Nanbeige4.3 33B vs Qwen 3.8 27B

Analysis

Nanbeige4.3 33B and Qwen 3.8 27B are both dense, Apache 2.0, reasoning-first agent models in the 30B class, released ten weeks apart (Qwen on 14 August 2026, Nanbeige on 24 October 2026). Nanbeige's own model card treats Qwen3.8-27B as the rival to beat, the strongest of the three models in its comparison table. A card can only say who scores higher on its rows. This page adds what it leaves out: what each model costs in memory, which cards run it at which context, and which differences a benchmark table will not show.

The architectures could hardly differ more for two dense models of this size. Nanbeige is a plain 52-layer Llama-style stack where every layer keeps a full KV cache (8 KV heads of width 128), about 213 KB for each token of context. Qwen 3.8 27B is a hybrid: only 16 of its 64 layers keep a cache that grows (4 KV heads of width 256) and the other 48 carry a fixed-size state, which comes to about 66 KB per token, a third as much. Weights are closer. Q4_K_M is 20.2 GB for Nanbeige against 16.4 GB for Qwen, so with an 8k context the totals are 24.5 GB and 19.0 GB, and that gap decides the 24 GB tier: Qwen fits an RTX 3090 or RTX 4090 at Q4_K_M, while Nanbeige just misses and has to drop to Q3_K_M. Context widens it. At the 262,144 tokens both can reach, the KV cache is 55.8 GB for Nanbeige against 17.2 GB for Qwen, and the full-window totals are 85.1 GB against 37.7 GB, the difference between needing a 96 GB RTX Pro 6000 and getting by with a 48 GB RTX 6000 Ada. On an RTX 5090 at Q4_K_M, Qwen reaches 131,072 tokens and Nanbeige 32,768. Speed follows the weights, since both read every parameter for each token: this site estimates 58.8 tok/s for Nanbeige against 68.5 for Qwen on an RTX 5090 at Q4_K_M, roughly a sixth faster for Qwen. On quality the picture reverses. On the three benchmark fields this site tracks, Nanbeige scores GPQA Diamond 89.5 against 89.2, Terminal-Bench 2.1 80.9 against 73.0 and SWE-Bench Pro 60.3 against 61.7, the Terminal-Bench figures coming from Terminus-family harnesses (Terminus 2 for Nanbeige, Terminus for Qwen). Across the 20 rows of Nanbeige's card it is higher on 18, and lower on SWE-Bench Pro and on GDPval (84.8 against 86.4). The widest gaps are on long agent tasks: DeepSWE 65.5 against 42.5, Terminal-Bench 4.0 23.8 against 6.3, Automation Bench 54.2 against 38.3 and DeepPlanning 63.3 against 49.2, while four of the 18 wins are by less than a point. These are Nanbeige's own runs and no independent reproduction had appeared when this page was written, but the card's Qwen figures for GPQA Diamond, SWE-Bench Pro and Terminal-Bench 2.1 match the ones Qwen publishes, unlike its figures for some other rivals. Qwen also reports LiveCodeBench v6 90.3, which Nanbeige's card has no counterpart for. Qwen reads images and video where Nanbeige is text only, and Qwen has an official Ollama tag (qwen3.8:27b) where Nanbeige runs from community GGUFs and Ollama's hf.co syntax.

Bottom line: Pick Nanbeige4.3 33B if agent quality is the priority and you have a 32 GB card, a 48 GB workstation card or a 64 GB Mac, with sessions short enough for the memory left over (about 32,000 tokens on a 32 GB card). Pick Qwen 3.8 27B if you have a 24 GB card, need long context on consumer hardware, want image or video input, or want a one-line Ollama install. Nanbeige's lead is widest on long, multi-step tasks and absent on SWE-Bench Pro, and every score here is a vendor's own, so run both on your own repository before committing.