Two efficient open-weight workhorses promise near-flagship skill without flagship cost. This Nanbeige vs. Qwen test pits a new 33B agent specialist against a proven 27B local all-rounder. Nanbeige4.3-33B is a 33-billion-parameter dense model with full vendor benchmark tables, which means every parameter works on every token, released October 24, 2026, under Apache 2.0, vendor-reported on its Hugging Face model card.
Qwen 3.8 27B is a 27.8-billion-parameter dense vision-language model from Alibaba, released August 14, 2026, under Apache 2.0, vendor-reported on its Hugging Face model card. Only independent testing shows which trade holds, and that testing is still thin.
The Split in One Minute
Vendor tests favor Nanbeige on terminal and long-horizon agent work, with Qwen holding coding fixes and office tasks, all tabled below. Independent tests tie the pair on the headline index. Qwen alone has a live API and vision input. Sources: the Nanbeige model page and the Qwen model page.
- Nanbeige leads Terminal-Bench 2.1 and DeepSWE in vendor runs. See vendor table.
- Qwen holds SWE-Bench Pro and GDPval in vendor runs. See vendor table.
- Science reasoning is a wash in vendor runs. See vendor table.
- The independent headline index lands both models at 20, with splits tabled below.
- Only Qwen has public API pricing and vision input. See price and spec tables.
Specs Side by Side
A context window is how much text, code, and history the model can see at once. Agentic or tool-use work means the model calls tools, runs commands, and chains steps to finish a job.
| Spec | Nanbeige4.3-33B | Qwen 3.8 27B |
|---|---|---|
| Type | Dense, 33.1 billion parameters | Dense, 27.8 billion parameters plus vision encoder |
| Input | Text only | Text, image, and video |
| Context window | 524,288 tokens | 262,144 tokens native, extendable to about 1 million with YaRN |
| License | Apache 2.0 | Apache 2.0 |
| Thinking control | Thinking on by default, preserve_thinking recommended true for agents | Thinking on by default, reasoning_effort xhigh, medium, or low, preserve_thinking on by default |
| Release | October 24, 2026 | August 14, 2026 |
Nanbeige is the newer teacher model for a planned small-model line, vendor-reported on its Hugging Face card. Our Qwen 3.8 breakdown covers the Qwen family in depth, including the 2.4-trillion-parameter flagship.
Vendor Scoreboard: Where Nanbeige Leads and Where It Does Not
Every figure in this section is vendor-reported on the Nanbeige Hugging Face card, run at the highest thinking effort with the same harness per benchmark. No independent reproduction was found, verified by checking Artificial Analysis leaderboards in full. Treat gaps under one point as ties.
| Benchmark, vendor-reported | Nanbeige4.3-33B | Qwen 3.8 27B |
|---|---|---|
| Terminal-Bench 2.1 | 80.9 | 73.0 |
| Terminal-Bench 4.0 | 23.8 | 6.3 |
| DeepSWE | 65.5 | 42.5 |
| SWE-Bench Pro | 60.3 | 61.7 |
| GDPval, revised rubrics | 84.8 | 86.4 |
| GPQA Diamond | 89.5 | 89.2 |
| AIME 2026 | 95.2 | 94.5 |
| Tau3-Bench | 48.8 | 48.5 |
The pattern is clear when translated to work. Nanbeige takes terminal coding and long software runs, so command-line tasks need fewer retries. Qwen takes patch-style coding fixes and office cowork tasks. See table.
Two harness notes matter, both vendor-reported on the respective Hugging Face cards. Nanbeige ran SWE-Bench Pro with mini-swe-agent v2.4.6. Qwen ran SWE-Bench Pro in the Claude Code harness at 256K context. Different harnesses can shift scores, so the 1.4-point SWE-Pro gap is thin evidence on its own.
What Independent Testing Shows So Far
Independent coverage is partial, and that is the honest headline. Every figure in the table below is Artificial Analysis-verified as of November 9, 2026. Artificial Analysis lists both models in full.
| Independent headline, Artificial Analysis-verified | Nanbeige4.3-33B | Qwen 3.8 27B xhigh |
|---|---|---|
| Intelligence Index | 20 | 20 |
| Progress Index v1 | 306 | 303 |
| Cost per Intelligence Index task | No public API | $3.53 |
| Time per task | No public API | 39.2 min |
| Time to first answer | No public API | 122.64s |
| Output per Index task | 158K tokens | 110K tokens |
The pair lands on the same general ability score, but only Qwen publishes what that ability costs to serve.
The Artificial Analysis release breakdown sharpens the split. The table carries the sub-scores, all Artificial Analysis-verified November 9, 2026.
| Artificial Analysis | Nanbeige4.3-33B | Qwen 3.8 27B |
|---|---|---|
| Terminal-Bench 4.0 | 29.8% | 5.6% |
| Terminal-Bench-Science 0.1 | 1.0% | 0.5% |
| CritPt hard physics | 15.5% | 4.0% |
| AutomationBench-AA | 53.1% | 48.2% |
| AA-Briefcase Elo | 1162 | 1391 |
| GDP.pdf | 11.2% | 16.4% |
| FrontierCode 2.0 | 10.7% | 14.9% |
| HLE-Diamond | 6.9% | 15.3% |
| Hallucination rate | 18% | 29% |
| Correct answers | 12% | 17% |
| Attempt rate | 29% | 43% |
Nanbeige points at terminal execution with cautious answering, while Qwen points at breadth, code, and knowledge work. Nanbeige attempts 29 percent of questions against 43 percent for Qwen, Artificial Analysis-verified, which is why a low hallucination rate arrives with fewer correct answers.
Artificial Analysis is the only independent source with both models, and it scores them tied with different strengths.
Price and API Reality
Pricing is asymmetric. Qwen has a live menu. Nanbeige does not.
| Item | Nanbeige4.3-33B | Qwen 3.8 27B |
|---|---|---|
| Input per million tokens | No public price | $0.50 |
| Output per million tokens | No public price | $3.00 |
| Blended rate | Not listed, unverified | $0.47 at 7:2:1 cache, input, output mix |
| Cost per Intelligence Index task | Not listed, no public API | $3.53 |
Qwen is available through multiple API providers with cheaper third-party blends on some routes, Artificial Analysis-verified. Nanbeige shows weights for download with no provider table.
For broader market context on why Chinese open-weight labs price low, see our Qwen3.8-Flash pricing analysis.
Running Them at Home: Memory, Context, and Settings
Both models need real GPUs at full precision. Full weights are the model itself, and the KV cache is the running memory of the conversation, which grows with every token you feed in.
Qwen fits common local setups on consumer cards and Apple silicon, where speed depends mostly on memory bandwidth.
Nanbeige needs a datacenter-class card or multi-GPU setup for full quality. The table below carries the published weights and cache math.
| Local fit item | Nanbeige4.3-33B | Qwen 3.8 27B |
|---|---|---|
| Full weights | 66.2 GB BF16, estimated | 55.6 GB BF16, vendor-reported |
| KV cache per token | 213 KB FP16, estimated | 64 KB, third-party estimate |
| Cache at 8K / 32K | 1.7 GB / 7 GB, estimated | 0.5 GB / 2 GB, estimated |
For the same agent work, Nanbeige wants a larger card and a shorter context.
What Quantization Changes for Memory and Speed
Quantization means storing each weight in fewer bits, so the model takes less memory and moves fewer bytes per token. The figures below are arithmetic estimates from the parameter counts and the cache math above, not readings from any particular machine.
| Precision, estimated | Nanbeige4.3-33B | Qwen 3.8 27B |
|---|---|---|
| 8-bit weights | About 33 GB | About 28 GB |
| 8-bit practical memory at 8K context | About 35 GB, so a 40 to 48 GB card | About 29 GB, so a 32 GB card |
| 4-bit weights | About 17 GB | About 14 GB |
| 4-bit practical memory at 8K context | About 19 GB, workable on a 24 GB card | About 15 GB, workable on a 16 GB card |
| 4-bit practical memory at 32K context | About 24 GB, tight on a 24 GB card | About 16 GB, comfortable on a 24 GB card |
Speed tracks memory bandwidth more than raw compute, because every generated token reads the whole weight set once. Cutting weights from 16 bits to 4 bits cuts the bytes moved per token to roughly a quarter, which is why a 4-bit build usually decodes several times faster than full precision on the same card. That advantage narrows as context grows, since the cache reads start to compete with the weights.
Quality is the other side of the trade. Lower bit widths compress harder, and the loss tends to show up first on long reasoning chains and exact syntax, which are the two things agent work leans on most. An 8-bit build stays close to full precision for most jobs, 4-bit is the usual default, and below that the quality risk climbs faster than the memory savings justify.
Settings matter for both. Keep preserve_thinking on for agent loops, use temperature 1.0 with top_p 0.95 for agentic work, and set tensor parallelism to 4 for Nanbeige serving, all vendor-recommended on the Hugging Face cards. For tool calls, Nanbeige wants XML format while the qwen3 parsers under vLLM or SGLang suit Qwen.
Which One Fits Your Work
Match the model to the job, not the headline.
- Choose Nanbeige for text-only terminal agents and long tool chains if you have a larger card and can keep context short. See vendor and fit tables.
- Choose Qwen for vision tasks. See spec table.
- Choose Qwen for hosted API use today. See price table.
- Choose Qwen for tight local memory. See fit table.
- Choose neither on science research terminals until more data lands.
Vision deserves emphasis. Qwen handles STEM diagrams, documents, and hour-scale video, vendor-reported on its Hugging Face card. Nanbeige cannot process images, Artificial Analysis-verified on its model page. If your workflow pastes screenshots, charts, or photos, Qwen wins by default. For related context on small Qwen vision models, see our Qwen3-VL analysis.
Frequently Asked Questions
Is Nanbeige4.3-33B better than Qwen 3.8 27B?
Nanbeige leads most vendor agent tests but loses two. See vendor table. The pair ties the independent headline at 20. No single winner holds across all tasks.
Which is cheaper to use?
Nanbeige has no public API price found, unverified as of November 9, 2026, so the self-host cost is hardware plus power. Hosted use favors Qwen today.
Can I run them locally?
Yes, with different memory budgets. See fit table. Keeping context short keeps the cache small.
Does either handle images?
Qwen 3.8 27B takes text, image, and video, vendor-reported on its Hugging Face card. Nanbeige4.3-33B is text in and text out only, Artificial Analysis-verified. Pick Qwen for any visual workflow, since screenshots and charts rule Nanbeige out before scores matter.
What remains unverified?
The full vendor table has no independent reproduction found, verified by reading Artificial Analysis in full. Nanbeige pricing has no public source found, unverified. Treat vendor gaps as claims, not settled results.
Bottom Line
Nanbeige4.3-33B shows a sharp vendor-reported edge on terminal and long agent runs, paired with heavier memory needs and no public API, all vendor-reported or unverified as labeled above. Qwen 3.8 27B counters with vision input, a lighter memory footprint, live API pricing, and matching independent general scores at 20 on the Intelligence Index, Artificial Analysis-verified. Run your own task set before switching, keep thinking traces on for agents, and budget cache before chasing the 512K ceiling. The trade is real, but only your workload settles it.
