EverygenAIEverygenAI
返回博客

Nanbeige4.3 Comparison: A 33B Open Agent Model Worth Trying?

Nanbeige 4.3 comparison: Nanbeige4.3-33B vs Qwen and peers—HF agent scores, who should try it, and ecommerce AI video ads.

2026年10月30日Keke

New open models drop every week. Charts flood X. Someone says “SOTA.” Someone else says “wait for independent evals.” Meanwhile you still need a model that can run tools, keep a long thread, and help you ship work—not just win a screenshot war.

That is who this Nanbeige 4.3 comparison is for.

Not scoreboard collectors. Builders, agent tinkerers, and ecommerce teams who want a clear read on Nanbeige4.3—especially Nanbeige4.3-33B—as an open source agent model. We will stick to what the authors publish on Hugging Face. Label their benches as their harness. Then talk about when to try it vs Qwen-class peers, and how to turn strong LLM copy into product video ads without pretending a chat model is a camera.

What Nanbeige4.3-33B is

Nanbeige4.3-33B comes from Nanbeige LLM Lab—the LLM research team at BOSS Zhipin (Kanzhun). On Hugging Face it ships under Apache-2.0, so commercial use and modification are on the table if you follow the license.

Per the Nanbeige4.3-33B model card, it is a roughly 33B dense flagship. Not a tiny MoE slice. Full dense capacity aimed at agent work. The lab frames it as the teacher for the Nanbeige4.3 series. Big teacher first. Distill later into smaller models (they call out a planned Nanbeige4.3-3B path). So if you are comparing “what can I run today” vs “what will the series become,” this 33B is the heavy end.

What they say it is good at:

  • Agentic coding — terminal and software-engineering style agent tasks

  • Professional cowork — office / automation / job-style agent benches in their table

  • 512K context — up to 524,288 tokens

  • Thinking + tools — chat-template knobs: enable_thinking, preserve_thinking, and tools (they recommend tool_call_format="xml" for best tool calling; json works for compatibility)

  • Serving — Transformers, SGLang, and vLLM quickstarts on the card

Inference tips from the README: for agentic / tool-use jobs they suggest temperature 1.0, top_p 0.95, and a high max-new-tokens cap; for reasoning / chat they suggest temperature 0.6. They evaluate Nanbeige4.3-33B with preserve_thinking=true in the chat template. They recommend preserve_thinking=False for general chat, and True for multi-turn tool use, office tasks, and code-agent workflows.

In plain English: this is an open source agent model you can self-host, wire to tools, and keep a long working memory—if you have the GPUs.

Nanbeige4.3 comparison table (authors’ harness)

Here is the Nanbeige 4.3 comparison people actually search for: Nanbeige4.3 vs Qwen and other ~30B peers.

Important honesty first. The numbers below are from Nanbeige’s own README. They compare Nanbeige4.3-33B, Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B. All at highest thinking effort. Same inference hyperparameters and harness across models—their setup, not an outside lab. Treat them as author-reported signals. Not gospel.

Selected agentic coding and cowork rows from the HF table:

Benchmark (category)Nanbeige4.3-33BQwen3.8-27BMuse-Glimmer-30BGemma4-31BTerminal-Bench 2.1 (agentic coding)80.973.061.847.2Terminal-Bench 4.023.86.31.60.0DeepSWE65.542.54.42.7SWE-Bench Pro60.361.751.236.9SWE-Atlas-QnA50.841.917.73.3NL2Repo53.048.327.433.8SkillsBench61.156.439.728.9Automation Bench (cowork)54.238.315.29.7JobBench44.834.516.69.0GDPval84.886.475.871.4Claw-Eval pass³78.375.257.344.2Toolathlon-Verified70.468.547.227.8DeepPlanning63.349.221.419.4YC-bench**$1,025,437**$641,290$513,978$265,152GPQA Diamond (reasoning)89.589.283.591.3AIME 202695.294.594.789.2

Source: Nanbeige/Nanbeige4.3-33B README. Footnotes on that page matter—different benches use Terminus 2, mini-swe-agent, Claude Code harnesses, and so on. Same harness within each row, still author-run.

What the table actually says

On many agentic coding and cowork rows, Nanbeige4.3-33B leads in their report. Terminal-Bench 2.1 at 80.9 vs Qwen’s 73.0 is the headline people will quote. DeepSWE and Automation Bench show wider gaps in the same table.

It is not a clean sweep. On SWE-Bench Pro, Qwen3.8-27B edges ahead (61.7 vs 60.3). On GDPval, Qwen is slightly higher (86.4 vs 84.8). On GPQA Diamond, Gemma4-31B leads their table (91.3). So a fair Nanbeige4.3 vs Qwen takeaway is: strong author-reported agent lean for Nanbeige on several coding/cowork benches; still check the rows that matter to your job.

Also remember: peer names and sizes are not identical (27B / 30B / 31B / 33B). Same “~30B class” story. Not a same-parameter clone fight.

Who should use which

Try Nanbeige4.3-33B if you want an open agent teacher you can deploy

Reach for Nanbeige4.3-33B when you care about:

  • Apache-2.0 weights you can host yourself

  • Long 512K threads for repos, SOPs, or multi-tool chats

  • Explicit thinking / preserve-thinking / tool templates

  • Serving stacks you already know: Transformers, SGLang, vLLM

  • Being early on a teacher model that may seed smaller distilled agents later

Good fit for agentic coding experiments, internal cowork agents, and teams that already operate open weights.

Lean Qwen-class (e.g. Qwen3.8-27B in their table) if your evals or stack already favor it

If your harness, plugins, or team muscle memory are Qwen-shaped, do not rip it out for one launch chart. In Nanbeige’s own table, Qwen wins a couple of important rows. Run both on your prompts. Keep the one you edit less.

Keep Muse-Glimmer / Gemma4 as peers, not afterthoughts

Their README includes them as same-class peers. On several agent rows they trail in the author table; on GPQA Diamond, Gemma leads. If your workload is more pure reasoning than terminal agents, re-check those columns before you standardize.

Honest gaps

Author benches are not independent third-party leaderboards. A 33B dense model needs real hardware. Long context is not free in VRAM or latency. And no text agent—open or closed—reliably ships brand-safe product video by itself. Logos warp. Labels invent themselves. Claims drift. That is a different job.

From open agent briefs to ecommerce ad creative AI

Here is the bridge most model roundups skip.

An open source agent model like Nanbeige4.3-33B is excellent at the thinking layer of ads: hooks, scripts, benefit angles, audience framing, objection handling, A/B copy packs. Long context helps you dump a PDP, brand voice notes, and past winners into one thread. Tool calling helps structure the brief.

It is not a dedicated video engine.

Finished ecommerce ad creative AI still needs:

  1. Product inputs (photo or public product URL)

  2. Scene / hook drafts you can review

  3. Generation settings (model, duration, resolution, credit cost)

  4. A human check for product fidelity and claims

That is the loop on EveryGenAI: add product → review creative drafts → generate clips only after you approve. Built for Shopify and ecommerce sellers who need store and social creative. Standalone workflow. No Shopify install required to try it. No auto-publish to your store or ads account—you download what you select.

Soft truth: use Nanbeige (or Qwen) for the brain. Use a product-first generator for the pixels. Keep a human for the brand and legal check.

Suggested workflow: Nanbeige → EveryGenAI → human check

  1. Brief in Nanbeige4.3-33B. Paste product copy, audience, placement, and 2–3 past winners. Ask for five hooks, three benefit angles, and a 15-second UGC-style script. Keep enable_thinking on if you want reasoning; use preserve_thinking=true if you are iterating with tools across turns.

  2. Tighten claims. Force the model to stick to verified features on the PDP. Delete anything you cannot prove.

  3. Move into EveryGenAI. Upload the product photo or share a public product URL. Map the best hook and scene into the creative review step. Approve drafts before you spend credits.

  4. Generate a small batch. One or two clips first. Inspect product shape, logo, label text, and motion.

  5. Human check before spend. Verify fidelity and claims. Add your captions, audio, or end card. Then test in ads or organic.

This keeps text cheap relative to video credits. It also keeps ecommerce ad creative AI output closer to something you can actually publish.

FAQ

Is Nanbeige4.3-33B really open source?

The Hugging Face card lists Apache-2.0 for Nanbeige/Nanbeige4.3-33B. That is a permissive open license for the weights as published—read the license text for your use case (including commercial). “Open” still means you bring GPUs, ops, and evals.

How does Nanbeige4.3 vs Qwen look on coding agents?

In Nanbeige’s own harness table, Nanbeige4.3-33B scores 80.9 on Terminal-Bench 2.1 vs 73.0 for Qwen3.8-27B, and leads several other agentic coding / cowork rows. Qwen edges ahead on SWE-Bench Pro and GDPval in that same table. Cite the HF README; do not treat one launch chart as forever.

What is the context length?

Up to 524,288 tokens (512K) per the model card. Useful for long repos and multi-turn agent threads. Also heavier to serve—budget memory and latency.

Can Nanbeige4.3-33B output finished video ads alone?

No. It can draft hooks, scripts, shot lists, and creative rationales. Finished ecommerce clips still need a product-to-clip workflow with draft review and fidelity checks. Pair the model with a tool like EveryGenAI, then inspect every frame.

Who built Nanbeige?

Nanbeige LLM Lab is the LLM research team at BOSS Zhipin (Kanzhun), per the Nanbeige org page. Contact listed on the model card: nanbeige@kanzhun.com.

Closing: try the agent, then ship the ad

So—is Nanbeige4.3 worth trying?

If you want a 33B open agent model with Apache-2.0 weights, 512K context, thinking/tool templates, and author-reported strength on several coding and cowork benches vs Qwen-class peers, yes—worth a serious try on your own harness. If you need one independent leaderboard to crown a king, wait for outside evals and keep Qwen (or another peer) in the A/B.

Either way: draft the brief. Review the creative. Generate the clip. Check the product.

When you are ready to turn a product link or photo into reviewable ad concepts and generated clips, try EveryGenAI.

Suggested URL slug: /blog/nanbeige4-3-comparison

Primary keywords: Nanbeige4.3, Nanbeige4.3-33B, Nanbeige 4.3 comparison
Secondary keywords: open source agent model, vs Qwen, ecommerce ad creative AI

Link suggestions:

  1. External (canonical): Nanbeige4.3-33B on Hugging Face

  2. External: Nanbeige org (BOSS Zhipin Nanbeige LLM Lab)

  3. External: Nanbeige4.3-33B-Base (if readers want the base for post-training)

  4. Internal: EveryGenAI homepage / product → drafts → clips workflow

  5. Internal (if you add related posts): link to other model comparison posts or a “how to write product video ad scripts with AI” guide on everygenai.ai