Nanbeige4.3-33B vs Qwen3.8-27B

Official Nanbeige4.3-33B benchmark figure from the Nanbeige Hugging Face card — Nanbeige/Nanbeige4.3-33B card figures/nanbeige43-benchmarks.png. Source · License: Apache-2.0.

Verdict. Pick Nanbeige4.3-33B if you want Nanbeige / BOSS Zhipin’s ungated 33B text teacher with a documented 512K context and published vLLM/SGLang qwen3-parser flags. Pick Qwen3.8-27B if you want Qwen’s Apache-2.0 27B native VLM (image and video) with a LICENSE file on disk, 262K native context, and the larger Hub snapshot. Nanbeige’s official table names Qwen3.8-27B as the lead ~30B opponent; those rows stay vendor. We have no sourced bake-off.

Compare Nanbeige4.3-33B and Qwen3.8-27B.

Dimension	Nanbeige4.3-33B	Qwen3.8-27B	Source
License	Apache-2.0 (card tag; LICENSE blob 404)	Apache-2.0 (LICENSE file + card tag)	Hugging Face cards (2026-10-26)
Gating	No	No	Hugging Face API
Developer	Nanbeige (BOSS Zhipin)	Qwen	HF org page / official card
Parameters	33B on the card; HF safetensors 33,118,977,024	27B on the card; HF safetensors 27,781,427,952	Official cards / HF API
Architecture	Dense llama / LlamaForCausalLM (52 layers)	qwen3_5 VLM / Qwen3_5ForConditionalGeneration (64 layers)	Official configs
Modalities	text-generation	image-text-to-text (image + video + text)	Hugging Face API
Context length	524,288 tokens (card + config)	262,144 native; extensible to 1,000,000 (card)	Official cards / configs
Serving named	Transformers, SGLang, vLLM with qwen3 / qwen3_coder parsers; TP=4	SGLang, vLLM, TokenSpeed (cookbook links; no one-liner copied here)	Official cards
Hugging Face likes	365 (2026-10-26)	18,393 (2026-10-26)	Hugging Face API
Quality vs ChatGPT / Cursor	Not independently tested	Not independently tested	n/a

When to choose which

Text-only 33B agent teacher with 512K context: Nanbeige4.3-33B — Card job is teacher for later 3B distillation. Not a quality rank.
Native image and video understanding: Qwen3.8-27B — pipeline_tag image-text-to-text. Nanbeige is text-only.
LICENSE file on the weight repo: Qwen3.8-27B — Both cards tag apache-2.0. Only Qwen’s LICENSE blob resolved on 2026-10-26.
Published vLLM/SGLang flag set on the card: Nanbeige4.3-33B — Nanbeige pastes TP=4 recipes with qwen3 parsers. Qwen points at cookbooks.

Read the entity pages: Nanbeige4.3-33B and Qwen3.8-27B. Both hang off ChatGPT alternatives and closed LLM APIs. Nanbeige also hangs off Claude Code alternatives and Cursor alternatives.

Why this comparison exists

Searchers who want a ~30B open agent/coding checkpoint now have Nanbeige’s 33B teacher (created 2026-10-24) next to Qwen’s 27B VLM (created 2026-08-05), which is the first opponent column on the Nanbeige card. This is identification, not a bake-off. Muse-Glimmer-30B and Gemma4-31B remain off-site.

Sourced snapshot (2026-10-26)

Field	Nanbeige4.3-33B	Qwen3.8-27B
Host	huggingface.co/Nanbeige/Nanbeige4.3-33B (not gated)	huggingface.co/Qwen/Qwen3.8-27B (not gated)
HF likes / downloads	365 / 0	18,393 / 7,762,014
License	Apache-2.0 (card tag; LICENSE 404)	Apache-2.0 (LICENSE file)
Size	33B (card); ~33.1B safetensors	27B (card); ~27.8B safetensors
Job	text-generation; agent teacher	image-text-to-text; coding / agents / VLM
Context	524,288	262,144 native (card also names 1M)
Official serve path	vLLM / SGLang one-liners, TP=4	SGLang / vLLM / TokenSpeed cookbooks
Quality vs ChatGPT	Not independently tested	Not independently tested

License

Both cards tag apache-2.0. Qwen’s LICENSE file is Apache License 2.0. Nanbeige’s LICENSE path 404’d on 2026-10-26. Not legal advice.

How to choose

Need a 33B text teacher with 512K and a pasted vLLM/SGLang recipe: Nanbeige4.3-33B.
Need a 27B native VLM with a LICENSE blob and image/video input: Qwen3.8-27B. Engine page: vLLM.
Need a 27B decision post-train of Qwen3.8-27B: Cloudflare Clef.
Need Apache-2.0 39.8B/2.5B local coding MoE with GGUF/Ollama: Flint — Flint vs Ornith-1.5-122B.
Need a 28.8B/4.3B compact preview MoE: Raon-1-mini.

See also AI Chat & Agents and Dev & Coding.

FAQ

Which is better at coding agents?

Nanbeige’s card table is self-reported and uses one harness per row. On that table Nanbeige leads Terminal-Bench 2.1 (80.9 vs 73.0) and DeepSWE (65.5 vs 42.5); Qwen3.8-27B leads SWE-Bench Pro (61.7 vs 60.3). We do not merge those into a ranking.

Which is easier to host?

BF16 usedStorage is ~61.7 GiB (Nanbeige) vs ~51.8 GiB (Qwen). Nanbeige’s official recipes use tensor parallel 4. Neither card gave us a measured single-GPU VRAM number. Untested here.

Why this pairing?

The Nanbeige README compares against Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B. Only Qwen3.8-27B is a page on this site, so it is the same-site opponent. Muse-Glimmer-30B and Gemma4-31B are not pages here.

Is Cloudflare Clef the same as Qwen3.8-27B?

No. Clef is a 27B decision post-train of Qwen3.8-27B. It scores schema options; it does not generate free text. See Cloudflare Clef.

Sources & verification

Last verified 2026-10-26 · Last updated 2026-10-26· Edited by forkalt editors · Editorial policy

License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.