turintech/MiniCPM5-2B-Artemis-GGUF

TurinTech models, built with Artemis

MiniCPM5 2B for local chat, tool calls and maths, compressed with Artemis

Run it

llama-server -hf turintech/MiniCPM5-2B-Artemis-GGUF:Q4_K_S --jinja -c 8192

Ollama: ollama run turintech/minicpm5-2b-artemis  ·  LM Studio: turintech/minicpm5-2b-artemis

Needs llama.cpp b11617 or newer. Thinking is on by default; every result below was measured with it off, so turn it off per request ("chat_template_kwargs": {"enable_thinking": false}).

Request example and a local download

curl http://localhost:8080/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "messages": [{"role": "user", "content": "In two sentences, why is the sky blue?"}],
  "chat_template_kwargs": {"enable_thinking": false},
  "temperature": 0,
  "max_tokens": 512
}'

The reply is in choices[0].message.content. Tool calls use the same endpoint with a tools list and come back in choices[0].message.tool_calls. To run a downloaded file:

hf download turintech/MiniCPM5-2B-Artemis-GGUF MiniCPM5-2B-Artemis-Q4_K_S.gguf --local-dir .
llama-server -m MiniCPM5-2B-Artemis-Q4_K_S.gguf --jinja -c 8192

Results

The record of how the file was chosen does not mention these item sets (How it was tested); bold marks the best in each column. After correcting for 16 planned comparisons, two gaps hold, both over OpenBMB's own Q4_K_M: tool calls (184 against 160) and GSM8K (209 against 179). Both come partly from answer format. OpenBMB's file made no tool call 46 times, ours 23. GSM8K allowed 768 tokens and counts a cut-off reply whose last number is right; 151 of OpenBMB's replies ran out of room, 70 of ours. A stricter scoring of other GSM8K problems on GB10 (4,096 tokens, finished replies only) gave 215 of 250 against 180, with 20 and 59 replies over the limit. MMLU-Pro allowed 8 output tokens for the answer letter; a reply that began working the problem instead was cut off and counts as wrong, ours 17 times and the original 40. Full results, memory and method.

Files

Text only. Peak GPU memory measured on NVIDIA RTX 5090, llama.cpp b11617, all layers on the GPU, 8,192-token context, one slot, prompt cache off (--cache-ram 0), first 25 items of each task.

Pick this model

• Good for tool calls and maths with thinking off, in a 1.53 GB download.
• Pick another file when closeness to the original or instruction following matters more: bartowski's Q8_0 (2.68 GB) has a KL of .0015 from the original on WikiText-2, against our .0530. On IFEval every public file followed a few more prompts than ours (bartowski's Q4_K_S 468 against 458), a gap the tests did not separate.

How it was made

Built with Artemis, TurinTech's optimisation platform. The file is a per-tensor mix of standard llama.cpp quant types on a Q4_K_S base, within the size of bartowski's Q4_K_S, so stock llama.cpp, Ollama and LM Studio run it.

How it was tested

Every file in the Results table ran on one NVIDIA RTX 5090 with llama.cpp b11617, the same chat template, greedy decoding and thinking off. The first card records that this file was chosen on BFCL simple tool calls (items 0 to 199), GSM8K (test items 0 to 249) and closeness to the original on two texts. The Results table uses other item sets: BFCL live_simple, GSM8K test items 1069 to 1318, a new MMLU-Pro sample and IFEval. That record does not mention them, so it neither shows nor rules out a part for them in the choice. The results page has the method, the full tables, and the GB10 measurements that the Ollama and LM Studio pages quote.

Closeness to the original

The closeness numbers from the tables on the results page, in one fixed layout with the protocol behind each table.

Protocols and standard tables

• Mean KLD: the mean KL divergence between the reference's and the file's next-token distributions, computed and averaged as each protocol's Tool line says. Lower is closer.
• 99th-percentile KLD: the 99th percentile of the per-token values. Lower is closer.
• Same top p: llama.cpp's name for top-1 agreement, the share of scored tokens where the file and the reference pick the same most likely token. Higher is closer.
• PPL and PPL ratio: the file's perplexity on the text, and that perplexity divided by the reference's (1 means the same). They measure prediction of the text, not closeness.

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

Protocol rtx5090-2026-10-27

One NVIDIA RTX 5090, 27 October 2026.

• Reference: BF16 GGUF converted with llama.cpp b11617 from openbmb/MiniCPM5-2B revision f97400052a43d642bbc6e9975e2397e3ae6a6b52, with its beginning-of-sequence setting set to off as in the original (5,039,005,088 bytes, SHA-256 b6bdbcdeb5bd76aee3251d40de4af8858de50e19028d19a24ee379280a0a4545); its saved log-probabilities for each text are the reference for every file
• Text wiki: WikiText-2 test; dataset: Salesforce/wikitext at b08601e04326c79dfdd32d625aee71d232d685c3; config: wikitext-2-raw-v1; split: test; chunks: 60; context (tokens): 512; text SHA-256: bbf94c53a05abe9ee670d3b6343608095822c85e26de37c70b24fc571964574a
• Text chat: OpenAssistant validation conversations (English threads as User/Assistant text); dataset: OpenAssistant/oasst1 at fdf72ae0827c1cda404aff25b6603abec9e3399b; split: validation; chunks: 60; context (tokens): 512; text SHA-256: 0a3499a7eaead8f4a40655a868e7faee1ee664e258d09a77cd8f3d97216fc8b7
• Tool: llama-perplexity --kl-divergence against the reference's saved log-probabilities, chunks of 512 tokens; Mean KLD and Same top p are llama.cpp's own summary lines
• llama.cpp build: b11617 (commit 14310190cd0a0c831fe949ac8cbedd8abdc90695)
• Command: llama-perplexity -m <file.gguf> -f <text> --kl-divergence-base <reference-text.kld> --kl-divergence -c 512 --chunks 60 -t 16 -ngl 99; the reference logs were written with the same command without --kl-divergence
• Hardware: one NVIDIA RTX 5090 (AMD EPYC 7V13 host)
• Settings: 16 threads, all layers on the GPU; the KL command sets no flash attention or KV cache flags, so llama-perplexity's defaults applied (the flash attention and F16 KV settings in the results page are for the task runs)
• KL unit: nats (natural log)
• Measured by: TurinTech downloaded the bartowski and OpenBMB files at the revisions below, checked each by SHA-256 and measured every file in this table with the same reference, binary and machine
• Selection: the two texts were used in choosing this file, so these columns are not independent tests
• Source, on the results page: the table columns "KL wiki / chat", "Same next token, wiki / chat"
• Comparison files: bartowski: bartowski/MiniCPM5-2B-GGUF at 31fee0334c4e, MiniCPM5-2B-Q4_K_S.gguf, SHA-256 2588c13c110b78da40f2ece9cfb0fe3d190116752eb1b169e0f5052d0fddd2a1; OpenBMB: openbmb/MiniCPM5-2B-GGUF at 2079a22f3bea, MiniCPM5-2B-Q4_K_M.gguf, SHA-256 ec2d5801640099e97d8d7e8003ad4d81f336e757811f03a26173dddf386602fd; bartowski: bartowski/MiniCPM5-2B-GGUF at 31fee0334c4e, MiniCPM5-2B-IQ4_NL.gguf, SHA-256 a2bd3ddc534368d5b42b0aa00636665d169e6838e86e12e892445c959b9fce41; bartowski: bartowski/MiniCPM5-2B-GGUF at 31fee0334c4e, MiniCPM5-2B-Q4_K_M.gguf, SHA-256 967d467b5a73f2def7f339f8ad724a984fe1caa323006577cd9c26adb755f0de; bartowski: bartowski/MiniCPM5-2B-GGUF at 31fee0334c4e, MiniCPM5-2B-Q8_0.gguf, SHA-256 a62c4b475bb5998391d520d29f2617dbdc197a46b9dcca9ca693a96bfdde5b51

Protocol rtx5090-2026-10-27, text wiki (WikiText-2 test):

Protocol rtx5090-2026-10-27, text chat (OpenAssistant validation conversations (English threads as User/Assistant text)):

File checksums

---

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 OpenBMB (licence); see LICENSE and NOTICE. An independent TurinTech release, not affiliated with or endorsed by OpenBMB. 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.