Nanbeige4.3-33B

reasoning
coding

Nanbeige4.3-33B is an open dense language model with 33.1 billion parameters, developed by the Nanbeige laboratory and released under the Apache 2.0 license. Technically, it is a classic decoder based on the LlamaForCausalLM architecture, comprising 52 layers with a hidden dimension of 7168. Its attention mechanism is based on Grouped-Query Attention (GQA) with 56 query heads and 8 key/value heads. The model uses SwiGLU activation in the MLP layers, a vocabulary of 166,912 tokens, and a context window of 524,288 tokens (512K).

The developers deliberately avoided using hybrid attention types and the innovative Looped Transformer architecture itself, which had been introduced in the previous Nanbeige4.2-3B version. The reason is that the looped architecture is well suited to compact models, as it allows a balance between quality and computational cost through layer reuse. However, for a 33B model this method is not suitable, and the emphasis was placed on a classic dense Transformer with GQA. This decision provides much greater output stability and predictability of reasoning, which is a critically important property, since this model is positioned as a “teacher model.” Its answers and reasoning trajectories can be used for knowledge distillation into smaller systems, so architectural simplicity and reliability are more important here than experiments with sparse attention.

The uniqueness of Nanbeige4.3-33B lies not only in its size but also in its specialization. Unlike the previous versions 4.0, 4.1, and 4.2 (these are 3B-parameter models), this model was initially designed for solving agentic tasks and automation. It supports flexible control of the thinking mode via the enable_thinking and preserve_thinking parameters, which allows the model to explicitly build a chain of reasoning and preserve it in the dialogue history for multi-step scenarios.

According to the provided benchmarks, Nanbeige4.3-33B demonstrates impressive results in agentic tasks, confidently outperforming competitors such as Qwen3.8-27B and Gemma4-31B. For example, in Terminal-Bench 2.1 (which tests an agent’s ability to work in the command line), the model scores 80.9 points versus 73.0 for Qwen3.8-27B. In the DeepSWE benchmark (solving complex engineering tasks), it reaches 65.5 points, more than one and a half times surpassing Qwen3.8-27B (42.5), and in Automation Bench (office process automation) — 54.2 points versus 38.3. At the same time, for the sake of objectivity, it is worth noting that in some tests competitors still outperform Nanbeige4.3-33B (Qwen3.8-27B is slightly better in SWE-Bench Pro with a score of 61.7 versus 60.3, and Gemma4-31B leads in GPQA Diamond — 91.3 versus 89.5).

Nanbeige4.3-33B is ideally suited for creating autonomous programmer agents capable of writing code, debugging it, and managing repositories directly from the terminal. Strong results in Automation Bench and JobBench make it an excellent core for digital assistants in office work (Agentic Cowork), capable of performing multi-step tasks and calling external tools. The 512K-token context allows the model to be used for analyzing substantial data arrays. And, of course, the model’s main purpose is to act as a generator of high-quality data (teacher model) for training compact versions.


Announce Date: 24.10.2026
Parameters: 34B
Context: 525K
Layers: 52
Attention Type: Full Attention
Developer: Nanbeige
Transformers Version: 4.51.1
vLLM Version: >=0.17.0
License: Apache 2.0

Public endpoint

Use our pre-built public endpoints for free to test inference and explore Nanbeige4.3-33B capabilities. You can obtain an API access token on the token management page after registration and verification.
Model Name Context Type GPU Status Link

API access to Nanbeige4.3-33B endpoints

There are no public endpoints for this model yet.

Private endpoint

An endpoint is a server with pre-installed software that automatically downloads neural network weights and launches an OpenAI-compatible API to handle incoming requests.

A private endpoint is created within the user's project, belongs exclusively to the user, and is billed per server configuration in accordance with the infrastructure pricing.

We recommend creating private endpoints when you need to:

  • achieve maximum endpoint performance,
  • access the full context window for processing large requests and conversations,
  • ensure maximum security by processing data in a dedicated, isolated environment.

Recommended server configurations for creating a private inference endpoint Nanbeige4.3-33B

Prices:
Name GPU Price, hour TPS Max Concurrency
rtx3090-4.16.96.320
262,144.0
tensor
4 $2.97 1.000 Launch
rtx3090-4.16.128.160.nvlink
262,144.0
tensor
4 $3.01 1.000 Launch
h100nvl-1.16.96.160
262,144.0
1 $4.11 1.177 Launch
rtx5090-3.16.96.160
262,144.0
pipeline
3 $4.34 1.008 Launch
teslaa100-2.24.96.160.nvlink
262,144.0
tensor
2 $4.61 2.269 Launch
teslaa100-2.24.128.160.nvlink
524,288.0
tensor
2 $4.67 1.134 Launch
h200-1.16.128.160
524,288.0
1 $4.74 1.002 Launch
teslaa100-2.24.256.160
524,288.0
tensor
2 $4.93 1.134 Launch
rtx5090-4.16.128.160
262,144.0
tensor
4 $5.74 1.552 Launch
rtx4090-6.44.256.160
262,144.0
pipeline
6 $5.83 1.620 Launch
dedicated-rtx3090-8.64.128.960-1
524,288.0
tensor
8 $6.04 1.202 Launch
rtx4090-8.44.256.160
524,288.0
tensor
8 $7.51 1.199 Launch
h100-2.24.256.160
524,288.0
tensor
2 $7.84 1.133 Launch
h100nvl-2.24.192.240
524,288.0
tensor
2 $8.17 1.379 Launch
rtx5090-6.44.256.160
524,288.0
pipeline
6 $8.86 1.210 Launch
h200-2.24.256.160.nvlink
524,288.0
tensor
2 $9.40 2.206 Launch
rtx5090-8.44.256.160
524,288.0
tensor
8 $11.54 1.754 Launch
Prices:
Name GPU Price, hour TPS Max Concurrency
teslaa100-2.24.96.160.nvlink
262,144.0
tensor
2 $4.61 2.037 Launch
h200-1.16.128.160
262,144.0
1 $4.74 1.772 Launch
teslaa100-2.24.192.160.nvlink
524,288.0
tensor
2 $4.80 1.018 Launch
teslaa100-2.24.256.160
524,288.0
tensor
2 $4.93 1.018 Launch
rtx5090-4.16.128.160
262,144.0
tensor
4 $5.74 1.319 Launch
rtx4090-6.44.256.160
262,144.0
pipeline
6 $5.83 1.388 Launch
dedicated-rtx3090-8.64.128.960-1
262,144.0
tensor
8 $6.04 2.172 Launch
rtx4090-8.44.256.160
524,288.0
tensor
8 $7.51 1.083 Launch
h100-2.24.256.160
524,288.0
tensor
2 $7.84 1.017 Launch
h100nvl-2.24.192.240
524,288.0
tensor
2 $8.17 1.263 Launch
rtx5090-6.44.256.160
524,288.0
pipeline
6 $8.86 1.094 Launch
h200-2.24.256.160
524,288.0
tensor
2 $9.40 2.090 Launch
h200-2.24.256.160.nvlink
524,288.0
tensor
2 $9.40 2.090 Launch
rtx5090-8.44.256.160
524,288.0
tensor
8 $11.54 1.637 Launch
Prices:
Name GPU Price, hour TPS Max Concurrency
h200-1.16.128.240
262,144.0
1 $4.74 1.221 Launch
teslaa100-2.24.256.240
262,144.0
tensor
2 $4.93 1.486 Launch
teslaa100-2.24.256.320.nvlink
262,144.0
tensor
2 $4.94 1.486 Launch
dedicated-rtx3090-8.64.128.960-1
262,144.0
tensor
8 $6.04 1.622 Launch
teslaa100-3.32.384.240
524,288.0
pipeline
3 $7.36 1.337 Launch
rtx4090-8.44.256.240
262,144.0
tensor
8 $7.52 1.617 Launch
h100-2.24.256.240
262,144.0
tensor
2 $7.85 1.484 Launch
h100nvl-2.24.192.240
262,144.0
tensor
2 $8.17 1.976 Launch
rtx5090-6.44.256.240
262,144.0
pipeline
6 $8.86 1.638 Launch
teslaa100-4.16.256.240
524,288.0
tensor
4 $9.14 2.080 Launch
h200-2.24.256.240
524,288.0
tensor
2 $9.41 1.814 Launch
h200-2.24.256.240.nvlink
524,288.0
tensor
2 $9.41 1.814 Launch
teslaa100-4.32.384.320.nvlink
524,288.0
tensor
4 $9.50 2.080 Launch
rtx5090-8.44.256.240
524,288.0
tensor
8 $11.55 1.362 Launch
h100-3.32.384.240
524,288.0
pipeline
3 $11.73 1.335 Launch
h100nvl-3.24.384.480
524,288.0
pipeline
3 $12.38 1.691 Launch
h100-4.16.256.240
524,288.0
tensor
4 $14.96 2.077 Launch
h100nvl-4.32.384.320
524,288.0
tensor
4 $16.22 2.569 Launch

Related models

Need help?

Contact our dedicated neural networks support team at support@immers.cloud or send your request to the sales department at sale@immers.cloud.

We use cookies and web analytics services to ensure the proper functioning of the website, analyze web-site traffic, and improve the quality of our services.
By continuing to use the website, you consent to the Privacy Policy and consent to the processing of cookie files and technical data.