Tested models

← README

Models Reika has been run against, in daily use and in evaluations. Local models are what Reika was designed around, and where the behavior work is tested. Day-to-day use is mostly on API models, which also serve as the comparison for what a local quant can and cannot hold.

Quantization is the one in parentheses. Where a note names modes, that is where the model was used rather than the only thing it can do.

Local

• Ling 3.0 Tiny (Q4_K_M)
• Qwen3.8 Flash Next (UD-IQ3_XXS; chat mode; minimal mode; via SSD streaming)
• DeepSeek V4 Flash 0731 (UD-IQ3_XXS; chat mode; minimal mode; via SSD streaming)
• Muse Glimmer 30B (IQ3_XXS)
• Qwen3.8 27B (UD-IQ3_XXS)
• Qwen3.6 35B A3B (UD-IQ2_M)
• Qwen3.6 35B A3B (UD-IQ2_XXS)
• Ornith 1.0 35B (UD-IQ2_M)
• Laguna XS 2.1 (IQ2_M)
• KAT Coder V2.5 Dev (IQ2_M)
• North Mini Code 1.0 (UD-IQ3_XXS)
• Gemma 4 26B A4B IT (UD-IQ3_XXS)
• GPT-OSS 20B (MXFP4 Q4_K_M)
• Ornith 1.5 9B (Q5_K_M)
• Qwen3.5 9B (MLX Q4)
• LFM2.5 8B A1B (MLX Q4)
• Granite 4.1 8B (MLX Q4)

API

• MiMo V2.6 Pro (OpenCode Go)
• MiMo V2.6 Flash (OpenCode Go)
• GLM 5.3 Flash (OpenCode Go)
• DeepSeek V4.1 Flash (OpenCode Go)
• GLM 5.2 (OpenRouter, OpenCode Go)
• Kimi K3 (MoonshotAI)
• Kimi K2.6 (MoonshotAI)
• Kimi K2.5 (MoonshotAI, OpenRouter)
• DeepSeek V4 Flash (OpenRouter, OpenCode Go)
• Laguna XS 2.1 (OpenRouter)
• Laguna M.1 (OpenRouter)
• StepFun 3.5 Flash (OpenRouter, NVIDIA NIM)
• StepFun 3.7 Flash (NVIDIA NIM)