refactr

Supported models

Newest first. Scores come from each model's own card and were checked against the cards on 10 November 2026.

ModelProviderLicenceSizeMemoryNodes a copyScores
Nemotron 3.5 Super VLNVIDIAOpenMDW-1.1120B12B active82 GB1GPQA Diamond 89.5SWE-bench Verified 71.1Terminal-Bench 2.1 60.7
Kolibri-1Aleph AlphaApache 2.078B3.5B active88 GB1GPQA Diamond 84.3LiveCodeBench v6 85.9IFBench 78.1
DeepSeek V4.1 Flash 1017DeepSeekMIT552B backbone16B active216 GB2Terminal-Bench 4.0 42.3DeepSWE v1.1 77.0HLE 42.6
Kimi K3.1Moonshot AIModified MIT2.8T104B activeOn a callOn a callGPQA Diamond 93.5Terminal-Bench 2.1 88.3 (Kimi K3 card)
Qwen3.8 27BQwenApache 2.027B33 GB1LiveCodeBench v6 90.3Terminal-Bench 2.1 73.0GPQA Diamond 89.2IFBench 79.5
GLM-5.3 FlashZ.aiMIT320B18B active200 GB2Terminal-Bench 2.1 84.3SWE-bench Verified 92.0 (Vals AI)
Muse Glimmer 30BMetaApache 2.030B31 GB1SWE-bench Verified 76.0MCP-Atlas 75.5CharXiv-R 78.8
Qwen3.6 35BQwenApache 2.035B3B active45 GB1SWE-bench Verified 73.4GPQA Diamond 86.0MMMU 81.7
Gemma 4 26BGoogleApache 2.025B4B active37 GB1MMMU-Pro 73.8GPQA Diamond 82.3
Mistral Small 4Mistral AIApache 2.0119B6.5B active72 GB1GPQA Diamond 71.2
Qwen3 Coder NextQwenApache 2.080B3B active55 GB1SWE-bench Verified 70.6Terminal-Bench 2.0 36.2
gpt-oss 120BOpenAIApache 2.0117B5B active70 GB1SWE-bench Verified 62.4GPQA Diamond 80.1
Qwen3 Embedding 8BQwenApache 2.08B10 GB1MTEB multilingual retrieval 70.88

Reading the table

  • Size. Each model is listed at the setup that keeps its quality within about half a point to two points of the original.
  • Memory. The weights plus room for 8 conversations of 32K tokens.
  • Nodes a copy. A model too big for one node spreads each copy across several linked nodes.
NextHow suggestions are made