Compressed decision models
Collection
Third-party decision models in smaller GGUF builds from TurinTech, built with Artemis. See each card for tests and hardware. • 12 items • Updated • 12
TurinTech models, built with Artemis
d3 27B for routing, triage and typed checks on text and images on a 24 GB GPU, compressed with Artemis
| Download | GPU memory | Same answer as original |
|---|---|---|
| 14.95 GB | 16.4 GiB (original over two GPUs 54.6 GiB) | 97.1% (ggml-org Q4_K_M 98.4%) |
Run it
This is a decision model: it answers typed questions (yes or no, pick one, a score) about text or images with a
probability for each answer, in one pass, on llama-server's /v1/systemone endpoint. It does not chat. Needs llama.cpp
b11572 (12 October 2026, which added d3 support) or newer (tested at fcac190).
llama-server -hf turintech/d3-27B-Artemis-GGUF:IQ4_XS --port 8080 -c 16384 -ub 4096 -b 4096 -np 1
LM Studio: turintech/d3-27b-artemis
Request example, settings, images and LM Studio
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?"}
}
}'
Keep -ub 4096 -b 4096: d3 reads the whole prompt in one batch, so an input longer than the batch size is rejected.
-np 1 keeps one request per batch; with several slots, llama.cpp can change a d3 answer when a shorter request
shares the batch (ggml-org/llama.cpp#30402, open as of 27 October 2026). -c 16384 stops the server from reserving
memory for the full training context. To run a downloaded file:
hf download turintech/d3-27B-Artemis-GGUF d3-27B-Artemis-IQ4_XS.gguf --local-dir .
llama-server -m d3-27B-Artemis-IQ4_XS.gguf --port 8080 -c 16384 -ub 4096 -b 4096 -np 1
For images, load ggml-org's vision projector for this model and raise the batch size, since several large images can
exceed 4,096 tokens; the image results on the results page were measured this way, with the projector at revision
de96701, which the download below pins. Send images as data
URLs in a top-level images array of the /v1/systemone request.
hf download ggml-org/vllm-d3-GGUF mmproj-vllm-d3-BF16.gguf --revision de96701 --local-dir .
llama-server -m d3-27B-Artemis-IQ4_XS.gguf --port 8080 --mmproj mmproj-vllm-d3-BF16.gguf -c 16384 -ub 8192 -b 8192 -np 1
There is no Ollama package. The commands above are the tested paths; an automatic integration button is not a compatibility test.
LM Studio (0.4.26 tested, CPU engine 2.64.0) loads the Hub entry
turintech/d3-27b-artemis as a decision model with the settings
above (a 16,384-token context, batch sizes of 4,096 and one parallel session); keep the @iq4_xs label when
downloading, since without one LM Studio can pick a different file, then send decision requests to /v1/systemone on
port 1234 with the model named. The chat window and /v1/chat/completions do not run it as a decision model.
lms get turintech/d3-27b-artemis@iq4_xs --gguf
lms load turintech/d3-27b-artemis
curl http://localhost:1234/v1/systemone -H 'Content-Type: application/json' -d '{
"model": "turintech/d3-27b-artemis",
"state": "I was charged twice for the same order.",
"questions": {
"angry": {"type": "noul", "instructions": "Is the customer angry?"}
}
}'
The LM Studio test, on 25 October 2026 with headless llmster 0.0.25-1 on Linux x86-64, downloaded
d3-27B-Artemis-IQ4_XS.gguf from Hugging Face, matched its SHA-256, loaded it through a local copy of the Hub
definition and got answers to a yes or no and a choice question on /v1/systemone. It was a compatibility check with
text only; the numbers on this card come from llama-server. Of the buttons under "Use this model", llama.cpp starts
the file as a chat model, which is not supported, and LM Studio imports it without the context and batch settings
above; the commands here are the tested paths.
Results
| File | Size | Same answer as original | KL to original (lower is better) | Accuracy | Accuracy change, points | Confident answers changed |
|---|---|---|---|---|---|---|
| TurinTech IQ4_XS | 14.95 GB | 97.1% | 0.0029 | 74.85% | +0.25 | 0 |
| TurinTech Q3_K_S | 12.26 GB | 95.7% | 0.0102 | 74.30% | -0.30 | 0 |
| ggml-org Q4_K_M | 19.40 GB | 98.4% | 0.0014 | 74.65% | +0.05 | 0 |
| BF16 original | 53.81 GB | reference | reference | 74.60% | reference | reference |
The test decisions were used to choose the published files, so they favour them; bold marks the best displayed value in each column, an observed extremum, not a significance claim. ggml-org's Q4_K_M (19.40 GB) stays closer to the original than our files on these decisions. On the 1,000 train-split decisions our files lose 1.6 and 1.7 points of accuracy, and both ranges lie entirely below zero, where the Q4_K_M loses 0.4. In the GB10 run on the results page, ggml-org's Q8_0 (28.60 GB) gave the same answer as the original on 99.4% of the test decisions, its Q4_K_M on 98.0% and our IQ4_XS on 97.2%. Every number in the table has a 95% range on the results page. Full results, memory and method.
Files
| File | Size | Peak GPU memory | Use it for |
|---|---|---|---|
d3-27B-Artemis-IQ4_XS.gguf |
14.95 GB | 16.4 GiB | Default |
d3-27B-Artemis-Q3_K_S.gguf |
12.26 GB | 14.1 GiB | Smaller: 16 GB GPUs |
mmproj-vllm-d3-BF16.gguf |
931 MB | Vision companion (mmproj) from ggml-org/vllm-d3-GGUF: pass it with --mmproj for image inputs |
Images need ggml-org's projector mmproj-vllm-d3-BF16.gguf with --mmproj and -ub 8192 -b 8192 (Run it); over the 330-image run the GPU peak was 19.0 GiB for IQ4_XS and 16.7 GiB for Q3_K_S. The original needs 54.6 GiB for text, the total over the two RTX 5090s it was split across. Peak GPU memory measured on NVIDIA RTX 5090, llama.cpp fcac190 (b11609), -c 16384 -ub 4096 -b 4096 -np 1, all layers on the GPU, the server's peak over the whole 14 October 2026 suite; images: --mmproj mmproj-vllm-d3-BF16.gguf with -ub 8192 -b 8192, the server's peak over the 330-image run of the same suite.
Pick this model
- Good for routing, triage and typed checks on text and images on a 24 GB GPU: the default file peaked at 16.4 GiB of GPU memory for text and 19.0 GiB with images on an RTX 5090, and keeps 97.1% of the original's test answers.
- Pick another file when a 16 GB GPU is the target: our Q3_K_S (12.26 GB) needs 14.1 GiB for text and keeps 95.7% of the original's answers. When the closest match to the original matters more: ggml-org's Q4_K_M (19.40 GB) gives the same answer on 98.4% of the test decisions. In the GB10 run on the results page, its Q8_0 (28.60 GB) gave the same answer on 99.4% of the test decisions and its Q4_K_M on 98.0%.
How it was made
d3 is the 27B decision model of Decision 3.0, made by the vLLM Semantic Router Team (original card, vLLM Semantic Router), built on Qwen/Qwen3.8-27B; it reads text and, with a vision projector, images. These files were built with Artemis, TurinTech's optimisation platform. Each file is a per-tensor mix of standard llama.cpp quant types, so its label is a size class (the smallest standard quant whose typical bits per weight are at least the file's; the bits per weight and tensor types are on the results page). Stock llama.cpp and LM Studio run the files; images use ggml-org's projector, which is not part of this repository. The weights are converted to GGUF and quantized for llama.cpp; nothing else changes from the original.
How it was tested
Every file in the Results table, the BF16 original included, ran through the same client on NVIDIA RTX 5090 32 GB
GPUs with llama.cpp b11609 (fcac190) at the settings above, with one slot (-np 1) and one client worker, on 14
October 2026; each file on one GPU, and the original, which does not fit one, split over two: 2,000 typed decisions
from the test split of LocalLLaMA/typed-decisions, a 300-item sample of the public Decision Index benchmarks, 330
image questions with ggml-org's BF16 projector, option reordering, both request shapes and a repeat run. The test
decisions were used to choose the published files, so they favour them. The results page has the
method, how the files were chosen, every table with its 95% ranges, the train-split decisions, the memory, a second
run on an NVIDIA GB10 on 28 and 29 October (text only, adding ggml-org's Q8_0), and the files' revisions and
checksums.
Agreement with the original
The agreement numbers from the tables on the results page, 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 rtx5090-fcac190-20261014
NVIDIA RTX 5090 GPUs, 14 October 2026.
- Reference: TurinTech's own BF16 GGUF conversion of vllm-sr/d3 (53,808,283,872 bytes, the same size as ggml-org's BF16 at revision de96701), served split over two RTX 5090s (
-ts 1,1, all 65 layers on the GPUs) on the same build and settings; our records do not hold its SHA-256 or the base model revision. Its answer probabilities on this machine and build are the reference for every file - Panel
test: typed-decisions test split (selection panel); dataset: LocalLLaMA/typed-decisions; split: test; cases: 400; decisions: 2,000; text SHA-256: 4f294f218ea1da27f3efef936359389c62ea4d3973a41457732990f1d31b647c - Panel
heldout: typed-decisions train split, 200 cases drawn outside the calibration window (seed 20261010); dataset: LocalLLaMA/typed-decisions; split: train; cases: 200; decisions: 1,000 - Panel
decision_index: Decision Index stratified sample (ANLI, ARC-Challenge, ARC-Easy, BANKING77, CLINC150 with out-of-scope, HellaSwag, MMLU, MMLU-Pro, When2Call, WinoGrande; 30 items each); dataset: ANLI, ARC, BANKING77, CLINC150, HellaSwag, MMLU, MMLU-Pro, When2Call and WinoGrande (300-item stratified sample); cases: 300; decisions: 300; text SHA-256: 9cb4301cc7d7d96dde9663263c53020912ad628201e5609ab2d65bf16ebb652e - Panel
images: 330 image questions (210 BLINK validation, 60 CV-Bench, 60 RealWorldQA) with ggml-org's BF16 vision projector; dataset: BLINK, CV-Bench, RealWorldQA; split: per source; cases: 330; decisions: 330; text SHA-256: 46d96d8d9556847a4fd2aecfa930fc47accc15995b0d1f0d83928dee413cf5b1 - Tool: llama-server
/v1/systemone, which returns a probability for every allowed answer, with all of a case's questions in one request; image questions carry their image as a data URL in the request. Same answer: the file's most likely answer is the original's. Mean KL: KL divergence from the original's answer probabilities to the file's, averaged over decisions. Accuracy: against the panel's labels, which are synthetic for typed-decisions and each source's own answer for the Decision Index sample and the images. Accuracy change: the file's accuracy minus the original's on the same decisions, in points, exact to two decimals on 2,000 decisions and to one on 1,000. Confident answers changed: decisions where the original gave an answer with more than 80% probability and the file chose another. Ranges on the results page are 95% case-bootstrap intervals (1,000 resamples, seed 20261010) - llama.cpp build: b11609 (commit fcac19079f3f2985fe74ed845f0880e7e9a8030a)
- Command:
llama-server -m <file.gguf> -c 16384 -ub 4096 -b 4096 -np 1 -ngl 99(the BF16 original adds-ts 1,1over two GPUs) with our decision suite as the client, one worker, one request at a time; image questions with-ub 8192 -b 8192 --mmproj mmproj-vllm-d3-BF16.gguf(ggml-org's BF16 projector, 931,146,240 bytes, SHA-256 e196a47584d33044d1f9b4aefd25326e35ec1388118da31160e383c0f46f13d2) - Hardware: NVIDIA GeForce RTX 5090 32 GB GPUs (driver 580.167.08): each file on one GPU, the BF16 original over two
- Settings: 16,384-token context, batch 4,096 (8,192 with images), one slot, all layers on the GPU
- Measured by: TurinTech measured every file in these tables, including ggml-org's Q4_K_M at revision de96701, with the same build and settings. Our files were the published files under their earlier size names (the same bytes, SHA-256 below); the earlier card's closeness data records them at revision beef0a18. Our records do not hold the SHA-256 of the ggml-org Q4_K_M copy this run measured
- Selection: the test decisions were used to choose the published files (a quick check on 100 of the test cases ranked nine candidate files of ours, then the full suite ran on the two published), so they are not an independent test; the verdict that accepted the files also required train-split accuracy within 2 points of the original, and 170 of the 1,000 train-split decisions were inside the calibration text of every file; the run measured the Decision Index sample and the images on both files, and no record says whether they informed the choice
- Source, on the results page: the table columns "Same answer as original, % of 2,000", "KL to original", "Accuracy, %", "Accuracy change, points", "Confident answers changed, of 1,109", "Same answer as original, % of 1,000", "Confident answers changed, of 558", "Same answer as original, %", "Accuracy, % of 300", "Image accuracy, % of 330", "Confident answers changed"
- Notes: The same two files of ours, under their current names, were measured again on the GB10 (protocol gb10-20261028), where the original was ggml-org's BF16 at revision de96701.
- Comparison files: ggml-org:
ggml-org/vllm-d3-GGUFatde9670136f5d,vllm-d3-Q4_K_M.gguf
Protocol rtx5090-fcac190-20261014, panel test (typed-decisions test split (selection panel)):
| File | Repository | Measured by | Size (GB) | Size (bytes) | Same answer as BF16 (%) | Mean KL to BF16 (nats) | Accuracy (%) | Accuracy change (points) | Confident answers changed |
|---|---|---|---|---|---|---|---|---|---|
| BF16 original | vllm-sr/d3 |
reference | 53.81 | 53808283872 | reference | reference | 74.60 | reference | reference |
d3-27B-Artemis-IQ4_XS.gguf |
turintech/d3-27B-Artemis-GGUF at beef0a18c1fe |
TurinTech | 14.95 | 14949286144 | 97.1 | 0.0029 | 74.85 | +0.25 | 0 |
d3-27B-Artemis-Q3_K_S.gguf |
turintech/d3-27B-Artemis-GGUF at beef0a18c1fe |
TurinTech | 12.26 | 12255531264 | 95.7 | 0.0102 | 74.30 | -0.30 | 0 |
vllm-d3-Q4_K_M.gguf |
ggml-org/vllm-d3-GGUF at de9670136f5d |
TurinTech | 19.40 | 19403686112 | 98.4 | 0.0014 | 74.65 | +0.05 | 0 |
Protocol rtx5090-fcac190-20261014, panel heldout (typed-decisions train split, 200 cases drawn outside the calibration window (seed 20261010)):
| File | Repository | Measured by | Size (GB) | Size (bytes) | Same answer as BF16 (%) | Mean KL to BF16 (nats) | Accuracy (%) | Accuracy change (points) | Confident answers changed |
|---|---|---|---|---|---|---|---|---|---|
| BF16 original | vllm-sr/d3 |
reference | 53.81 | 53808283872 | reference | reference | 76.5 | reference | reference |
d3-27B-Artemis-IQ4_XS.gguf |
turintech/d3-27B-Artemis-GGUF at beef0a18c1fe |
TurinTech | 14.95 | 14949286144 | 96.8 | 0.0029 | 74.9 | -1.6 | 0 |
d3-27B-Artemis-Q3_K_S.gguf |
turintech/d3-27B-Artemis-GGUF at beef0a18c1fe |
TurinTech | 12.26 | 12255531264 | 95.1 | 0.0091 | 74.8 | -1.7 | 0 |
vllm-d3-Q4_K_M.gguf |
ggml-org/vllm-d3-GGUF at de9670136f5d |
TurinTech | 19.40 | 19403686112 | 97.9 | 0.0016 | 76.1 | -0.4 | 0 |
Protocol rtx5090-fcac190-20261014, panel decision_index (Decision Index stratified sample (ANLI, ARC-Challenge, ARC-Easy, BANKING77, CLINC150 with out-of-scope, HellaSwag, MMLU, MMLU-Pro, When2Call, WinoGrande; 30 items each)):
| File | Repository | Measured by | Size (GB) | Size (bytes) | Same answer as BF16 (%) | Accuracy (%) |
|---|---|---|---|---|---|---|
| BF16 original | vllm-sr/d3 |
reference | 53.81 | 53808283872 | reference | 88.7 |
d3-27B-Artemis-IQ4_XS.gguf |
turintech/d3-27B-Artemis-GGUF at beef0a18c1fe |
TurinTech | 14.95 | 14949286144 | 98.7 | 89.3 |
d3-27B-Artemis-Q3_K_S.gguf |
turintech/d3-27B-Artemis-GGUF at beef0a18c1fe |
TurinTech | 12.26 | 12255531264 | 95.7 | 89.0 |
vllm-d3-Q4_K_M.gguf |
ggml-org/vllm-d3-GGUF at de9670136f5d |
TurinTech | 19.40 | 19403686112 | 99.0 | 89.7 |
Protocol rtx5090-fcac190-20261014, panel images (330 image questions (210 BLINK validation, 60 CV-Bench, 60 RealWorldQA) with ggml-org's BF16 vision projector):
| File | Repository | Measured by | Size (GB) | Size (bytes) | Same answer as BF16 (%) | Mean KL to BF16 (nats) | Accuracy (%) | Confident answers changed |
|---|---|---|---|---|---|---|---|---|
| BF16 original | vllm-sr/d3 |
reference | 53.81 | 53808283872 | reference | reference | 75.2 | reference |
d3-27B-Artemis-IQ4_XS.gguf |
turintech/d3-27B-Artemis-GGUF at beef0a18c1fe |
TurinTech | 14.95 | 14949286144 | 98.5 | 0.0026 | 75.2 | 0 |
d3-27B-Artemis-Q3_K_S.gguf |
turintech/d3-27B-Artemis-GGUF at beef0a18c1fe |
TurinTech | 12.26 | 12255531264 | 95.8 | 0.0086 | 74.8 | 0 |
vllm-d3-Q4_K_M.gguf |
ggml-org/vllm-d3-GGUF at de9670136f5d |
TurinTech | 19.40 | 19403686112 | 97.6 | 0.0017 | 73.9 | 0 |
Protocol gb10-20261028
One NVIDIA GB10, 28 and 29 October 2026.
- Reference: ggml-org's BF16 GGUF
vllm-d3-BF16.gguf(53,808,283,872 bytes, SHA-256 e04fd64d983231e6ba922a3b260f2b73fde727ccb9f5c8d4b735e14048171931) of vllm-sr/d3, from ggml-org/vllm-d3-GGUF revision de96701, served whole on this machine on 28 October. Its answer probabilities on this machine and build are the reference for every file - Panel
test: typed-decisions test split (selection panel); dataset: LocalLLaMA/typed-decisions; split: test; cases: 400; decisions: 2,000; text SHA-256: 4f294f218ea1da27f3efef936359389c62ea4d3973a41457732990f1d31b647c - Panel
decision_index: Decision Index stratified sample (ANLI, ARC-Challenge, ARC-Easy, BANKING77, CLINC150 with out-of-scope, HellaSwag, MMLU, MMLU-Pro, When2Call, WinoGrande; 30 items each); dataset: ANLI, ARC, BANKING77, CLINC150, HellaSwag, MMLU, MMLU-Pro, When2Call and WinoGrande (300-item stratified sample); cases: 300; decisions: 300; text SHA-256: 9cb4301cc7d7d96dde9663263c53020912ad628201e5609ab2d65bf16ebb652e - Tool: llama-server
/v1/systemone, which returns a probability for every allowed answer, with all of a case's questions in one request. Same answer: the file's most likely answer is the original's. Mean KL: KL divergence from the original's answer probabilities to the file's, averaged over decisions. Accuracy: against the panel's labels, which are synthetic for typed-decisions and each source's own answer for the Decision Index sample. Accuracy change: the file's accuracy minus the original's on the same decisions, in points, exact to two decimals on 2,000 decisions and to one on 1,000. Confident answers changed: decisions where the original gave an answer with more than 80% probability and the file chose another. Ranges on the results page are 95% case-bootstrap intervals (1,000 resamples, seed 20261010) - llama.cpp build: b11609 (commit fcac19079f3f2985fe74ed845f0880e7e9a8030a), built for CUDA 13.0 (architecture 121)
- Command:
llama-server -m <file.gguf> -c 16384 -ub 4096 -b 4096 --parallel 1 -ngl 99 -t 16with our decision suite as the client, one worker (one slot because of llama.cpp issue #30402); text only, no image pass - Hardware: one NVIDIA GB10 (HP ZGX Nano, 128 GB of memory shared by the CPU and GPU, driver 580.142), one file on disk at a time
- Settings: 16,384-token context, batch 4,096, one slot, all layers on the GPU
- Measured by: TurinTech downloaded every file, ours at revision 6ec98bd and ggml-org's BF16, Q4_K_M and Q8_0 at revision de96701, checked each by SHA-256 against the Hub listing, and measured all of them with the same build, settings and machine, each file against the saved run of the original
- Selection: the test decisions were used to choose the published files, so they are not an independent test; the Decision Index sample is as described under protocol rtx5090-fcac190-20261014; images were not measured here, because the image pass with the original froze the machine twice and no further image pass ran in this campaign
- Source, on the results page: the table columns "Same answer as original, % of 2,000, GB10", "KL to original, GB10", "Accuracy, %, GB10", "Accuracy change, points, GB10", "Confident answers changed, of 1,107", "Same answer as original, %, GB10", "Accuracy, % of 300, GB10"
- Comparison files: ggml-org:
ggml-org/vllm-d3-GGUFatde9670136f5d,vllm-d3-Q4_K_M.gguf, SHA-256 a5a341e8611dc5d916c8f4c693dbef4a68e57eb17d37914fd56dcb5a9221eb2d; ggml-org:ggml-org/vllm-d3-GGUFatde9670136f5d,vllm-d3-Q8_0.gguf, SHA-256 82130d83cb6f31d43a742a08f737a78c087e408776115b6f28e5d3bc27617805
Protocol gb10-20261028, panel test (typed-decisions test split (selection panel)):
| File | Repository | Measured by | Size (GB) | Size (bytes) | Same answer as BF16 (%) | Mean KL to BF16 (nats) | Accuracy (%) | Accuracy change (points) | Confident answers changed |
|---|---|---|---|---|---|---|---|---|---|
vllm-d3-BF16.gguf |
ggml-org/vllm-d3-GGUF at de9670136f5d |
reference | 53.81 | 53808283872 | reference | reference | 74.5 | reference | reference |
d3-27B-Artemis-IQ4_XS.gguf |
turintech/d3-27B-Artemis-GGUF at 6ec98bddb009 |
TurinTech | 14.95 | 14949286144 | 97.2 | 0.0029 | 74.6 | +0.10 | 0 |
d3-27B-Artemis-Q3_K_S.gguf |
turintech/d3-27B-Artemis-GGUF at 6ec98bddb009 |
TurinTech | 12.26 | 12255531264 | 95.9 | 0.0102 | 74.3 | -0.20 | 0 |
vllm-d3-Q4_K_M.gguf |
ggml-org/vllm-d3-GGUF at de9670136f5d |
TurinTech | 19.40 | 19403686112 | 98.0 | 0.0014 | 74.6 | +0.10 | 0 |
vllm-d3-Q8_0.gguf |
ggml-org/vllm-d3-GGUF at de9670136f5d |
TurinTech | 28.60 | 28595765472 | 99.4 | 0.0001 | 74.5 | +0.00 | 0 |
Protocol gb10-20261028, panel decision_index (Decision Index stratified sample (ANLI, ARC-Challenge, ARC-Easy, BANKING77, CLINC150 with out-of-scope, HellaSwag, MMLU, MMLU-Pro, When2Call, WinoGrande; 30 items each)):
| File | Repository | Measured by | Size (GB) | Size (bytes) | Same answer as BF16 (%) | Accuracy (%) |
|---|---|---|---|---|---|---|
vllm-d3-BF16.gguf |
ggml-org/vllm-d3-GGUF at de9670136f5d |
reference | 53.81 | 53808283872 | reference | 89.0 |
d3-27B-Artemis-IQ4_XS.gguf |
turintech/d3-27B-Artemis-GGUF at 6ec98bddb009 |
TurinTech | 14.95 | 14949286144 | 98.3 | 89.3 |
d3-27B-Artemis-Q3_K_S.gguf |
turintech/d3-27B-Artemis-GGUF at 6ec98bddb009 |
TurinTech | 12.26 | 12255531264 | 95.7 | 88.7 |
vllm-d3-Q4_K_M.gguf |
ggml-org/vllm-d3-GGUF at de9670136f5d |
TurinTech | 19.40 | 19403686112 | 99.3 | 89.7 |
vllm-d3-Q8_0.gguf |
ggml-org/vllm-d3-GGUF at de9670136f5d |
TurinTech | 28.60 | 28595765472 | 99.7 | 88.7 |
File checksums
| File | Bytes | sha256 |
|---|---|---|
d3-27B-Artemis-IQ4_XS.gguf |
14,949,286,144 | 4ae3fc00c3f7e7475796263a2b2c6311d2090f07f1dbcae9a175faa215366568 |
d3-27B-Artemis-Q3_K_S.gguf |
12,255,531,264 | 34c640d250afb383a63aee495f2c0d9ab01d8a1f309789ffd35f3f7e298e3224 |
Artemis is TurinTech's optimisation platform: teams use it to make their own models smaller and faster, or to optimise their code.
All our models · LinkedIn · Questions and requests
Apache-2.0, as released by vLLM Semantic Router Team (licence); see LICENSE and NOTICE. An independent TurinTech release, not affiliated with or endorsed by vLLM Semantic Router Team. 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.
- Downloads last month
- 3,418
Hardware compatibility
Log In to add your hardware
3-bit
4-bit