurinTech measured every file in this column on the same kind of machine; each file is compared with the original run there
• Selection: the test results informed the choice of files, so they are not an independent test
• Source on this card: the table column "Same answer: A100 80 GB"
• Comparison files: ggml-org: ggml-org/Clef-Flash-GGUF, Clef-Flash-Q4_K_M.gguf; bartowski: bartowski/Cloudflare_clef-flash-GGUF, Cloudflare_clef-flash-Q4_K_M.gguf; bartowski: bartowski/Cloudflare_clef-flash-GGUF, Cloudflare_clef-flash-IQ3_M.gguf

Protocol a100-text-20261012, panel test (typed-decisions test split (all 400 cases)):

Protocol rtx5090-20261012

NVIDIA RTX 5090, 12 October 2026.

• Reference: the BF16 original run on the same GPU
• Panel test: typed-decisions test split (all 400 cases); dataset: LocalLLaMA/typed-decisions; split: test; cases: 400; decisions: 2000; text SHA-256: 4f294f218ea1da27f3efef936359389c62ea4d3973a41457732990f1d31b647c
• Tool: llama-server /v1/systemone, which returns a probability for every allowed answer; all of a case's questions in one request
• llama.cpp build: b11540 (commit 1e6f04a75e91da4bb66b5248f3fd4bdeff169d41)
• Command: llama-server -m <file.gguf> -c 16384 -ub 4096 -b 4096 --parallel 2 -ngl 99 (flags as recorded) with our decision suite, two workers
• Hardware: one NVIDIA RTX 5090 32 GB
• Settings: 16,384-token context, batch 4,096, two slots, all layers on the GPU
• Measured by: TurinTech measured every file in this column on the same GPU; each file is compared with the original run there
• Selection: the test results informed the choice of files, so they are not an independent test
• Source on this card: the table column "RTX 5090 32 GB"
• Comparison files: ggml-org: ggml-org/Clef-Flash-GGUF at 4a7a08c09bc6, Clef-Flash-Q4_K_M.gguf, SHA-256 3243a51d7a7bb205fb4fcf402a7df763f7d31e6645a9c10e5754c4395ab85b66; bartowski: bartowski/Cloudflare_clef-flash-GGUF at 5fcdd9ba6d3d, Cloudflare_clef-flash-Q4_K_M.gguf, SHA-256 cf747351878a0e569fd9d6681395bd723e111f63a4f6fe6e26e7563dfdfa9b0a; bartowski: bartowski/Cloudflare_clef-flash-GGUF at 5fcdd9ba6d3d, Cloudflare_clef-flash-IQ3_M.gguf, SHA-256 0204fc3f3f3dbc8ef4a978463a7cbaf5eb21a338f3b5c017dcdeda7be2116703

Protocol rtx5090-20261012, panel test (typed-decisions test split (all 400 cases)):

Protocol gb10-20261012

NVIDIA GB10, 12 October 2026.

• Reference: the BF16 original run on the same machine
• Panel test: typed-decisions test split (all 400 cases); dataset: LocalLLaMA/typed-decisions; split: test; cases: 400; decisions: 2000; text SHA-256: 4f294f218ea1da27f3efef936359389c62ea4d3973a41457732990f1d31b647c
• Tool: llama-server /v1/systemone, which returns a probability for every allowed answer; all of a case's questions in one request
• llama.cpp build: b11540 (commit 1e6f04a75e91da4bb66b5248f3fd4bdeff169d41)
• Command: llama-server -m <file.gguf> -c 16384 -ub 4096 -b 4096 --parallel 2 -ngl 99 -t 20 (flags as recorded) with our decision suite, two workers
• Hardware: one NVIDIA GB10 (128 GB unified memory)
• Settings: 16,384-token context, batch 4,096, two slots, all layers on the GPU
• Measured by: TurinTech measured every file in this column on the same machine; each file is compared with the original run there
• Selection: the test results informed the choice of files, so they are not an independent test
• Source on this card: the table column "GB10 (DGX Spark)"
• Comparison files: ggml-org: ggml-org/Clef-Flash-GGUF, Clef-Flash-Q4_K_M.gguf; bartowski: bartowski/Cloudflare_clef-flash-GGUF, Cloudflare_clef-flash-Q4_K_M.gguf; bartowski: bartowski/Cloudflare_clef-flash-GGUF, Cloudflare_clef-flash-IQ3_M.gguf

Protocol gb10-20261012, panel test (typed-decisions test split (all 400 cases)):

Protocol rtx3060-20261012

NVIDIA RTX 3060, 12 October 2026.

• Reference: the BF16 GGUF converted by TurinTech (SHA-256 c433d75a632ff43bdb191d1f3596a1f6e9717deb2e352eacadfd67b3d5710068), run on the same machine with 18 of 32 blocks on the GPU
• Panel test: typed-decisions test split (all 400 cases); dataset: LocalLLaMA/typed-decisions; split: test; cases: 400; decisions: 2000; text SHA-256: 4f294f218ea1da27f3efef936359389c62ea4d3973a41457732990f1d31b647c
• Tool: llama-server /v1/systemone, which returns a probability for every allowed answer; all of a case's questions in one request
• llama.cpp build: b11540 (commit 1e6f04a75e91da4bb66b5248f3fd4bdeff169d41)
• Command: llama-server -m <file.gguf> -c 16384 -ub 4096 -b 4096 -t 12 --parallel 2 -ngl 99 (the BF16 original with -ngl 18)
• Hardware: one NVIDIA RTX 3060 12 GB with an AMD Ryzen 9 5900X
• Settings: 16,384-token context, batch 4,096, two slots, our files and the public file fully on the GPU
• Measured by: TurinTech measured every file in this column on the same machine; each file is compared with the original run there
• Selection: the test results informed the choice of files, so they are not an independent test
• Source on this card: the table column "RTX 3060 12 GB"
• Comparison files: ggml-org: ggml-org/Clef-Flash-GGUF, Clef-Flash-Q4_K_M.gguf, SHA-256 3626116623ca8d89641952471fac9ce12af84fb3517520029a2ac6f5146868a2

Protocol rtx3060-20261012, panel test (typed-decisions test split (all 400 cases)):

File names, sizes and SHA-256 checksums

File names

File names end in a standard llama.cpp quant label, so llama.cpp, Ollama and LM Studio can pick a file by tag. Each file is a per-tensor mix of quant types, so the label is a size class: the smallest standard llama.cpp quant whose typical bits per weight are at least the file's. The general.file_type field inside the files reads Q4_K_M, the base type given to llama-quantize; the table shows the tensor types each file holds.

The earlier names were removed from the repository on 26 October 2026; the files under the names above hold the same bytes.

File checksums

Custom versions

We also make fine-tuned and further optimised versions on request: tuned on specific decisions and data, built for the target hardware (CPU servers, laptops, Apple Silicon, edge devices), or compressed from a private model. Book a call with TurinTech.

About and licence

Clef-Flash is Cloudflare's 9B decision model. The model accepts text (and optionally images) with a set of typed questions, and it returns a probability for every allowed answer in a single pass, instead of generating text. See the original model card.

This is an independent TurinTech conversion, not affiliated with or endorsed by Cloudflare. The files are standard GGUF and run on stock llama.cpp. For images, use the vision projector from ggml-org/Clef-Flash-GGUF (see below). Changes from the original: the weights are converted to GGUF and quantized for llama.cpp. See NOTICE.

Apache-2.0 (see LICENSE and NOTICE), the same as Cloudflare/clef-flash, which is post-trained from Qwen/Qwen3.5-9B.

When serving these files to others, for example through an API, an inference provider or OpenRouter, we'd appreciate a mention of TurinTech (turintech.ai) as the source of the quantization.

These files keep improving.