tinyanalysis

Labs

Every lab tracked, grouped by where it is based and ordered by its best Small LM Index score: United States (14) · China (14) · Europe (12) · Rest of Asia (5) · Other (4).

United States

PrismML US

1 model · best index 74.6 (Ternary Bonsai 2 27B)

Ternary Bonsai 2 27B27.4B74.62026-09-17

NVIDIA US

8 models · best index 73.3 (Nemotron 3.5 Bolt 74B-A9B)

Nemotron 3.5 Bolt 74B-A9B74B / 9B73.32026-10-12
Nemotron 3.5 Lightning31.6B / 3.6B63.62026-08-11
Nemotron 3 Nano 30B-A3B31.6B / 3.5B62.62025-12-15
Nemotron Labs Dory 1.0 4B4B60.02026-10-16
Nemotron Nano 9B v28.9B52.52025-08-18
Nemotron 3.5 Bolt E4B8B—2026-10-05
Nemotron 3 Nano Omni 30B-A3B31.6B / 3.5B—2026-04-28
Nemotron 3 Nano 4B4B—2026

Meta US

5 models · best index 72.6 (Muse Glimmer 30B)

Muse Glimmer 30B29.6B72.62026-08-09
Llama 3.3 70B70.6B40.82024-12-06
Llama 3.1 8B8B21.32024-07-23
Llama 3.2 3B3.2B—2024-09-25
Llama 3.2 1B1.2B—2024-09-25

Google US

13 models · best index 70.1 (Gemma 4 31B)

Gemma 4 31B30.7B70.12026-04-02
Gemma 4 26B-A4B25.2B / 3.8B67.52026-04-02
Gemma 4 12B11.9B56.92026-06-03
DiffusionGemma 26B-A4B25.8B / 4B52.62026-06-09
Gemma 4 E4B8B38.22026-04-02
Gemma 3 27B27.4B32.72025-03-12
Gemma 4 E2B5.1B31.92026-04-02
Gemma 3 12B12.2B28.42025-03-12
Gemma 3n E4B7.8B18.22025-06-26
Gemma 3 4B4.3B17.52025-03-12
Gemma 3n E2B5.4B14.32025-06-26
Gemma 3 1B1B6.12025-03-12
Gemma 3 270M270M—2025-08-14

IBM (Granite) US

3 models · best index 67.0 (Granite 4.2 30B)

Granite 4.2 30B30B67.02026-08-07
Granite 4.2 8B8B63.52026-08-07
Granite 4.2 3B3B54.82026-08-07

OpenAI US

2 models · best index 64.0 (gpt-oss-20b)

gpt-oss-20b20.9B / 3.6B64.02025-08-05
gpt-oss-safeguard-20b20.9B / 3.6B—2025-10-29

Liquid AI US

10 models · best index 61.4 (LFM2.5 24B-A2B)

LFM2.5 24B-A2B24B / 2B61.42026-10-13
LFM3 1.7B1.7B60.02026-10-27
LFM2.5 2.6B2.7B43.22026-07-28
LFM2.5 1.2B Thinking1.2B30.12026-01-20
LFM2.5 1.2B Instruct1.2B20.62026-01-06
LFM2.5 350M350M12.62026-03-31
LFM3-Omni 2B2.2B—2026-10-21
LFM3-VL 2B1.9B—2026-10-19
LFM2.5-VL 3B3.1B—2026-08-11
LFM2.5 8B-A1B8.5B / 1B—2026-05-28

ServiceNow (Apriel) US

1 model · best index 60.9 (Apriel 1.6 15B Thinker)

Apriel 1.6 15B Thinker14.9B60.92025-12-04

Ai2 (OLMo) US

2 models · best index 57.0 (OLMo 3.1 32B Think)

OLMo 3.1 32B Think32.2B57.02025-12-10
OLMo 3 7B Think7.3B43.42025-11-20

Microsoft (Phi) US

5 models · best index 54.4 (Phi-4-reasoning-plus)

Phi-4-reasoning-plus14.7B54.42025-04-30
Phi-414.7B44.72024-12-12
FrogNano 4B4.7B—2026-09-17
Phi-4-reasoning-vision-15B15B—2026-03
Phi-4-mini3.8B—2025-02-26

Nous Research (Hermes) US

1 model · best index 51.4 (Hermes 4.3 36B)

Hermes 4.3 36B36.2B51.42025-11-17

poolside US

1 model

Laguna XS.233B / 3B—2026-04-28

Arcee AI (Trinity) US

1 model

Trinity Mini26.1B / 3B—2025-12-01

Essential AI (Rnj) US

1 model

Rnj-1.5 Instruct8.3B—2026-04-22

China

Nanbeige (Kanzhun) CN

2 models · best index 77.8 (Nanbeige4.3 33B)

Nanbeige4.3 33B33.1B77.82026-10-24
Nanbeige4.2 3B4.2B70.42026-07-21

Alibaba (Qwen) CN

16 models · best index 77.0 (Qwen3.8 27B)

Qwen3.8 27B27B77.02026-08-14
Qwen3.6 27B27B74.02026-04-22
Qwen3.6 35B-A3B35B / 3B72.22026-04-16
Qwen3.5 27B27B70.62026-02-24
Qwen3.5 35B-A3B35B / 3B69.92026-02-24
Qwen3.5 9B9B63.92026-03-02
Qwen3-Next 80B-A3B Thinking80B / 3B63.32025-09-11
Qwen3 30B-A3B Thinking 250730.5B / 3.3B59.32025-07-30
Qwen3.5 4B4B57.82026-03-02
Qwen3 4B Thinking 25074B55.72025-08-06
Qwen3 32B32.8B52.42025-04-29
Qwen3 8B8.2B48.52025-04-29
Qwen3.5 2B2.3B33.02026-03-02
Qwen3.5 0.8B900M10.72026-03-02
Qwen3-Coder-Next79.7B / 3B—2026-01-30
Qwen3-Coder 30B-A3B30.5B / 3.3B—2025-07-31

SparkLLM (XHToken) CN

3 models · best index 67.5 (Spark-X2.6 4B)

Spark-X2.6 4B4.1B67.52026-10-19
Spark-X2.5 4B4.1B65.52026-08-24
Spark-X2.6 1.7B1.7B50.12026-10-19

Z.ai (GLM) CN

2 models · best index 66.1 (GLM-4.7-Flash)

GLM-4.7-Flash31B / 3B66.12026-01-19
GLM-4.6V-Flash9B—2025-12-08

ByteDance Seed CN

1 model · best index 60.3 (Seed-OSS 36B Instruct)

Seed-OSS 36B Instruct36.2B60.32025-08-20

OpenBMB (MiniCPM) CN

9 models · best index 57.5 (MiniCPM5 2B)

MiniCPM5 2B2.5B57.52026-09-06
MiniCPM-V 4.7 35B-A3B35.2B / 3B54.52026-10-07
MiniCPM-SALA9.5B52.62026-02-11
MiniCPM-V 4.7 1B1.6B—2026-10-06
MiniCPM5 1B1.1B—2026-05-21
MiniCPM-V 4.6 Thinking1.3B—2026-05-08
MiniCPM-V 4.61.3B—2026-04-13
MiniCPM-o 4.59.4B—2026-02-03
AgentCPM-Explore 4B4B—2026-01-11

DeepSeek CN

4 models · best index 53.1 (DeepSeek-R1-0528-Qwen3-8B)

DeepSeek-R1-0528-Qwen3-8B8.2B53.12025-05-29
DeepSeek-R1-Distill-Qwen-32B32.8B50.92025-01-20
DeepSeek-R1-Distill-Llama-70B70.6B50.12025-01-20
DeepSeek-R1-Distill-Qwen-7B7.6B39.82025-01-20

Tencent (Hunyuan) CN

1 model · best index 52.5 (Hunyuan 7B Instruct)

Hunyuan 7B Instruct7.5B52.52025-08-04

InclusionAI (Ling) CN

4 models · best index 51.9 (Ling 3.0 Tiny)

Ling 3.0 Tiny7.9B / 1.4B51.92026-08-10
LLaDA2.2 mini16.3B / 1.4B29.72026-09-05
Ring-mini-2.016.3B / 1.4B—2025-09-19
Ling-mini-2.016.3B / 1.4B—2025-09-09

Xiaomi (MiMo) CN

3 models · best index 40.7 (MiMo-7B-RL)

MiMo-7B-RL7.8B40.72025-04-30
MiMo-V2.6-Distill-Qwen-9B9.4B—2026-09
MiMo-VL-7B-RL8.3B—2025-05-30

Moonshot AI (Kimi) CN

2 models

Kimi Linear 48B-A3B48B / 3B—2025-10-30
Kimi-VL-A3B-Thinking-250616.4B / 2.8B—2025-06-21

StepFun CN

1 model

Step3-VL-10B10B—2026-01

Baidu (ERNIE) CN

3 models

ERNIE 4.5 VL 28B-A3B Thinking28B / 3B—2025-11-07
ERNIE 4.5 21B-A3B Thinking21.8B / 3B—2025-09-08
ERNIE 4.5 0.3B360M—2025-06-30

Europe

Aleph Alpha DE

1 model · best index 68.3 (Kolibri-1)

Kolibri-178.1B / 3.5B68.32026-10-03

Mistral AI FR

7 models · best index 60.0 (Ministral 3 14B)

Ministral 3 14B14B60.02025-12-02
Magistral Small 1.224B55.92025-09-17
Ministral 3 8B8B55.22025-12-02
Ministral 3 3B3B49.42025-12-02
Mistral Small 3.224B38.62025-06-20
Devstral Small 224B—2025-12-09
Voxtral Small 24B24B—2025-07-15

Hugging Face (SmolLM) FR

1 model · best index 30.4 (SmolLM3 3B)

SmolLM3 3B3.1B30.42025-07-08

Swiss AI Initiative (Apertus) CH

2 models · best index 20.1 (Apertus 1.5 8B)

Apertus 1.5 8B8.9B20.12026-07-24
Apertus 1.5 70B72B—2026-07-24

EuroLLM (UTTER consortium) PT

2 models · best index 19.2 (EuroLLM 22B Instruct)

EuroLLM 22B Instruct22.6B19.22025-12-05
EuroLLM 9B Instruct (2512)9.2B—2026-01-26

Tilde LV

1 model

TildeOpen 30B30.7B—2026-02-04

SpeakLeash (Bielik) PL

1 model

Bielik 11B v3 Instruct11.2B—2025-11-07

OpenLLM-France (Luciole) FR

2 models

Luciole 23B Thinking23.2B—2026-09-09
Luciole 8B Instruct8.1B—2026-07-07

PleIAs FR

1 model

Baguettotron320M—2025-11-10

Rest of Asia

LG AI Research (EXAONE) KR

2 models · best index 62.0 (EXAONE 4.0 32B)

EXAONE 4.0 32B32B62.02025-07-15
EXAONE 4.0 1.2B1.3B36.12025-07-15

Sarvam AI IN

1 model · best index 61.6 (Sarvam 30B)

Sarvam 30B32.2B / 2.4B61.62026-03-03

Kakao (Kanana) KR

1 model · best index 50.7 (Kanana-2 30B-A3B Thinking)

Kanana-2 30B-A3B Thinking30.7B / 3B50.72026-01-14

Trillion Labs KR

1 model · best index 43.6 (Tri-21B Think)

Tri-21B Think20.7B43.62026-02-18

Other

Edge0 country not published

2 models · best index 62.8 (Edge0 35B-A3B (preview))

Edge0 35B-A3B (preview)34.7B / 3B62.82026-09-08
Edge0 8B-A1B (preview)7.9B / 1.4B47.22026-09-08

LLM360 / MBZUAI (K2) AE

1 model · best index 57.9 (K2-V2 Instruct)

K2-V2 Instruct70B57.92025-12-03

TII (Falcon) AE

1 model · best index 57.5 (Falcon H1R 7B)

Falcon H1R 7B7.6B57.52026-01-05

Cohere CA

1 model

North Mini Code30B / 3B—2026-06-17

Labs with no open models under 100B

Tracked so you know they were checked, not forgotten.

AnthropicNo open-weight models. Claude Haiku is the small tier, but its parameter count is undisclosed.
MiniMaxOpen-weight releases (M2, M2.5, ...) are ~200B+ total parameters (MoE, ~10B active).
Reflection AIFirst open-weight model Beam (Oct 2026) is 501B total / 23B active.
Thinking Machines LabInkling (Jul 2026) is 975B total / 41B active; a smaller variant is still well above 100B.
SpaceXAI (formerly xAI)Grok 4.x models are proprietary; earlier open checkpoints (Grok-1, Grok-2) are 270B+.
TypeSafe AIJev models are proprietary; parameter counts are unpublished.
Atomic ChatAn app for running other labs' open models locally (GGUF, MLX, ONNX). It publishes quantized copies of others' models, not models of its own.
OctenBuilds search infrastructure and embedding models (Octen-Embedding 0.6B-8B), not chat or reasoning language models.
Fastino LabsBuilds small task-specific models (GLiNER2.5 extraction, GLiDE decision model), not general-purpose language models.