Nanbeige4.3 33B vs Ornith 1.5 35B-A3B (MoE)

Analysis

Nanbeige4.3 33B and Ornith 1.5 35B-A3B take opposite routes to a local coding agent. Nanbeige is dense: all 33.1B parameters work on every token (24 October 2026, Apache 2.0, text only). Ornith 1.5 35B-A3B is a mixture of experts that computes with about 3B of its 35B (19 August 2026, MIT, with an image and video tower in its weights). Their weight files are almost the same size, so the memory bill at a short context is nearly identical. The differences that matter are speed, long context and what each vendor claims.

Weights first. Q4_K_M is 20.2 GB for Nanbeige and 21.3 GB for Ornith, and with an 8k context the totals are 24.5 GB and 24.1 GB, so both just miss a 24 GB card at Q4_K_M and both fit it at Q3_K_M. Past that they diverge. Ornith is a Qwen3.5-style hybrid: only 10 of its 40 layers keep a cache that grows (2 KV heads of width 256), about 20 KB per token, against Nanbeige's 213 KB with all 52 layers caching. At the 262,144 tokens both can reach, the cache is 55.8 GB for Nanbeige against 5.4 GB for Ornith, and the full-window totals are 85.1 GB against 29.9 GB. On an RTX 5090 at Q4_K_M, Ornith holds the whole 262,144-token window and Nanbeige holds 32,768. Speed splits the same way. A mixture of experts reads only its active experts for each token, and this site estimates Ornith at 191 tok/s on an RTX 3090 at Q3_K_M against 40.7 for Nanbeige (4.7 times), and 220 against 58.8 on an RTX 5090 at Q4_K_M (3.7 times). Quality is closer than the speed gap suggests, and it favors the dense model. On the three benchmark fields both report, Nanbeige scores GPQA Diamond 89.5 against 89.2, SWE-Bench Pro 60.3 against 59.6 and Terminal-Bench 2.1 80.9 against 67.8, and both Terminal-Bench figures use the Terminus 2 harness. That 13.1-point lead is the one real separation; the other two gaps are under a point, too small to separate with two vendors' separate runs. Ornith also reports SWE-bench Verified 79.0, which Nanbeige's card has no figure for. Neither vendor's comparison includes the other model: Nanbeige's table names Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, while Ornith AI's published comparison, as recorded on this site, names Gemma 4 31B, Muse Glimmer 30B and Qwen 3.6 35B-A3B. The figures here come from two separate sets of vendor runs, and neither had been independently reproduced when this page was written. On tooling, Ornith has an Ollama tag (ornith-1.5:35b) where Nanbeige runs from community GGUFs.

Bottom line: Pick Ornith 1.5 35B-A3B if you want speed or long context: it decodes 3.7 to 4.7 times faster on this site's estimates, caches about a tenth as much at 262,144 tokens, matches Nanbeige on GPQA Diamond and SWE-Bench Pro, and carries an image and video tower. Pick Nanbeige4.3 33B if command-line agent work is the job and you can spend the speed: it leads Terminal-Bench 2.1 by 13.1 points on the same harness, on a 32 GB card at a short context or a 64 GB Mac. Both sets of scores are vendor-reported, so run both on your own repository.