# Nanbeige4.3-33B ![Organization](https://img.shields.io/badge/Org-Nanbeige-555) ![Context](https://img.shields.io/badge/Context-512K%20tokens-1f6feb) ![License](https://img.shields.io/badge/License-Apache--2.0-2da44e) ![Modality](https://img.shields.io/badge/Modality-Text-8957e5) ![Released](https://img.shields.io/badge/Released-Oct%202026-555) > **Quick answer:** Nanbeige4.3-33B is Nanbeige Team's flagship teacher model in the Nanbeige4.3 series — a 33B open-weight (Apache-2.0) model designed to anchor distillation into smaller agentic models. It supports up to 512K tokens of context and leads most same-class (~30B) open rivals on agentic coding and cowork benchmarks, scoring 80.9% on Terminal-Bench 2.1, 65.5% on DeepSWE, and 95.2% on AIME 2026. ## At a Glance **Where Nanbeige4.3-33B leads** - 80.9% Terminal-Bench 2.1 and 23.8% Terminal-Bench 4.0 — ahead of Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B - 65.5% DeepSWE — clear lead over same-class rivals (next best 42.5%) - 95.2% AIME 2026 and 84.0% AA-LCR — strong reasoning/long-context scores at this parameter count - Designed as a teacher model for distilling agentic capability into smaller Nanbeige4.3 models (e.g. an upcoming 3B) **Where it lags** - 60.3% SWE-Bench Pro — edged out by Qwen3.8-27B (61.7%) - 84.8% GDPval — behind Qwen3.8-27B (86.4%) - GPQA Diamond (89.5%) trails Gemma4-31B (91.3%) **Best for:** teams wanting an open-weight, ~30B-class agentic coding/cowork model with a long context window, or a base for distilling smaller task-specific agents. ## What Nanbeige4.3-33B Is Nanbeige4.3-33B is the flagship model of Nanbeige Team's Nanbeige4.3 series, positioned as a high-capacity "teacher" model whose agentic capabilities are intended to be distilled into smaller models later in the series (e.g. a planned 3B variant). Rather than chasing raw knowledge benchmarks, the release focuses on agentic coding and "cowork" tasks — multi-step, tool-using work across a wide range of professional occupations. The model supports a 524,288-token (512K) context window and ships a configurable chat template with `enable_thinking` and `preserve_thinking` controls for reasoning and multi-turn tool use. Nanbeige reports it outperforms other open models in the ~30B-parameter class (Qwen3.8-27B, Muse-Glimmer-30B, Gemma4-31B) on most of the agentic benchmarks it reports, including a notable lead on JobBench, where it reportedly outperforms baselines across most of 35 evaluated occupations. ## Specifications | Field | Value | |---|---| | Organization | Nanbeige | | Parameters | 33B (dense) | | Context window | 524,288 tokens (512K) | | Architecture | Dense transformer (llama-family tags per HF card) | | License | Apache-2.0 | | Release date | 2026-10-25 | | Modality | Text | ## Pricing Nanbeige4.3-33B is released as open weights (self-hosted); no hosted API pricing has been announced. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [Terminal-Bench 2.1](/benchmarks/tbench/terminalbench-21) | 80.9 | [Nanbeige4.3-33B model card](https://huggingface.co/Nanbeige/Nanbeige4.3-33B) | 2026-10 | | [Terminal-Bench 4.0](/benchmarks/anthropic/terminal-bench-40) | 23.8 | [Nanbeige4.3-33B model card](https://huggingface.co/Nanbeige/Nanbeige4.3-33B) | 2026-10 | | [DeepSWE](/benchmarks/datacurve/deepswe) | 65.5 | [Nanbeige4.3-33B model card](https://huggingface.co/Nanbeige/Nanbeige4.3-33B) | 2026-10 | | [SWE-Bench Pro](/benchmarks/swe-bench-team/swe-bench-pro) | 60.3 | [Nanbeige4.3-33B model card](https://huggingface.co/Nanbeige/Nanbeige4.3-33B) | 2026-10 | | [SWE Atlas – Codebase Q&A](/benchmarks/raghavendra-et-al-scale-ai/swe-atlas-codebase-qa) | 50.8 | [Nanbeige4.3-33B model card](https://huggingface.co/Nanbeige/Nanbeige4.3-33B) | 2026-10 | | [NL2Repo](/benchmarks/nl2repo) | 53.0 | [Nanbeige4.3-33B model card](https://huggingface.co/Nanbeige/Nanbeige4.3-33B) | 2026-10 | | [SciCode](/benchmarks/scicode) | 52.6 | [Nanbeige4.3-33B model card](https://huggingface.co/Nanbeige/Nanbeige4.3-33B) | 2026-10 | | [SkillsBench](/benchmarks/benchflow-team-advised-by-dawn-song/skillsbench) | 61.1 | [Nanbeige4.3-33B model card](https://huggingface.co/Nanbeige/Nanbeige4.3-33B) | 2026-10 | | [AutomationBench](/benchmarks/automationbench) | 54.2 | [Nanbeige4.3-33B model card](https://huggingface.co/Nanbeige/Nanbeige4.3-33B) | 2026-10 | | [JobBench](/benchmarks/jobbench) | 44.8 | [Nanbeige4.3-33B model card](https://huggingface.co/Nanbeige/Nanbeige4.3-33B) | 2026-10 | | [Claw-Eval](/benchmarks/cross-lab-agent-evaluation-reference-cited-by-meta-kimi-qwen-tencent/claw-eval) | 78.3 | [Nanbeige4.3-33B model card](https://huggingface.co/Nanbeige/Nanbeige4.3-33B) (General subset only, 157 tasks) | 2026-10 | | [Toolathlon-Verified](/benchmarks/hkust-nlp/toolathlon-verified) | 70.4 | [Nanbeige4.3-33B model card](https://huggingface.co/Nanbeige/Nanbeige4.3-33B) | 2026-10 | | [GPQA Diamond](/benchmarks/rein-et-al/gpqa-diamond) | 89.5 | [Nanbeige4.3-33B model card](https://huggingface.co/Nanbeige/Nanbeige4.3-33B) | 2026-10 | | [AIME 2026](/benchmarks/mathematical-association-of-america/aime-2026) | 95.2 | [Nanbeige4.3-33B model card](https://huggingface.co/Nanbeige/Nanbeige4.3-33B) | 2026-10 | | [AA-LCR](/benchmarks/artificial-analysis/aa-lcr) | 84.0 | [Nanbeige4.3-33B model card](https://huggingface.co/Nanbeige/Nanbeige4.3-33B) | 2026-10 | *Scores above are reported by Nanbeige Team and shown for context. They are not Benchgen measurements. Figures depend on harness, tools, and effort settings — see the source for methodology.* **Excluded from the table above (bucket-C, documented per skill procedure):** - **GDPval** — vendor reports 84.8 on a 0–100 "revised rubrics" judge score; our local `gdpval-aa-v2` entity is Artificial Analysis's Elo-rating methodology (anchored to a human baseline of 1,000) — different scale/methodology, not comparable. Excluded rather than mis-mapped. - **DeepPlanning**, **EnterpriseOps-Gym** — no local Benchgen entity exists for either; vendor table gives only a score and "evaluated using their official repositories," with no paper/announcement URL discoverable — insufficient public methodology info to stand up a new benchmark config. - **YC-bench** — metric is a dollar figure ($1,025,437), not a comparable percentage/scale; no local entity, and the metric type doesn't fit our standard score format. - **Tau3-Bench** (plain, as named by vendor) — ambiguous generic name vs. our existing domain-specific `tau3-banking` entity; per standing policy, never silently narrow/widen scope to a domain-specific match. Excluded, not guessed. ## Nanbeige4.3-33B vs Alternatives | Model | Context | Terminal-Bench 2.1 | GPQA Diamond | Price (in/out per 1M) | |---|---|---|---|---| | **Nanbeige4.3-33B** | 512K | 80.9 | 89.5 | Open weights | | [Qwen3.8-27B](/models/alibaba/qwen3-8-27b) | — | 73.0 | 89.2 | — | | [Ornith-1.5-397B](/models/ornith/ornith-1-5-397b) | 262K (1M w/ YaRN) | 86.1 | 92.8 | Open weights | Nanbeige4.3-33B trades flagship-scale ceiling (vs. a 397B model like Ornith-1.5) for strong same-class ~30B agentic performance, open weights, and a long 512K native context window. ## How Nanbeige4.3-33B Performs on Real Agent Tasks Nanbeige4.3-33B's benchmark spread leans heavily toward agentic coding (Terminal-Bench, DeepSWE, SWE-Bench Pro, SWE-Atlas) and agentic "cowork" tasks (AutomationBench, JobBench, Toolathlon-Verified) rather than pure knowledge recall — consistent with its stated design goal as a teacher model for distilling agentic skills. A public leaderboard score establishes general task-completion capability; it does not establish reliability on a specific domain's repeated workflows, which is what builders evaluating a self-hosted ~30B coding/cowork model should verify against their own task distribution. ## Use Nanbeige4.3-33B via API ```python from transformers import AutoModelForCausalLM, AutoTokenizer model_id = "Nanbeige/Nanbeige4.3-33B" tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=False, trust_remote_code=True) model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto", trust_remote_code=True) messages = [{"role": "user", "content": "Summarize this contract clause..."}] inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device) output_ids = model.generate(inputs, max_new_tokens=1024) print(tokenizer.decode(output_ids[0][inputs.shape[-1]:], skip_special_tokens=True)) ``` ## Related Models & Benchmarks - Compare: [Qwen3.8-27B](/models/alibaba/qwen3-8-27b) · [Ornith-1.5-397B](/models/ornith/ornith-1-5-397b) - Benchmarks: [Terminal-Bench 2.1](/benchmarks/tbench/terminalbench-21) · [DeepSWE](/benchmarks/datacurve/deepswe) · [GPQA Diamond](/benchmarks/rein-et-al/gpqa-diamond) ## Frequently Asked Questions

What is Nanbeige4.3-33B?

It's Nanbeige Team's flagship 33B open-weight teacher model in the Nanbeige4.3 series, designed for agentic coding and professional cowork tasks and intended to anchor distillation into smaller models.

What is Nanbeige4.3-33B's context window?

Up to 524,288 tokens (512K) — roughly 400,000 words of text in a single request.

How much does Nanbeige4.3-33B cost?

It's released under an open-weight license — free to download; you pay only your own hosting cost.

Is Nanbeige4.3-33B open source?

Yes — Apache-2.0.

What is Nanbeige4.3-33B's knowledge cutoff?

Not disclosed by Nanbeige Team at release.

--- *Specs and scores sourced from Nanbeige Team's official Hugging Face model card; third-party benchmark scores attributed inline. Last updated 2026-10-26.*

Nanbeige4.3-33B — Evaluations

Evaluations