Start with a known result
For your first local answer, use Qwen2 0.5B. For a specific model, choose its checkpoint below. Each profile connects the publisher card with the CKE artifact, launch path and measured scope.
Text and code
- Gemma3 270M IT
- Gemma4 E4B IT
- Qwen2 0.5B Instruct
- Qwen3 0.6B
- Qwen3.5 0.8B
- Qwen3.5 35B-A3B
- Qwen3.6 27B
- Qwen3.8 27B dense
- Nemotron Nano 9B v2
- GLM4 9B
- Nanbeige 4.1 3B / Llama-family
- Cohere2 Command R
- Cohere North Mini Code
- Cohere Command A+ 218B/25B-active
- Laguna-XS 2.1
- Kimi-VL A3B text decoder
- Muse-Glimmer 30B text decoder
- Instella-MoE 16B-A3B
- GPT-2
Images and recordings
- Qwen3-VL 8B
- Qwen3.6-VL
- Whisper Tiny, Base, or Small
- Parakeet TDT 0.6B v3
- Cohere Transcribe 2B
- North Micro Vision Instruct
Special-purpose models and previews
- Inkling 2 Mini — experimental Q8_0
- Qwen3.8 Flash Next
- Kokoro 82M
- Ming3 TTS 0.5B
- Qwen3 TTS CustomVoice
- Clef Flash decision head
- RWKV-7
- Gemma4 speculative assistant
Compare machines
Check your hardware before downloading larger weights. Compare long-context evidence if you plan to use long documents or code.
Compare compute, model formats, servers and agent harnesses. Each model card gives its tested boundary and the next check.