Nanbeige4.3-33B

Nanbeige4.3-33B

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

Pricing

Nanbeige4.3-33B is released as open weights (self-hosted); no hosted API pricing has been announced.

Public Benchmark Scores

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

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

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 · Ornith-1.5-397B
• Benchmarks: Terminal-Bench 2.1 · DeepSWE · GPQA Diamond

Frequently Asked Questions

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