Discoveries / model catalog

Find a smaller model.

TurinTech models, built with Artemis. Each release documents its files, tested settings and quality trade-offs.

Explore Artemis discoveries

Two kinds of model

Chat models write answers. Decision models pick one.

Compressed LLMs

Chat and tools: Gemma, Qwen

user: What is the weather in Paris?

get_weather({ "city": "Paris", "unit": "celsius" })

Browse chat models

Compressed decision models

Typed decisions: Kev, d3, Clef

question: How should this support ticket be resolved?

refund
0.91
replace
0.07
escalate
0.02
Browse decision models

Illustrative examples of each output type.

The method behind the results

How we made these models smaller

We give Artemis something to improve, a model or an engine's code, and a goal. It explores many methods and finds something new. TurinTech checks it, and we apply it with Artemis again and again. That is how we release many models quickly, and the same workflow produced our merged pull requests.

discovery · chat models

Input

Qwen3.5 9B

Qwen BF16 · 17.92 GB

ArtemisAsk

New discovery

Input: Qwen3.5 9B

Goal: A smaller file whose answers stay close to the original.

Smaller models appear here

The search and its timing are an illustration; a real Discovery runs for hours. Bars show each file as a share of the original download. Agreement comes from each card's own tests, which are small samples; some also informed which file was chosen.Browse the models

Step 1 of 5: Ask. We give Artemis a chat model and a goal.

Compression, before and after

Smaller files, with measured trade-offs.

Compare selected releases with their original downloads. Select a model for details, including quality trade-offs.

Compressed LLMs

Compressed decision models

Compressed LLMs

Gemma 4 E4B

3.5× smaller than BF16

Google BF1614.90 GB
TurinTech file, built with Artemis4.22 GB
How close to the original
98.4% same next token
Runtime memory
3.3 GiB of GPU memory plus 4.4 GiB of system memory

File size is not memory in use. Agreement shows how often the smaller file answers like the original; it is not accuracy, and for decision models the test questions also helped choose the file. The card has the full details.

Open the Gemma 4 E4B card for its full tables and comparisons with public files

Download sizes below are not runtime memory requirements. See each card for tested settings.

Released models, recommended download sizes, and public model cards. Sizes are downloads, not runtime memory.
Model / original authorCategoryRecommended downloadGet the model
Clef
Original by Cloudflare
Quality trade-offs

Smaller variants change some answers the original was confident about. Image use needs a separate projector. Read the evaluation.

Compressed decision models
14.82 GBGGUF

Clef-Artemis-IQ4_XS.gguf

Apache-2.0

Hugging Face Ollama
Clef-Flash
Original by Cloudflare
Quality trade-offs

Quality differs between CPU and GPU. Smaller variants drift more on many-option questions. Read the evaluation.

Compressed decision models
4.94 GBGGUF

Clef-Flash-Artemis-IQ4_XS.gguf

Apache-2.0

Hugging Face Ollama
d3-27B
Original by vLLM Semantic Router Team
Quality trade-offs

Public Q4_K_M is closer on text evaluations. Use the documented single-request settings. Read the evaluation.

Compressed decision models
14.95 GBGGUF

d3-27B-Artemis-IQ4_XS.gguf

Apache-2.0

Hugging Face
d3-flash
Original by vLLM Semantic Router Team
Quality trade-offs

Public Q4_K_M is slightly closer on test decisions and images. Use the documented single-request settings. Read the evaluation.

Compressed decision models
4.87 GBGGUF

d3-flash-Artemis-IQ4_XS.gguf

Apache-2.0

Hugging Face
Gemma 4 E2B
Original by Google
Quality trade-offs

Text only. The larger public Unsloth file is closer to the original. Runtime memory has not been measured. Read the evaluation.

Compressed LLMs
2.31 GBGGUF

gemma-4-E2B-it-Artemis-Q3_K_L.gguf

Apache-2.0

Hugging Face Ollama
Gemma 4 E4B
Original by Google
Quality trade-offs

Text only. Smaller variants drift more during longer answers; see the card for comparisons with public files. Read the evaluation.

Compressed LLMs
4.22 GBGGUF

gemma-4-E4B-it-Artemis-Q4_0.gguf

Apache-2.0

Hugging Face Ollama
GLM-5.2 MaxMin108
Original by Z.ai
Quality trade-offs

Substantial world-knowledge loss versus the parent. Download every shard; this is a large model for machines with substantial memory. Read the evaluation.

Compressed LLMs
102.84 GBGGUF · 6 shards

GLM-5.2-MaxMin108-IQ2_M-fast

MIT

Hugging Face
Kev-27B
Original by Jared Palmer
Quality trade-offs

Public Q6_K has higher decision agreement. Public Q4_K_M is closer on the benchmark sample, but has lower agreement on the selection panel. Read the evaluation.

Compressed decision models
13.91 GBGGUF

Kev-27B-Artemis-IQ4_XS.gguf

Apache-2.0

Hugging Face Ollama
Kev-4B
Original by Jared Palmer
Quality trade-offs

The ggml-org Q4_K_M file preserves more accuracy on the additional evaluation. Selection-panel agreement is not accuracy. Read the evaluation.

Compressed decision models
2.54 GBGGUF

Kev-4B-Artemis-Q4_K_M.gguf

Apache-2.0

Hugging Face Ollama
Kev-9B
Original by Jared Palmer
Quality trade-offs

The ggml-org Q4_K_M file is closer to the original. The recommended variant used validation results during selection. Read the evaluation.

Compressed decision models
4.78 GBGGUF

Kev-9B-Artemis-IQ3_M.gguf

Apache-2.0

Hugging Face Ollama
pplx-decider-27B
Original by Perplexity
Quality trade-offs

Use Q4_K_S or the exact filename; BF16 and Q8_0 tags select vision projectors. Plain Q4_K_M scores higher on the Decision Index sample. Read the evaluation.

Compressed decision models
15.28 GBGGUF

pplx-decider-27B-Artemis-Q4_K_S.gguf

Apache-2.0

Hugging Face Ollama
Qwen3.5 0.8B
Original by Qwen (Alibaba Cloud)
Quality trade-offs

Text only. Tool-calling results used the same 200 tasks for file selection, so they are not an independent evaluation. Not better on GSM8K math. Read the evaluation.

Compressed LLMs
0.569 GBGGUF

Qwen3.5-0.8B-Artemis-Q5_K_M.gguf

Apache-2.0

Hugging Face Ollama
Qwen3.5 2B
Original by Qwen (Alibaba Cloud)
Quality trade-offs

Text only. Tool-calling gains against Unsloth Q4_K_M, but no significant gain against bartowski Q4_K_M. Unsloth Q4_K_M scored higher on MMLU-Pro; no GSM8K separation. Read the evaluation.

Compressed LLMs
1.395 GBGGUF

Qwen3.5-2B-Artemis-Q5_K_M.gguf

Apache-2.0

Hugging Face Ollama
Qwen3.5 4B
Original by Qwen (Alibaba Cloud)
Quality trade-offs

Text only. No significant tool-calling difference against matched public files; this does not establish equivalence. Original comparisons ran on Apple M3 Max; separate GB10 and RTX 5090 checks are on the card. Chat validation informed selection. Read the evaluation.

Compressed LLMs
2.714 GBGGUF

Qwen3.5-4B-Artemis-Q4_K_M.gguf

Apache-2.0

Hugging Face Ollama
Qwen3.5 9B
Original by Qwen (Alibaba Cloud)
Quality trade-offs

Text only. No task gain or equivalence established. The larger Unsloth UD-Q4_K_XL is closer to the original than the default file. Tested on one NVIDIA GB10. Read the evaluation.

Compressed LLMs
5.661 GBGGUF

Qwen3.5-9B-Artemis-Q4_K_M.gguf

Apache-2.0

Hugging Face Ollama

Sizes are recommended downloads in decimal GB, not GPU or system-memory requirements. For sharded models, the size includes every required shard. Model cards contain the current files, measurements and limitations.

Decision models return probabilities for typed questions; chat models generate text. Use the model card’s runtime and request settings. An Ollama link means a published TurinTech model, not a claim that every GGUF works in every app.

Compressed decision modelsCompressed LLMsIndependent conversions; upstream authors retain credit.TurinTech on GitHubRuntime: llama.cppTurinTech on Hugging Face