SYLIN

Nanbeige

Model · Language & code models

33B dense open model, 66.3 GB; claimed to beat Qwen 3.8 and Gemma 4 on agentic and knowledge-work benchmarks

AccessOpen weightsNeeds~8 GBDownload8 GBLicenseApache-2.0

Versions

VersionReleasedRank
Nanbeige 4.3LATESTNov 1
Nanbeige 4.2Jul 23

Evidence

  1. Nov 1
    Claimed higher scores than Qwen 3.8 on Terminal-Bench and Deep suite
  2. Jul 26
    Nanbeige 4.2 3B looped-transformer agentic model, 8 GB, claimed to beat Gemma 4 and Qwen 3.5 4B/9B on agentic and coding benchmarks
  3. Jul 26
    Vendor charts show it topping agent and reasoning benchmarks over Qwen3.5-9B and Gemma 4 12B; Gemma slightly ahead on SciCode
  4. Jul 23
    Looped transformer reuses one layer's weights over several passes, aiming for larger-model reasoning depth without extra parameters
  5. Jul 23
    Via Hermes Agent it found and fixed a planted Redis pub/sub channel bug in a FastAPI app, verified end-to-end, but took about an hour

Cost to run

Alternatives