Nanbeige4.3-33B 33B Model for Coding and Agents: 12GB+ VRAM

At a Glance
| Item | Value |
|---|---|
| Repository | Nanbeige/Nanbeige4.3-33B |
| Family guide | Nanbeige4.3 guide (1 articles) |
| Published | 2026-10-24 |
| License | apache-2.0 |
| Formats | safetensors |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
Nanbeige4.3-33B has been released. This model is positioned as the flagship teacher model in the Nanbeige4.3 series, developed with the aim of providing state-of-the-art agent performance among open models of the same scale.
It excels in complex coding tasks and specialized collaborative work (cowork tasks), functioning not only as a standalone high-capability agent but also as a foundation for distilling advanced agent capabilities into smaller models planned for future release (such as Nanbeige4.3-3B).
Specifications
- Architecture: LlamaForCausalLM (llama)
- Context length: Up to 524,288 tokens (512K)
Performance
Comparison results published in the model card against major open models of the same scale (approximately 30B parameter band)—Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B—are shown below. All measurements were executed at each model’s highest thinking effort.
→ Scroll horizontally to see all columns
| Category | Benchmark | Nanbeige4.3-33B | Qwen3.8-27B | Muse-Glimmer-30B | Gemma4-31B |
|---|---|---|---|---|---|
| Agentic Coding | Terminal-Bench 2.1 | 80.9 | 73.0 | 61.8 | 47.2 |
| Agentic Coding | Terminal-Bench 4.0 | 23.8 | 6.3 | 1.6 | 0.0 |
| Agentic Coding | DeepSWE | 65.5 | 42.5 | 4.4 | 2.7 |
| Agentic Coding | SWE-Bench Pro | 60.3 | 61.7 | 36.9 | 51.2 |
| Agentic Coding | SWE-Atlas-QnA | 50.8 | 41.9 | 17.7 | 3.3 |
| Agentic Coding | NL2Repo | 53.0 | 48.3 | 27.4 | 33.8 |
| Agentic Coding | Scicode | 52.6 | 46.7 | 44.9 | 45.5 |
| Agentic Coding | SkillsBench | 61.1 | 56.4 | 39.7 | 28.9 |
| Agentic Cowork | Automation Bench | 54.2 | 38.3 | 15.2 | 9.7 |
| Agentic Cowork | JobBench | 44.8 | 34.5 | 16.6 | 9.0 |
| Agentic Cowork | GDPval | 84.8 | 86.4 | 75.8 | 71.4 |
| Agentic Cowork | Claw-Eval pass^3 | 78.3 | 75.2 | 57.3 | 44.2 |
| Agentic Cowork | Toolathlon-Verified | 70.4 | 68.5 | 47.2 | 27.8 |
| Agentic Cowork | DeepPlanning | 63.3 | 49.2 | 21.4 | 19.4 |
| Agentic Cowork | EnterpriseOps-Gym | 41.0 | 40.2 | 30.5 | 27.3 |
| Agentic Cowork | YC-bench | $1,025,437 | $641,290 | $513,978 | $265,152 |
| Agentic Cowork | Tau3-Bench | 48.8 | 48.5 | 33.7 | 18.6 |
| Reasoning | GPQA Diamond | 89.5 | 89.2 | 83.5 | 91.3 |
| Reasoning | AIME 2026 | 95.2 | 94.5 | 94.7 | 89.2 |
| Reasoning | AA-LCR | 84.0 | 82.0 | 83.3 | 72.0 |
According to the measurement results in the model card, Nanbeige4.3-33B significantly outpaces other models across many evaluation items. Particularly in the “Agentic Coding" domain, it records 80.9 on Terminal-Bench 2.1, which measures the ability to actually execute commands in a terminal to perform tasks, and 23.8 on the difficult Terminal-Bench 4.0, demonstrating a high advantage over comparison models. It also achieves high figures in Automation Bench, JobBench, and DeepPlanning within the “Agentic Cowork" field, showing strengths in agent-driven business support and automation tasks.
On the other hand, on SWE-Bench Pro, which measures the ability to fix real GitHub issues via repository editing, it scores 60.3, falling slightly short of Qwen3.8-27B’s 61.7. Similarly, Gemma4-31B (91.3) and Qwen3.8-27B (86.4) take the top spots on GDPval and GPQA Diamond (which handles PhD-level science problems), showing instances where it yields to other models on specific reasoning or evaluation metrics.
Additionally, regarding the base model Nanbeige4.3-33B-Base, comparison results with Gemma4-31B-Base are published on the model card. It is shown to consistently outperform Gemma4-31B-Base across STEM reasoning, mathematics, coding, and general knowledge benchmarks such as MMLU-Pro (66.9 vs 59.3), GPQA (62.0 vs 41.5), GSM8K (92.9 vs 88.9), and MBPP (74.6 vs 72.5).
Strengths and Use Cases
Nanbeige4.3-33B specializes in utilization as an agent handling complex coding and diverse specialized tasks. According to the model card, the model has strengths in agent tasks assuming a wide range of industries and professions, showing results that exceed existing baselines in the majority of the 35 professions evaluated in JobBench.
Specifically, it is suited for the following use cases and operating environments:
- Autonomous Coding Agents: Suited for terminal operations, repository analysis and modification, and executing multi-step development workflows.
- Specialized Business and Office Work Collaborative Tasks: Adapted for use cases that invoke external tools through multi-turn dialogue to assist with business automation and planning.
- Distillation Base for Small Agent Models: Positioned as the flagship teacher model in the Nanbeige4.3 series, with the purpose of distilling advanced capabilities into upcoming small agents such as Nanbeige4.3-3B explicitly stated.
Additionally, the model’s tokenizer incorporates a chat template supporting reasoning capabilities and tool calling, allowing flexible configuration according to the use case.
enable_thinking: Controls whether to output the reasoning process (thinking) in the current response (enabled by default). Set toFalsewhen normal responses are required.preserve_thinking: Specifies whether to retain previous thinking logs in multi-turn conversation history.Falseis recommended for general chat and QA, whileTrueis recommended for multi-turn workflows such as tool calling, office tasks, and code agents.- Tool Calling: Passing
toolsto the arguments enables the tool usage template.xml(most recommended) andjsonare available as tool calling formats.
Supported languages include English and Chinese tags, and it also supports long-text analysis and long-term task tracking utilizing a long context of up to 512K (524,288 tokens).
Regarding inference settings, the model card indicates recommended hyperparameters according to the task type:
- Agent and tool usage tasks: Temperature 1.0, Top_p 0.95, Max New Tokens 65,536
- Reasoning and dialogue tasks: Temperature 0.6, Top_p 0.95, Max New Tokens 262,144
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 33.1B parameters
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 12GB (RTX 4070 / 3060 12GB, etc.) | IQ2_XS | 9.9GB | 11.9GB |
| 16GB (RTX 5060 Ti 16GB / 4060 Ti 16GB, etc.) | Q2_K | 12.6GB | 15.1GB |
| 24GB (RTX 4090 / 3090, etc.) | Q4_1 | 19.7GB | 23.7GB |
| 32GB (RTX 5090, etc.) | Q6_K_S | 25.8GB | 31.0GB |
| 48GB (RTX 6000 Ada / A6000, etc.) | Q8_0 | 32.8GB | 39.3GB |
| 80GB class (A100 / H100) | BF16 | 61.7GB | 74.0GB |
Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: registered, vLLM: registered, MLX (mlx-lm): registered.
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. File sizes are measured from the converted build bartowski/Nanbeige_Nanbeige4.3-33B-GGUF. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
Runs in Ollama, LM Studio and llama.cpp via a converted build.
The publisher ships safetensors, but bartowski/Nanbeige_Nanbeige4.3-33B-GGUF provides a GGUF build you can use.
License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.
Compression: the IQ2_XS build measures 2.57 bits per weight — about 16% the size of the original 16-bit weights, calculated by this site from the actual file sizes.
Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Recent Models in the Same Size Class
Models with 15–40B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.
→ Scroll horizontally to see all columns
| Model | Parameters | Smallest VRAM tier | License | Our article |
|---|---|---|---|---|
| LiquidAI/LFM2.5-24B-A2B-Base | 24.0B | 80GB | lfm1.0 | LiquidAI Releases LFM2.5-24B-A2B for On-Device Agents (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B-FP8 | 24.0B | 32GB | lfm1.0 | LFM2.5-24B-A2B-FP8 Text Generation Model: 32GB+ VRAM (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B-MLX-bf16 | 24.0B | 80GB | lfm1.0 | LFM2.5-24B-A2B-MLX-bf16 Reasoning Model for Apple Silicon: 80GB+ VRAM (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B-MLX-8bit | 24.0B | 32GB | lfm1.0 | LFM2.5-24B-A2B-MLX-8bit Text Generation Model: 32GB+ VRAM (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B | 24.0B | 12GB | lfm1.0 | LFM2.5-24B-A2B Agent-Optimized MoE Reasoning Model: 12GB+ VRAM (2026-11-17) |
How to Get It
Nanbeige4.3-33B is available on Hugging Face under the Apache 2.0 license, and safetensors weights can be downloaded directly. No gating application or prior consent procedures are required for use.
Transformers, SGLang, and vLLM are supported as inference engines.
Example Execution with Transformers
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': 'Which number is bigger, 9.11 or 9.8?'}]
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)
response = tokenizer.decode(output_ids[0][inputs.shape[-1]:], skip_special_tokens=True)
print(response)
Example Launch with SGLang
MODEL_PATH=/path/to/your/Nanbeige4.3-33B
python -m sglang.launch_server \
--model-path ${MODEL_PATH} \
--host 0.0.0.0 \
--port 8000 \
--tp-size 4 \
--mem-fraction-static 0.8 \
--reasoning-parser qwen3 \
--tool-call-parser qwen3_coder
Example Launch with vLLM
MODEL_PATH=/path/to/your/Nanbeige4.3-33B
vllm serve ${MODEL_PATH} \
--host 0.0.0.0 \
--port 8000 \
--tensor-parallel-size 4 \
--gpu-memory-utilization 0.8 \
--enable-auto-tool-choice \
--reasoning-parser qwen3 \
--tool-call-parser qwen3_coder
When launching inference servers with SGLang or vLLM, it is guided to specify --reasoning-parser qwen3 and --tool-call-parser qwen3_coder as parser options.
Quantized and Converted Variants
→ Scroll horizontally to see all columns
| Added | Publisher | Format | Repository | Smallest VRAM tier (build, est. memory) |
|---|---|---|---|---|
| 2026-11-17 | bartowski | GGUF (imatrix) | bartowski/Nanbeige_Nanbeige4.3-33B-GGUF | IQ2_XS 11.9GB (fits in 12GB VRAM) |
| 2026-11-17 | mradermacher | GGUF (imatrix) | mradermacher/Nanbeige4.3-33B-i1-GGUF | IQ2_XS 11.4GB (fits in 12GB VRAM) |
| 2026-11-17 | Nanbeige | INT4 | Nanbeige/Nanbeige4.3-33B-INT4 | – |
| 2026-11-17 | Nanbeige | FP8 | Nanbeige/Nanbeige4.3-33B-FP8 | FP8 39.7GB (fits in 48GB VRAM) |
| 2026-11-17 | Nanbeige | INT8 | Nanbeige/Nanbeige4.3-33B-INT8 | INT8 40.2GB (fits in 48GB VRAM) |
File sizes of each build:
- Available builds in bartowski/Nanbeige_Nanbeige4.3-33B-GGUF: IQ2_XXS 9.4GB / IQ2_XS 9.9GB / IQ2_S 10.4GB / IQ2_M 11.4GB / Q2_K 12.6GB / IQ3_XXS 13.5GB / Q3_K_S 14.2GB / IQ3_XS 14.2GB / Q3_K_M 14.8GB / Q3_K_L 16.1GB / IQ3_M 16.6GB / IQ4_XS 17.4GB / Q4_0 18.0GB / Q4_K_S 18.4GB / Q4_K_M 19.4GB / IQ4_NL 19.4GB / Q4_1 19.7GB / Q4_K_L 20.8GB / Q5_K_S 22.0GB / Q5_K_M 23.4GB / Q6_K_S 25.8GB / Q6_K 27.0GB / Q6_K_L 28.3GB / Q8_0 32.8GB / BF16 61.7GB
- Available builds in mradermacher/Nanbeige4.3-33B-i1-GGUF: IQ1_S 7.1GB / IQ1_M 7.7GB / IQ2_XXS 8.7GB / IQ2_XS 9.5GB / IQ2_S 10.1GB / IQ2_M 10.9GB / Q2_K_S 11.0GB / Q2_K 11.8GB / IQ3_XXS 12.2GB / IQ3_XS 13.1GB / Q3_K_S 13.7GB / IQ3_S 13.7GB / IQ3_M 14.2GB / Q3_K_M 15.1GB / Q3_K_L 16.4GB / IQ4_XS 16.8GB / Q4_0 17.7GB / Q4_K_S 17.8GB / Q4_K_M 18.7GB / Q4_1 19.5GB / Q5_K_S 21.4GB / Q5_K_M 21.9GB / Q6_K 25.3GB
- Available builds in Nanbeige/Nanbeige4.3-33B-FP8: FP8 33.1GB
- Available builds in Nanbeige/Nanbeige4.3-33B-INT8: INT8 33.5GB
In addition, 42 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.
This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.
Related Articles
- Mage-VL-Cowork Vision-Language Model: 12GB+ VRAM, GGUF Builds
- Youtu-Parsing-Omni Omnimodal Analysis Model: 12GB+ VRAM
- Hemmingway-1 Text Generation Model: 12GB+ VRAM, GGUF Builds
- ZoomGain-4B: 4GB+ VRAM, GGUF Builds
What to Read Next
- Find models by VRAM (This model runs from the 12GB tier) → Other models that run on a 12GB GPU
- Explore the same model family → Nanbeige4.3 family overview (1 articles, 5 converted builds)
- Engines that run this model → llama.cpp / Ollama / vLLM
- What IQ2_XS, Q2_K, Q4_1 mean and where to get this model → GGUF format guide and models / FP8 format guide and models / Quantization and model-format glossary
- Other models for the same task → Other text generation models
Sources
Update History
- 2026-11-17: Added converted builds to “Quantized and Converted Variants”: bartowski/Nanbeige_Nanbeige4.3-33B-GGUF, mradermacher/Nanbeige4.3-33B-i1-GGUF, Nanbeige/Nanbeige4.3-33B-INT4, Nanbeige/Nanbeige4.3-33B-FP8, Nanbeige/Nanbeige4.3-33B-INT8
- 2026-11-17: Updated the hardware requirements table with the actual file sizes of bartowski/Nanbeige_Nanbeige4.3-33B-GGUF.
- 2026-11-17: Changed the title to show what the article covers (VRAM requirements, file list, etc.).

