AI-assisted summary, reviewed by NullTX editors.

Nanbeige4.3-33B matches Qwen3.8-27B at 20 on the Artificial Analysis Intelligence Index but uses 158K output tokens per task against 110K. Here is what that costs locally.

Two models just posted the same score. One needed 158,000 output tokens per task to get there. The other needed 110,000.

Nanbeige4.3-33B, a dense 33-billion-parameter reasoning model from the foundation model team at BOSS Zhipin, scores 20 on the Artificial Analysis Intelligence Index v5. According to Artificial Analysis's evaluation, published November 5, that matches Qwen3.8 27B at its xhigh reasoning setting. The model is open weights under Apache 2.0, accepts text only, and has a 524K-token context window.

Sources indicate that behind the tie sits a gap in effort. Artificial Analysis notes Nanbeige used 158K output tokens per Intelligence Index task against 110K for Qwen3.8 27B, about 44% more. That weighted average covers answer and reasoning tokens across all nine v5 evaluations, including agentic tool-use turns. For anyone running these models, tokens translate into time, electricity or money, so the same score doesn't mean the same cost.

The Intelligence Index v5 combines nine evaluations, among them AA-Briefcase, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode 2.0 and AA-Omniscience. A score of 20 puts Nanbeige well above the 6 of Muse Glimmer, a 30B model in the same comparison. Artificial Analysis also notes that Nanbeige sets a newly introduced Pareto-optimal point for Intelligence Index against total parameters, since both it and Qwen3.8 27B are dense models. A larger mixture-of-experts model like DeepSeek V4.1 Flash, by contrast, scores 33 while activating only 16B parameters per token.

The model card Nanbeige's lab published on Hugging Face describes the base model, positioned as the flagship teacher for the Nanbeige4.3 series. The lab notes it will distill agentic capabilities into smaller models, such as an upcoming 3B version. On base-model benchmarks, the card demonstrates 66.9 on MMLU-Pro against 59.3 for Gemma4-31B-Base. Those figures belong to the base model and are separate from the reasoning model's Intelligence Index result.

According to recent reports, the two models reach 20 by different routes. Nanbeige scores 30% on Terminal-Bench 4.0 against 6% for Qwen, and 16% on the CritPt v2 Hard Subset against 4%. It also edges Qwen on AutomationBench-AA, 53% to 48%.

According to recent reports, elsewhere, it falls behind. On AA-Briefcase v1.1 it scores 1162 Elo against 1391, on FrontierCode 2.0 11% against 15%, and on HLE-Diamond 7% against 15%. The knowledge test tells the most distinctive story. On AA-Omniscience v1.1, Nanbeige scores -4 against Qwen's -6, with an 18% hallucination rate against 29%. Yet it answers only 29% of questions against 43%, declines 71% of them and gets 12% correct against 17%.

For general question answering, Qwen's higher accuracy is the better result.

According to recent reports, output tokens aren't a vanity metric. A reasoning model writes out its thinking as tokens, and every one costs decode time. If both models ran at the same speed on the same hardware, Nanbeige would take about 44% longer to finish a typical task.

Local decoding adds a second factor. Generation is usually limited by memory bandwidth, because each new token requires reading the model's weights, and a 33B dense model reads about 20% more weight data per token than a 27B dense model. Assuming the same quantization and the same memory-bound hardware, those two effects compound to roughly 1.7 times the wall-clock time per task. That figure is a back-of-the-envelope calculation, not a measured result.

Recent findings suggest that nor is there a hosted alternative to soften the blow. Artificial Analysis lists no confirmed public first-party or third-party APIs for Nanbeige, so unlike Qwen, which is sold per token through Alibaba's API, the cost is the hardware you run it on and the time it takes.

Industry observers note that artificial Analysis describes the model as small enough to run locally on a laptop, which holds true with a caveat. The weights are published in BF16, which at 33 billion parameters means roughly 66 GB, far beyond almost any laptop. At 8-bit, the weights fall to about 33 GB, and at 4-bit to about 17 GB, plus memory for the context cache. A machine with 24 GB of unified memory or more can plausibly hold a 4-bit build, though a long context will push it.

It has been revealed that the vendor's own quickstart points the other way. The card's SGLang and vLLM examples both set tensor parallelism across four GPUs, which suggests the reference deployment is a multi-GPU server. Its Transformers example also enables trust_remote_code=True, which runs code from the repository. Reading that code first is the safer move, as is serving the model through vLLM or SGLang instead.

Quantization changes the model, too. The Intelligence Index numbers come from Artificial Analysis's own runs, so a 4-bit local build may score somewhat lower, or generate different token counts, than the version they evaluated.

Sources indicate that this release is a clean example of why a headline score isn't enough. Nanbeige delivers Qwen-level results at 33B, with strengths in terminal tasks and physics reasoning and a more cautious answering style. The price for those strengths is length.

Anyone choosing between the two should run both on their own workload and record three numbers: the score on their tasks, output tokens per task and seconds per task on their hardware. For terminal-heavy, agentic work, Nanbeige's profile may justify the longer runs. For knowledge-heavy or latency-sensitive work, Qwen's shorter, more accurate answers look stronger.

In a recent development, three signals are worth watching next: whether the 3B distilled version keeps the model's strengths, whether other labs report the same token behavior, and whether anyone publishes quantized local benchmarks. Until then, the simplest summary is also the most accurate: same score, different bill.

In a recent development, disclosure: This is not trading or investment advice. Always do your research before buying any cryptocurrency or investing in any services. Follow us on XΒ @nulltxnews.

Industry observers note that webAI's open-weight TwIL-8B posts 98.2 on MATH-500 and indicates it tops Gemma 4 31B. The card's own footnotes show why the comparison deserves a re-run.A model small enough to run from a roughly 5 GiB file on a CPU now claims a 98.2 on MATH-500. The most revealing line in its model card isn't a score, though. It's a footnote saying the competitor numbers were ne….

One email a week with our newly introduced explainers and the news behind them.

More explainers and analysis from NullTX.

β†’ Read original article