Knowledge
Open AI models to run yourself.
Which language models you may run yourself, what hardware they need and how good they are, including in German. The scores come from open sources (LMArena, Epoch AI, EuroEval, Hugging Face, models.dev); the selection and assessment are ours. We sell neither models nor hardware.
Updated weekly
New open models appear almost every week. That’s why we pull the scores from the sources weekly, review new models and add them to the selection where it makes sense. The rankings are those of the sources, not ours.
- models
- 36
- makers
- 17
Open AI models to run yourself.
Xiaomi · China
MiMo-V2.6-Pro
Free for commercial useRuns on: Server with several GPUs
Large Xiaomi model that accepts audio and video as well as text and images. So far it has mainly human ratings on LMArena; independent benchmarks are still missing.
- Human rating:
- 1491 · rank 1 of 226 open
- Runs on
- about 707 GB of memory at 4 bits (estimated)
- Context
- Context: 1 m tokens · Inputs: text, image, audio, video
All scores and links
LMArena by category
- Coding1532 · 1/221
- Maths1499 · 2/215
- Creative writing1465 · 1/224
- Following instructions1491 · 1/226
- Hard prompts1510 · 1/226
- Long prompts1500 · 1/205
- Multi-turn dialogue1463 · 8/224
- Building websites (WebDev)1626 · 6/58
- Understanding images (Vision)1264 · 7/49
4,075 votes · mimo-v2.6-pro
- Parameters
- 1,024.2bn, 42bn active
- Released
- September 2026
Moonshot AI · China
Kimi K3
Commercial use with conditionsRuns on: Server with several GPUs
Very large Moonshot AI model for agents, coding and long documents up to one million tokens; also accepts images. Runs only on servers with several GPUs.
- Human rating:
- 1475 · rank 2 of 226 open
- In German:
- 1489 · rank 2 of 154 open
- Epoch index:
- 157.4 · rank 1 of 127 open
- Runs on
- about 1,735 GB of memory at 4 bits
- Context
- Context: 1 m tokens · Inputs: text, image
All scores and links
LMArena by category
- Coding1507 · 3/221
- Maths1494 · 3/215
- Creative writing1456 · 4/224
- Following instructions1483 · 2/226
- Hard prompts1494 · 2/226
- Long prompts1493 · 2/205
- Multi-turn dialogue1488 · 1/224
- Building websites (WebDev)1655 · 2/58
32,795 votes · Kimi K3 (Max)
Tests according to Epoch AI
- GPQA Diamond (expert questions)90.8 %
- AIME (maths)97.2 %
- SimpleQA (factual knowledge)50.6 %
- Parameters
- 2,779.9bn
- Memory at 8 bits
- 1,561 GB
- Released
- July 2026
Licence: Commercial use allowed. If you offer the model as a service (e.g. via API) with over US$20m annual revenue, you need an agreement with Moonshot; very large products must display “Kimi K3”. Internal use is exempt.
Z.ai · China
GLM-5.3-Flash
Free for commercial useRuns on: Mac Studio with 512 GB
Smaller, faster version of GLM-5.3 with image input and one million tokens of context. With 4-bit quantisation it fits on a Mac Studio with 512 GB.
- Human rating:
- 1471 · rank 4 of 226 open
- In German:
- 1466 · rank 7 of 154 open
- Epoch index:
- 153.2 · rank 5 of 127 open
- Runs on
- about 230 GB of memory at 4 bits
- Context
- Context: 1 m tokens · Inputs: text, image
All scores and links
LMArena by category
- Coding1508 · 2/221
- Maths1505 · 1/215
- Creative writing1441 · 7/224
- Following instructions1478 · 3/226
- Hard prompts1490 · 3/226
- Long prompts1482 · 3/205
- Multi-turn dialogue1469 · 6/224
- Building websites (WebDev)1607 · 9/58
- Understanding images (Vision)1296 · 1/49
32,925 votes · glm-5.3-flash
Tests according to Epoch AI
- GPQA Diamond (expert questions)86.9 %
- AIME (maths)93.9 %
- Parameters
- 321.3bn, 18bn active
- Memory at 8 bits
- 341 GB
- Released
- August 2026
Z.ai · China
GLM-5.3
Free for commercial useRuns on: Server with several GPUs
Large Z.ai text model for agents, coding and long contexts of up to one million tokens.
- Human rating:
- 1471 · rank 3 of 226 open
- In German:
- 1493 · rank 1 of 154 open
- Runs on
- about 537 GB of memory at 4 bits
- Context
- Context: 1 m tokens · Inputs: text
All scores and links
LMArena by category
- Coding1495 · 7/221
- Maths1491 · 4/215
- Creative writing1456 · 3/224
- Following instructions1476 · 4/226
- Hard prompts1488 · 4/226
- Long prompts1481 · 5/205
- Multi-turn dialogue1472 · 4/224
- Building websites (WebDev)1622 · 7/58
22,543 votes · glm-5.3-max
- Parameters
- 753.3bn
- Memory at 8 bits
- 801 GB
- Released
- August 2026
Licence: Free for commercial use; only model-as-a-service providers with over US$10bn revenue need a security review by Z.ai.
DeepSeek · China
DeepSeek V4.1 Flash
Free for commercial useRuns on: Server with several GPUs
DeepSeek's faster line with up to one million tokens of context and image input. Only a small share of the parameters is active per request, which makes it cheaper to run than dense models of the same size.
- Human rating:
- 1463 · rank 7 of 226 open
- In German:
- 1476 · rank 3 of 154 open
- Epoch index:
- 155.6 · rank 2 of 127 open
- Runs on
- about 527 GB of memory at 4 bits (estimated)
- Context
- Context: 1 m tokens · Inputs: text, image
All scores and links
LMArena by category
- Coding1501 · 6/221
- Maths1478 · 7/215
- Creative writing1437 · 12/224
- Following instructions1473 · 6/226
- Hard prompts1482 · 6/226
- Long prompts1474 · 7/205
- Multi-turn dialogue1460 · 9/224
- Understanding images (Vision)1277 · 4/49
12,793 votes · deepseek-v4.1-flash-max
Tests according to Epoch AI
- GPQA Diamond (expert questions)86.4 %
- AIME (maths)98.3 %
- SimpleQA (factual knowledge)42.9 %
- Parameters
- 763.2bn, 13bn active
- Released
- September 2026
Moonshot AI · China
Kimi K2.6
Free for commercial useRuns on: Server with several GPUs
Large mixture-of-experts model from Moonshot AI for agents and coding tasks that also understands images and video as input.
- Human rating:
- 1455 · rank 9 of 226 open
- In German:
- 1451 · rank 12 of 154 open
- Epoch index:
- 150.9 · rank 8 of 127 open
- Runs on
- about 671 GB of memory at 4 bits
- Context
- Context: 262,000 tokens · Inputs: text, image, video
All scores and links
LMArena by category
- Coding1489 · 8/221
- Maths1475 · 9/215
- Creative writing1433 · 13/224
- Following instructions1451 · 10/226
- Hard prompts1470 · 10/226
- Long prompts1468 · 8/205
- Multi-turn dialogue1453 · 12/224
- Building websites (WebDev)1510 · 15/58
- Understanding images (Vision)1283 · 2/49
40,013 votes · kimi-k2.6
Tests according to Epoch AI
- GPQA Diamond (expert questions)87.7 %
- AIME (maths)96.1 %
- SimpleQA (factual knowledge)34.9 %
- Parameters
- 1,026.9bn, 32bn active
- Memory at 8 bits
- 595 GB
- Released
- April 2026
Licence: Modified MIT licence: free for commercial use; only products with over 100m monthly users or over US$20m monthly revenue must display “Kimi K2.6”.
Xiaomi · China
MiMo-V2.6-Flash
Free for commercial useRuns on: Mac Studio with 512 GB
The smaller version of MiMo-V2.6 with the same input types (text, image, audio, video). With 4-bit quantisation a Mac Studio with 512 GB is enough.
- Human rating:
- 1455 · rank 10 of 226 open
- In German:
- 1464 · rank 8 of 154 open
- Runs on
- about 214 GB of memory at 4 bits (estimated)
- Context
- Context: 1 m tokens · Inputs: text, image, audio, video
All scores and links
LMArena by category
- Coding1502 · 4/221
- Maths1474 · 10/215
- Creative writing1403 · 22/224
- Following instructions1456 · 8/226
- Hard prompts1479 · 8/226
- Long prompts1459 · 10/205
- Multi-turn dialogue1449 · 14/224
- Building websites (WebDev)1633 · 3/58
- Understanding images (Vision)1259 · 10/49
10,917 votes · mimo-v2.6-flash
- Parameters
- 310.8bn, 15bn active
- Released
- September 2026
DeepSeek · China
DeepSeek V4 Pro
Free for commercial useRuns on: Server with several GPUs
DeepSeek's large text model for demanding tasks such as analysis, maths and longer texts. Needs a server with a great deal of GPU memory.
- Human rating:
- 1451 · rank 11 of 226 open
- In German:
- 1458 · rank 10 of 154 open
- Epoch index:
- 155.3 · rank 3 of 127 open
- Runs on
- about 977 GB of memory at 4 bits
- Context
- Context: 1 m tokens · Inputs: text
All scores and links
LMArena by category
- Coding1469 · 15/221
- Maths1443 · 21/215
- Creative writing1446 · 6/224
- Following instructions1445 · 12/226
- Hard prompts1461 · 11/226
- Long prompts1457 · 12/205
- Multi-turn dialogue1467 · 7/224
- Building websites (WebDev)1582 · 13/58
57,594 votes · deepseek-v4-pro
Tests according to Epoch AI
- GPQA Diamond (expert questions)88.9 %
- AIME (maths)98.6 %
- SimpleQA (factual knowledge)52.9 %
- Parameters
- 1,650.5bn
- Memory at 8 bits
- 873 GB
- Released
- August 2026
NVIDIA · United States
Nemotron 3 Ultra
Free for commercial useRuns on: Server with several GPUs
Large NVIDIA text model with one million tokens of context under the permissive OpenMDW licence.
- Human rating:
- 1445 · rank 16 of 226 open
- In German:
- 1472 · rank 4 of 154 open
- Epoch index:
- 145.8 · rank 24 of 127 open
- Runs on
- about 414 GB of memory at 4 bits
- Context
- Context: 1 m tokens · Inputs: text
All scores and links
LMArena by category
- Coding1468 · 18/221
- Maths1456 · 18/215
- Creative writing1398 · 30/224
- Following instructions1418 · 26/226
- Hard prompts1452 · 16/226
- Long prompts1437 · 22/205
- Multi-turn dialogue1411 · 44/224
11,567 votes · nvidia-nemotron-3-ultra-550b-a55b-nvfp4
Tests according to Epoch AI
- GPQA Diamond (expert questions)80.5 %
- AIME (maths)86.7 %
- Parameters
- 560.5bn, 55bn active
- Memory at 8 bits
- 584 GB
- Released
- June 2026
Licence: OpenMDW 1.1: permissive licence for models without usage restrictions.
Google · United States
Gemma 4 31B
Free for commercial useRuns on: Computer with 32 GB
Compact Google model with image input under the Apache 2.0 licence. It ranks near the top of EuroEval's German comparison and runs on a computer with 32 GB.
- Human rating:
- 1443 · rank 17 of 226 open
- In German:
- EuroEval German: rank 6 of 140
- Epoch index:
- 142.9 · rank 29 of 127 open
- Runs on
- about 21 GB of memory at 4 bits
- Context
- Context: 262,000 tokens · Inputs: text, image
All scores and links
LMArena by category
- Coding1459 · 25/221
- Maths1464 · 15/215
- Creative writing1415 · 15/224
- Following instructions1433 · 16/226
- Hard prompts1448 · 22/226
- Long prompts1446 · 14/205
- Multi-turn dialogue1452 · 13/224
- Building websites (WebDev)1368 · 36/58
- Understanding images (Vision)1277 · 3/49
6,147 votes · gemma-4-31b
Tests according to Epoch AI
- GPQA Diamond (expert questions)67.7 %
- AIME (maths)73.3 %
- SimpleQA (factual knowledge)10.4 %
- Parameters
- 31.3bn
- Memory at 8 bits
- 33 GB
- Released
- April 2026
Alibaba (Qwen) · China
Qwen 3.8 27B
Free for commercial useRuns on: Computer with 32 GB
Compact model from the Qwen 3.8 series that understands images and calls tools. Quantised, it runs on a computer with 32 GB and reaches scores otherwise seen from much larger models.
- Human rating:
- 1441 · rank 19 of 226 open
- In German:
- 1436 · rank 22 of 154 open
- Epoch index:
- 151.6 · rank 7 of 127 open
- Runs on
- about 20 GB of memory at 4 bits
- Context
- Context: 262,000 tokens · Inputs: text, image
All scores and links
LMArena by category
- Coding1483 · 12/221
- Maths1457 · 17/215
- Creative writing1384 · 39/224
- Following instructions1437 · 13/226
- Hard prompts1459 · 12/226
- Long prompts1449 · 13/205
- Multi-turn dialogue1442 · 22/224
- Building websites (WebDev)1592 · 11/58
- Understanding images (Vision)1271 · 5/49
23,294 votes · qwen3.8-27b
- Parameters
- 27.8bn
- Memory at 8 bits
- 29 GB
- Released
- August 2026
Thinking Machines · United States
Inkling
Free for commercial useRuns on: Server with several GPUs
Large Thinking Machines model under the Apache 2.0 licence with image input and one million tokens of context.
- Human rating:
- 1441 · rank 18 of 226 open
- In German:
- 1447 · rank 16 of 154 open
- Epoch index:
- 148.6 · rank 13 of 127 open
- Runs on
- about 675 GB of memory at 4 bits
- Context
- Context: 1 m tokens · Inputs: text, image
All scores and links
LMArena by category
- Coding1465 · 20/221
- Maths1477 · 8/215
- Creative writing1387 · 37/224
- Following instructions1425 · 22/226
- Hard prompts1451 · 18/226
- Long prompts1434 · 25/205
- Multi-turn dialogue1437 · 24/224
- Building websites (WebDev)1414 · 29/58
40,648 votes · inkling
Tests according to Epoch AI
- GPQA Diamond (expert questions)84.3 %
- AIME (maths)88.9 %
- SimpleQA (factual knowledge)40.3 %
- Parameters
- 952.4bn, 41bn active
- Memory at 8 bits
- 857 GB
- Released
- July 2026
Alibaba (Qwen) · China
Qwen3.5 397B-A17B
Free for commercial useRuns on: Mac Studio with 512 GB
The large model of Alibaba's Qwen 3.5 series with image input. Only 17 billion parameters are active per request; quantised, it runs on a Mac Studio with 512 GB.
- Human rating:
- 1438 · rank 22 of 226 open
- In German:
- 1446 · rank 17 of 154 open
- Epoch index:
- 146.7 · rank 17 of 127 open
- Runs on
- about 281 GB of memory at 4 bits
- Context
- Context: 262,000 tokens · Inputs: text, image
All scores and links
LMArena by category
- Coding1464 · 22/221
- Maths1454 · 20/215
- Creative writing1402 · 24/224
- Following instructions1423 · 23/226
- Hard prompts1448 · 21/226
- Long prompts1442 · 20/205
- Multi-turn dialogue1446 · 16/224
- Building websites (WebDev)1400 · 32/58
- Understanding images (Vision)1263 · 8/49
93,567 votes · qwen3.5-397b-a17b
Tests according to Epoch AI
- GPQA Diamond (expert questions)81.8 %
- AIME (maths)88.9 %
- Parameters
- 403.4bn, 17bn active
- Memory at 8 bits
- 422 GB
- Released
- February 2026
Tencent · China
Hy3
Free for commercial useRuns on: Mac Studio with 512 GB
Mixture-of-experts text model from Tencent under the Apache 2.0 licence. Quantised, a Mac Studio with 512 GB is enough.
- Human rating:
- 1438 · rank 21 of 226 open
- In German:
- 1447 · rank 15 of 154 open
- Runs on
- about 210 GB of memory at 4 bits
- Context
- Context: 262,000 tokens · Inputs: text
All scores and links
LMArena by category
- Coding1465 · 21/221
- Maths1471 · 12/215
- Creative writing1402 · 23/224
- Following instructions1426 · 21/226
- Hard prompts1447 · 23/226
- Long prompts1441 · 21/205
- Multi-turn dialogue1435 · 25/224
- Building websites (WebDev)1507 · 17/58
12,208 votes · hy3
- Parameters
- 298.8bn, 21bn active
- Memory at 8 bits
- 318 GB
- Released
- July 2026
Google · United States
Gemma 4 26B-A4B
Free for commercial useRuns on: Computer with 32 GB
Mixture-of-experts variant of Gemma 4 with four billion active parameters. It responds faster than Gemma 4 31B with similar memory needs.
- Human rating:
- 1434 · rank 24 of 226 open
- In German:
- EuroEval German: rank 19 of 140
- Epoch index:
- 141.8 · rank 31 of 127 open
- Runs on
- about 20 GB of memory at 4 bits
- Context
- Context: 262,000 tokens · Inputs: text, image
All scores and links
LMArena by category
- Coding1446 · 36/221
- Maths1470 · 14/215
- Creative writing1401 · 25/224
- Following instructions1420 · 25/226
- Hard prompts1438 · 28/226
- Long prompts1427 · 28/205
- Multi-turn dialogue1441 · 23/224
- Building websites (WebDev)1360 · 39/58
- Understanding images (Vision)1260 · 9/49
6,094 votes · gemma-4-26b-a4b
Tests according to Epoch AI
- GPQA Diamond (expert questions)64.3 %
- AIME (maths)82.2 %
- Parameters
- 25.8bn, 4bn active
- Memory at 8 bits
- 27 GB
- Released
- April 2026
MiniMax · China
MiniMax-M3
Commercial use with conditionsRuns on: Mac Studio with 512 GB
Mixture-of-experts model from MiniMax for agents and coding with 512,000 tokens of context and image input. The licence requires a notice and a message to MiniMax for commercial use.
- Human rating:
- 1433 · rank 25 of 226 open
- In German:
- 1427 · rank 25 of 154 open
- Epoch index:
- 147.1 · rank 15 of 127 open
- Runs on
- about 306 GB of memory at 4 bits
- Context
- Context: 524,000 tokens · Inputs: text, image
All scores and links
LMArena by category
- Coding1471 · 14/221
- Maths1429 · 33/215
- Creative writing1404 · 18/224
- Following instructions1433 · 17/226
- Hard prompts1448 · 20/226
- Long prompts1446 · 16/205
- Multi-turn dialogue1443 · 21/224
- Building websites (WebDev)1482 · 18/58
- Understanding images (Vision)1253 · 12/49
63,973 votes · minimax-m3
Tests according to Epoch AI
- GPQA Diamond (expert questions)87.9 %
- AIME (maths)71.1 %
- Parameters
- 427bn, 23bn active
- Memory at 8 bits
- 454 GB
- Released
- June 2026
Licence: MiniMax Community Licence: for commercial use you must show “Built with MiniMax M3” and notify MiniMax once by email; above US$20m annual revenue written permission is required.
Mistral AI · France
Mistral Large 3
Free for commercial useRuns on: Server with several GPUs
Large model from Paris-based Mistral AI under the Apache 2.0 licence with image input. The largest open model from a European maker in this list.
- Human rating:
- 1428 · rank 28 of 226 open
- In German:
- 1436 · rank 23 of 154 open
- Runs on
- about 468 GB of memory at 4 bits
- Context
- Context: 262,000 tokens · Inputs: text, image
All scores and links
LMArena by category
- Coding1449 · 35/221
- Maths1415 · 49/215
- Creative writing1386 · 38/224
- Following instructions1403 · 38/226
- Hard prompts1430 · 35/226
- Long prompts1414 · 41/205
- Multi-turn dialogue1430 · 28/224
- Building websites (WebDev)1230 · 56/58
- Understanding images (Vision)1221 · 19/49
84,205 votes · mistral-large-3
- Parameters
- 675bn, 41bn active
- Memory at 8 bits
- 716 GB
- Released
- December 2025
Mistral AI · France
Mistral Medium 3.5
Commercial use with conditionsRuns on: Mac Studio with 128 GB
Mid-sized Mistral AI model with image input and 262,000 tokens of context; quantised, it fits a Mac Studio with 128 GB. The licence excludes large companies.
- Human rating:
- 1421 · rank 36 of 226 open
- In German:
- 1453 · rank 11 of 154 open
- Epoch index:
- 141.5 · rank 32 of 127 open
- Runs on
- about 86 GB of memory at 4 bits
- Context
- Context: 262,000 tokens · Inputs: text, image
All scores and links
LMArena by category
- Coding1461 · 24/221
- Maths1431 · 32/215
- Creative writing1374 · 45/224
- Following instructions1415 · 28/226
- Hard prompts1437 · 29/226
- Long prompts1416 · 38/205
- Multi-turn dialogue1424 · 33/224
- Building websites (WebDev)1264 · 53/58
- Understanding images (Vision)1223 · 18/49
11,778 votes · mistral-medium-3.5
- Parameters
- 127.7bn
- Memory at 8 bits
- 133 GB
- Released
- April 2026
Licence: Modified MIT licence: not permitted for companies with over US$20m monthly revenue; they need a commercial licence from Mistral AI.
Alibaba (Qwen) · China
Qwen3.5 122B-A10B
Free for commercial useRuns on: Mac Studio with 128 GB
Mid-sized Qwen model with image input and 262,000 tokens of context. With 4-bit quantisation it fits on a Mac Studio with 128 GB.
- Human rating:
- 1417 · rank 43 of 226 open
- In German:
- 1425 · rank 27 of 154 open
- Runs on
- about 88 GB of memory at 4 bits
- Context
- Context: 262,000 tokens · Inputs: text, image
All scores and links
LMArena by category
- Coding1436 · 43/221
- Maths1431 · 31/215
- Creative writing1368 · 47/224
- Following instructions1399 · 44/226
- Hard prompts1421 · 45/226
- Long prompts1410 · 45/205
- Multi-turn dialogue1417 · 37/224
- Building websites (WebDev)1361 · 38/58
- Understanding images (Vision)1245 · 15/49
29,732 votes · qwen3.5-122b-a10b
- Parameters
- 125.1bn, 10bn active
- Memory at 8 bits
- 130 GB
- Released
- February 2026
Thinking Machines · United States
Inkling Small
Free for commercial useRuns on: Mac Studio with 256 GB
Smaller version of Inkling with image input and 512,000 tokens of context. Quantised, it fits on a Mac Studio with 256 GB.
- Human rating:
- 1413 · rank 49 of 226 open
- In German:
- 1407 · rank 42 of 154 open
- Epoch index:
- 150.2 · rank 9 of 127 open
- Runs on
- about 188 GB of memory at 4 bits
- Context
- Context: 524,000 tokens · Inputs: text, image
All scores and links
LMArena by category
- Coding1452 · 32/221
- Maths1457 · 16/215
- Creative writing1331 · 69/224
- Following instructions1399 · 45/226
- Hard prompts1423 · 44/226
- Long prompts1401 · 50/205
- Multi-turn dialogue1414 · 40/224
- Building websites (WebDev)1410 · 30/58
- Understanding images (Vision)1235 · 17/49
32,304 votes · Inkling Small
Tests according to Epoch AI
- GPQA Diamond (expert questions)84.7 %
- AIME (maths)90 %
- SimpleQA (factual knowledge)19.1 %
- Parameters
- 266bn, 12bn active
- Memory at 8 bits
- 280 GB
- Released
- July 2026
OpenAI · United States
gpt-oss-120b
Free for commercial useRuns on: Mac Studio with 128 GB
OpenAI's open text model from 2025 with adjustable reasoning effort and tool calls. Fits on a Mac Studio with 128 GB or a single 80 GB GPU.
- Human rating:
- 1366 · rank 72 of 226 open
- In German:
- 1354 · rank 67 of 154 open
- Epoch index:
- 140 · rank 35 of 127 open
- Runs on
- about 72 GB of memory at 4 bits
- Context
- Context: 131,000 tokens · Inputs: text
All scores and links
LMArena by category
- Coding1379 · 75/221
- Maths1389 · 67/215
- Creative writing1275 · 96/224
- Following instructions1318 · 83/226
- Hard prompts1363 · 73/226
- Long prompts1319 · 89/205
- Multi-turn dialogue1340 · 82/224
30,206 votes · gpt-oss-120b
Tests according to Epoch AI
- GPQA Diamond (expert questions)67.7 %
- AIME (maths)88.9 %
- Parameters
- 116.8bn, 5.1bn active
- Memory at 8 bits
- 63 GB
- Released
- August 2025
NVIDIA · United States
Nemotron 3 Nano 30B-A3B
Commercial use with conditionsRuns on: Workstation with 64 GB
Small NVIDIA mixture-of-experts text model with three billion active parameters, built for fast agents and tool calls on your own hardware.
- Human rating:
- 1348 · rank 81 of 226 open
- In German:
- 1348 · rank 69 of 154 open
- Runs on
- about 28 GB of memory at 4 bits
- Context
- Context: 262,000 tokens · Inputs: text
All scores and links
LMArena by category
- Coding1378 · 77/221
- Maths1371 · 73/215
- Creative writing1269 · 100/224
- Following instructions1304 · 93/226
- Hard prompts1348 · 82/226
- Long prompts1300 · 99/205
- Multi-turn dialogue1319 · 93/224
15,770 votes · nvidia-nemotron-3-nano-30b-a3b-bf16
EuroEval German: rank 50 of 140
- Parameters
- 31.6bn, 3bn active
- Memory at 8 bits
- 34 GB
- Released
- December 2025
Licence: NVIDIA Open Model Licence: commercial use allowed with a licence notice; the licence ends if built-in safeguards are bypassed.
Mistral AI · France
Mistral Small 3.2
Free for commercial useRuns on: Computer with 32 GB
Compact Mistral AI model under the Apache 2.0 licence that runs on a computer with 32 GB. It ranks near the top of EuroEval's German comparison.
- Human rating:
- 1338 · rank 88 of 226 open
- In German:
- 1340 · rank 73 of 154 open
- Epoch index:
- 131.8 · rank 55 of 127 open
- Runs on
- about 16 GB of memory at 4 bits
- Context
- Context: 33,000 tokens · Inputs: text, image
All scores and links
LMArena by category
- Coding1362 · 87/221
- Maths1341 · 88/215
- Creative writing1305 · 84/224
- Following instructions1309 · 89/226
- Hard prompts1335 · 89/226
- Long prompts1326 · 83/205
- Multi-turn dialogue1343 · 78/224
- Understanding images (Vision)1142 · 26/49
17,462 votes · mistral-small-2506
Tests according to Epoch AI
- GPQA Diamond (expert questions)32.1 %
- AIME (maths)30.2 %
EuroEval German: rank 4 of 140
- Parameters
- 24bn
- Memory at 8 bits
- 25 GB
- Released
- June 2025
Allen Institute for AI · United States
Olmo 3.1 32B
Free for commercial useRuns on: Computer with 32 GB
Fully open model from the Allen Institute for AI: training data and training code are published alongside the weights. Useful when you need to trace what a model was trained on.
- Human rating:
- 1311 · rank 99 of 226 open
- In German:
- 1282 · rank 84 of 154 open
- Runs on
- about 22 GB of memory at 4 bits
All scores and links
LMArena by category
- Coding1347 · 93/221
- Maths1304 · 97/215
- Creative writing1265 · 103/224
- Following instructions1299 · 96/226
- Hard prompts1322 · 95/226
- Long prompts1312 · 94/205
- Multi-turn dialogue1310 · 97/224
11,817 votes · olmo-3.1-32b-instruct
EuroEval German: rank 65 of 140
- Parameters
- 32.2bn
- Memory at 8 bits
- 34 GB
OpenAI · United States
gpt-oss-20b
Free for commercial useRuns on: Computer with 32 GB
The small version of gpt-oss with adjustable reasoning effort and tool calls, for computers with 32 GB.
- Human rating:
- 1287 · rank 109 of 226 open
- In German:
- 1255 · rank 91 of 154 open
- Epoch index:
- 137.8 · rank 44 of 127 open
- Runs on
- about 13 GB of memory at 4 bits
- Context
- Context: 131,000 tokens · Inputs: text
All scores and links
LMArena by category
- Coding1307 · 105/221
- Maths1317 · 90/215
- Creative writing1201 · 133/224
- Following instructions1236 · 122/226
- Hard prompts1274 · 110/226
- Long prompts1250 · 121/205
- Multi-turn dialogue1268 · 115/224
10,419 votes · gpt-oss-20b
Tests according to Epoch AI
- GPQA Diamond (expert questions)47.7 %
- AIME (maths)65.2 %
- Parameters
- 20.9bn, 3.6bn active
- Memory at 8 bits
- 12 GB
- Released
- August 2025
Alibaba (Qwen) · China
Qwen 3.6 35B-A3B
Free for commercial useRuns on: Workstation with 64 GB
Qwen mixture-of-experts model with only three billion active parameters per request, so it responds quickly even on ordinary hardware; image input included.
- Epoch index:
- 144.1 · rank 26 of 127 open
- Runs on
- about 26 GB of memory at 4 bits
- Context
- Context: 262,000 tokens · Inputs: text, image
All scores and links
Tests according to Epoch AI
- GPQA Diamond (expert questions)79.8 %
- AIME (maths)86.7 %
- Parameters
- 36bn, 3bn active
- Memory at 8 bits
- 37 GB
- Released
- April 2026
Alibaba (Qwen) · China
Qwen3.5 9B
Free for commercial useRuns on: Laptop with 16 GB
Small Qwen model with image input and 262,000 tokens of context that runs on a laptop with 16 GB. Suited to summaries, drafts and simple assistant tasks.
- In German:
- EuroEval German: rank 53 of 140
- Epoch index:
- 139.4 · rank 38 of 127 open
- Runs on
- about 7 GB of memory at 4 bits
- Context
- Context: 262,000 tokens · Inputs: text, image
All scores and links
Tests according to Epoch AI
- GPQA Diamond (expert questions)72 %
- AIME (maths)61.6 %
- Parameters
- 9.7bn
- Memory at 8 bits
- 10 GB
- Released
- February 2026
Google · United States
Gemma 4 E2B
Free for commercial useRuns on: Laptop with 16 GB
The smallest Gemma 4 model, built for devices with little memory. Accepts text, images and audio.
- In German:
- EuroEval German: rank 82 of 140
- Runs on
- about 4 GB of memory at 4 bits
- Context
- Context: 33,000 tokens · Inputs: text, image, audio
No independent scores from LMArena or Epoch AI yet.
All scores and links
- Parameters
- 5.1bn
- Memory at 8 bits
- 5 GB
- Released
- April 2026
Google · United States
Gemma 4 E4B
Free for commercial useRuns on: Laptop with 16 GB
Small Gemma model for laptops and high-end phones that accepts text, images and audio.
- In German:
- EuroEval German: rank 57 of 140
- Runs on
- about 6 GB of memory at 4 bits
- Context
- Context: 33,000 tokens · Inputs: text, image, audio
No independent scores from LMArena or Epoch AI yet.
All scores and links
- Parameters
- 8bn
- Memory at 8 bits
- 8 GB
- Released
- April 2026
Mistral AI · France
Ministral 3 8B
Free for commercial useRuns on: Laptop with 16 GB
An even smaller version of Ministral 3 with image input that gets by with 16 GB of memory.
- In German:
- EuroEval German: rank 47 of 140
- Runs on
- about 6 GB of memory at 4 bits
- Context
- Context: 262,000 tokens · Inputs: text, image
No independent scores from LMArena or Epoch AI yet.
All scores and links
- Parameters
- 8.9bn
- Memory at 8 bits
- 9 GB
- Released
- December 2025
Swiss AI (EPFL, ETH Zürich, CSCS) · Switzerland
Apertus 8B
Free for commercial useRuns on: Laptop with 16 GB
Small version of Apertus for laptops, also with open training data.
- Runs on
- about 6 GB of memory at 4 bits (estimated)
No independent scores from LMArena or Epoch AI yet.
All scores and links
- Parameters
- 8.9bn
OpenGPT-X (Fraunhofer u. a.) · Germany
Teuken 7B
Free for commercial useRuns on: Laptop with 16 GB
Small model from the German OpenGPT-X project (Fraunhofer and others), trained with a large share of European languages. Listed is the commercially usable version 0.4.
- Runs on
- about 6 GB of memory at 4 bits
No independent scores from LMArena or Epoch AI yet.
All scores and links
- Parameters
- 7.5bn
- Memory at 8 bits
- 9 GB
Licence: Version 0.4 “commercial” is under Apache 2.0; the newer version 0.6 is licensed for non-commercial use only.
Mistral AI · France
Ministral 3 14B
Free for commercial useRuns on: Laptop with 16 GB
Small Mistral AI model with image input for laptops and local assistants.
- In German:
- EuroEval German: rank 37 of 140
- Runs on
- about 9 GB of memory at 4 bits
- Context
- Context: 100,000 tokens · Inputs: text, image
No independent scores from LMArena or Epoch AI yet.
All scores and links
- Parameters
- 13.9bn
- Memory at 8 bits
- 14 GB
- Released
- December 2025
UTTER-Konsortium (EU) · European Union
EuroLLM 22B
Free for commercial useRuns on: Computer with 32 GB
Model developed by a European research consortium that covers all 24 official EU languages.
- In German:
- EuroEval German: rank 42 of 140
- Runs on
- about 16 GB of memory at 4 bits
No independent scores from LMArena or Epoch AI yet.
All scores and links
- Parameters
- 22.6bn
- Memory at 8 bits
- 24 GB
Swiss AI (EPFL, ETH Zürich, CSCS) · Switzerland
Apertus 70B
Free for commercial useRuns on: Mac Studio with 128 GB
Open model from Switzerland (EPFL, ETH Zurich, CSCS) with published training data, trained on more than 1,000 languages. Proton's Lumo offers it as one of its models.
- Runs on
- about 50 GB of memory at 4 bits (estimated)
- Context
- Context: 100,000 tokens · Inputs: text, image, audio
No independent scores from LMArena or Epoch AI yet.
All scores and links
- Parameters
- 72bn
- Released
- July 2026
Meta · United States
Muse Spark Open
Commercial use with conditionsRuns on: Server with several GPUs
The first large open model of Meta's Muse series, released in October 2026, with image and video input. There are no independent measurements yet.
- Runs on
- about 581 GB of memory at 4 bits
- Context
- Context: 1 m tokens · Inputs: text, image, video
No independent scores from LMArena or Epoch AI yet.
All scores and links
- Parameters
- 822bn, 42bn active
- Released
- October 2026
Licence: Muse Model Licence: commercial use allowed, but every product must display “Built with Muse Spark Open”; offering it as a model service above US$20m revenue requires an agreement with Meta. Download only after accepting the terms on Hugging Face.
How to read the scores
What the numbers mean.
Human rating
On LMArena, people compare the answers of two anonymous models and pick the better one. This produces an Elo score, as in chess. A gap of 20 points is a small difference, 100 points a clear one. The rank counts open models only.
In German
The same rating, but only for prompts written in German. Because there are fewer votes, the uncertainty is higher. For smaller models we add the German rank from EuroEval, a European evaluation suite.
Epoch index
Epoch AI combines many tests (expert questions, maths, coding) into one index. It measures capabilities independently of human taste and so complements the human rating.
What is a benchmark? →Hardware and memory
We take the file size of the common 4-bit version plus 15 % for context and runtime, and assume that about three quarters of the memory is usable. Long documents need more memory.
What is quantization? →What to check
Before you run a model yourself.
Read the licence
MIT and Apache 2.0 allow commercial use without further obligations. Other licences require a notice in your product, a message to the maker, or exclude large companies. The assessment for each model is under “All scores and links”.
Origin and data protection
When a model runs on your own hardware, no data goes to the maker, even for models from China or the USA. It is different if you use the same model through the maker’s cloud.
To the AI glossary →Plan hardware realistically
Large models need hundreds of gigabytes of fast memory. For many office and admin tasks, a model that runs on a computer with 32 to 64 GB is enough. Test with your own documents before you buy hardware.
EU AI Act
If you run a model yourself, you generally become the deployer of an AI system under the EU AI Act. That includes AI literacy for staff and, depending on the use, transparency obligations.
More on the EU AI Act →Sources and method
Where the scores come from.
We take all scores unchanged from the sources; the selection of models, the licence assessment and the short texts are ours. Where a source has nothing on a model, the field stays empty.
- LMArena (Arena Leaderboard Dataset) ↗Human rating overall, in German, by category, WebDev and VisionCC BY 4.0As of 15 October 2026
- Epoch AI Benchmarking Hub ↗Epoch Capabilities Index, GPQA Diamond, AIME, SimpleQACC BY 4.0As of 19 October 2026
- EuroEval ↗German rank for smaller modelsMITAs of 17 April 2026
- Hugging Face ↗Parameters, licence, file sizes of the GGUF versionsModel page detailsAs of 19 October 2026
- models.dev ↗Context length, input types, release dateMITAs of 19 October 2026
FAQ
Frequently asked questions.
What is an open AI model?
Is a model from China GDPR-compliant if I run it locally?
What hardware do I need?
Why are some models listed without scores?
Can you help with choosing and running a model?
AI consulting
Your own language model for your company?
Together we find the model that fits your tasks, your hardware and your data protection requirements, test it with real documents and set up operation.
Looking for ready-made tools to run models, such as Ollama or LibreChat? To the AI tools directory →