cerebras/GLM-4.7-Flash-REAP-23B-A3B

Text Generation • 23B • Updated 5 days ago • 3.44k • 39

Note 30 on the Artificial Analysis Intelligence Index (Jan '26), beating GPT-OSS 20B, and only 3 points behind the larger 120B variant. More than HALF as intelligent as its big sibling, GLM 4.7 (Reasoning). Only 23B when pruned for "unused" experts. Uniquely good for its size, and MoE; only 3B params active per token. https://artificialanalysis.ai/models/glm-4-7-flash GGUF: unsloth/GLM-4.7-Flash-REAP-23B-A3B-GGUF https://huggingface.co/unsloth/GLM-4.7-Flash-REAP-23B-A3B-GGUF

janhq/Jan-v3-4B-base-instruct

Text Generation • 4B • Updated about 21 hours ago • 42 • 19

Note Beats Qwen3 4B Thinking... But it's not a thinking model. Just instruct! Same param count.

ServiceNow-AI/Apriel-1.6-15b-Thinker

Image-Text-to-Text • 15B • Updated Dec 22, 2025 • 2.89k • • 266

Note Doesn't usually overthink, massive improvement over the previous 1.5 model. Outstanding intelligence for a 15B model.

Alibaba-Apsara/DASD-4B-Thinking

Text Generation • 4B • Updated 13 days ago • 2.09k • 171

Note Born from a great paper. Visibly outperforms all models of similar size.

Nanbeige/Nanbeige4-3B-Thinking-2511

Text Generation • 4B • Updated Dec 17, 2025 • 2.82k • 176

Note Outperforms Qwen3 4B Thinking at a slightly smaller size.

ByteDance/Ouro-1.4B-Thinking

Text Generation • Updated 9 days ago • 1.82k • 28

Note On par with 3-4B models.

ByteDance/Ouro-2.6B-Thinking

Text Generation • Updated 9 days ago • 244 • 72

Note On par with 4-8B models.

tiiuae/Falcon-H1R-7B

Text Generation • 8B • Updated 7 days ago • 6.59k • 196

Note Overthinks, but good proof-of-concept. Similar in intelligence to Apriel 1.5 Thinker (a 15B model), but not as good at agentic tasks. A bit benchmaxxed, and not so great at general knowledge. Better with RAG.

AiAsistent/Gemma3-4B-Dark-Chain-of-Thought-CoT

Text Generation • 4B • Updated 24 days ago • 247 • 11

Note Experimental model.