TL;DR
This dataset compiles the model card data of Nanbeige4.3-33B, open-sourced on October 25, 2026 by BOSS Zhipin's Nanbeige lab: 33.1B dense parameters, BF16 weights, 512K context, Apache 2.0 license, a self-reported 80.9 on Terminal-Bench 2.1, leading 17 of 20 benchmark rows against three ~30B-class open models.
Data Description
Data sources are the official model card of Nanbeige LLM Lab and the Hugging Face repositories (released October 25, 2026), compiled by MarkTechPost and others. Comparison figures are self-reported by Nanbeige; the comparison models are Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B. Competitor values come from each lab's official model card and differ in places from Nanbeige's table (due to different harnesses); treat them as provisional until independent runs appear.
Core Data
Model Specification
| model | Description | Value |
|---|---|---|
| model_name | Model name | Nanbeige4.3-33B |
| parameter_count | Parameters | 33.1B (dense, non-MoE) |
| precision | Weight precision | BF16 |
| context_window | Context window | 524,288 tokens (512K) |
| license | License | Apache 2.0 |
| architecture | Architecture | LlamaForCausalLM (52 layers, hidden 7168, GQA 56Q/8KV, vocab 166,912) |
| variants | Released versions | Nanbeige4.3-33B, Nanbeige4.3-33B-Base |
| release_date | Release date | 2026-10-25 |
Benchmarks (self-reported by Nanbeige)
| field | Description | Value |
|---|---|---|
| terminal_bench_21 | Terminal-Bench 2.1 (agentic coding) | 80.9 |
| deep_swe | DeepSWE | 65.5 |
| swe_bench_pro | SWE-Bench Pro | 60.3 |
| automation_bench | Automation Bench (agentic office) | 54.2 |
| job_bench | JobBench | 44.8 |
| aime_2026 | AIME 2026 (math reasoning) | 95.2 |
| gpqa_diamond | GPQA Diamond | 89.5 |
Comparison vs. Competitors (self-reported, partial)
| field | Description | Nanbeige4.3-33B | Qwen3.8-27B |
|---|---|---|---|
| terminal_bench_21 | Terminal-Bench 2.1 | 80.9 | 73.0 |
| deep_swe | DeepSWE | 65.5 | 42.5 |
| swe_bench_pro | SWE-Bench Pro | 60.3 | 61.7 |
| yc_bench_earnings | YC-bench simulated earnings (USD) | 1,025,437 | 641,290 |
Field Descriptions
- Dense: no expert routing (MoE); all 33B parameters compute on every token.
- BF16 memory: per config.json, weights need about 66 GB VRAM; a full 512K-sequence KV cache is about 112 GB.
- Deployment: the model card provides Transformers, SGLang and vLLM recipes; server examples use 4-way tensor parallelism.
- Self-reported scope: all comparison scores were measured by Nanbeige under shared hyperparameters and harnesses; some competitor official card values differ from Nanbeige's table and await independent verification.