ModelsArchitectures & capability
Nanbeige releases Nanbeige4.3-33B, an Apache-2.0 open-weight model built for agentic coding
Nanbeige, the AI team backed by BOSS Zhipin, has released a dense 33B open-weight model under Apache 2.0. It is built for agentic coding, has a 512K-token context, and is positioned as a teacher for smaller distilled models.
huggingface.co (article, date unverified): On agentic benchmarks, Nanbeige reports clear leads over Qwen3.8-27B: 65.5 vs 42.5 on DeepSWE and 80.9 vs 73.0 on Terminal-Bench 2.1. [1] [2]
SOURCES[1] Nanbeige4.3-33B model card, Hugging Face huggingface.co[2] Nanbeige's New 33B Open Model Takes Aim at Qwen and Gemma… startupfortune.com
AT Pulse research deskPublished · Updated · 4 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- On agentic benchmarks, Nanbeige reports clear leads over Qwen3.8-27B: 65.5 vs 42.5 on DeepSWE and 80.9 vs 73.0 on Terminal-Bench 2.1. [1] [2]
- Original source
- huggingface.co article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- Studies
- arXiv 2602.13367 (Feb 2026) [3] date from arXiv ID
- When
- Original publication date unverified
- Covered
- Oct 26 edition
- Evidence
- 7 sources across 6 sites
- 1Nanbeige4.3-33B model card, Hugging Facehuggingface.co
- 2Nanbeige's New 33B Open Model Takes Aim at Qwen and Gemma…startupfortune.com
- 3Nanbeige4.1-3B: A Small General Model that Reasons…arxiv.org
- 4Nanbeige4.3-33B release note, AI TL;DRai-tldr.dev
- 5Nanbeige Releases Nanbeige4.3-33B, MarkTechPostmarktechpost.com
- 6Nanbeige4.3-33B, NVIDIA NeMo AutoModel docsdocs.nvidia.com
QuietHow it’s measured
Hacker News · 3 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
On agentic benchmarks, Nanbeige reports clear leads over Qwen3.8-27B: 65.5 vs 42.5 on DeepSWE and 80.9 vs 73.0 on Terminal-Bench 2.1. No one has independently reproduced these scores. The model does not win everywhere: it trails Qwen3.8-27B on SWE-Bench Pro (60.3 vs 61.7) and Gemma4-31B on GPQA Diamond (89.5 vs 91.3). [1] [2]
The model card describes the 33B model as a flagship teacher, with a distilled Nanbeige4.3-3B student planned. This builds on the lab's earlier small agentic models, such as Nanbeige4.1-3B, whose paper claimed tool-use runs of up to 600 turns. [1] [3] [7]
The architecture is a standard Llama-style design with 52 layers and grouped-query attention, so existing tooling should work. The card includes serving examples for Transformers, SGLang and vLLM, plus switchable reasoning traces. NVIDIA's NeMo AutoModel docs list full-parameter SFT and LoRA fine-tuning recipes for H100 nodes. [1] [6]
a permissively licensed 33B model with standard Llama architecture and existing fine-tuning recipes.
Read the full assessment
Implication: if its agentic-coding margins hold on teams' own repositories and scaffolds, it could make self-hosted coding agents a credible alternative to API models.
Executive brief
Nanbeige, an AI team backed by the Chinese recruiting platform BOSS Zhipin, has released a dense 33B open-weight model under Apache 2.0. It reports 80.9 on Terminal-Bench 2.1 and 65.5 on DeepSWE. The second figure is 23 points above the 42.5 attributed to Qwen3.8-27B (Startup Fortune). All of these scores come from Nanbeige; the reviewed sources contain no independent reproduction. The model card names it a "flagship teacher" meant to be distilled into smaller agentic models. It supports 512K tokens of context and has switchable reasoning traces (Hugging Face model card). The card lists no limitations.
What changed and event timeline
Nanbeige4.1-3B paper
Nanbeige posted a technical report for a 3B model. It claims tool-use runs of up to 600 turns and wins over Qwen3-4B, which shows the lab's agentic focus before 4.3 ().
Nanbeige4.3-33B released (reported date)
Secondary coverage gives this date for the Apache-2.0 release on Hugging Face. The model card itself has no release date (;).
Trade-press coverage
MarkTechPost covered the release as an open agentic coding model (). Startup Fortune pointed out the benchmarks where the model loses ().
- Also
Date unknown: NVIDIA NeMo AutoModel support
NVIDIA's nightly documentation lists the model with full-parameter SFT and LoRA fine-tuning recipes for H100 nodes ().
Capabilities and access
- Model: Nanbeige4.3-33B, dense, about 33.1B parameters, BF16 (NVIDIA docs).
- Context: 524,288 tokens.
- Tool calls: XML and JSON formats (Startup Fortune).
Read the full section
- Model: Nanbeige4.3-33B, dense, about 33.1B parameters, BF16 (NVIDIA docs).
- Context: 524,288 tokens. The chat template has
enable_thinkingandpreserve_thinkingflags, so reasoning traces can be turned on or off and optionally kept across turns (model card). - Tool calls: XML and JSON formats (Startup Fortune).
- Serving: the card includes examples for Transformers, SGLang and vLLM. Community quantizations are also available (model card).
Technical analysis for researchers and developers
- Architecture: a standard
LlamaForCausalLMwith 52 layers, hidden size 7,168, a SwiGLU MLP (intermediate size 22,016) and grouped-query attention with 56 query heads and 8 KV heads (NVIDIA docs). - Fine-tuning: NVIDIA's recipes run full SFT on two 8×H100 nodes with 16-way context parallelism, or rank-16 LoRA on one node.
- Evaluation: the reviewed sources do not document training data, training procedure, evaluation harnesses, scaffolds, sampling settings or thinking-mode settings.
Read the full section
- Architecture: a standard
LlamaForCausalLMwith 52 layers, hidden size 7,168, a SwiGLU MLP (intermediate size 22,016) and grouped-query attention with 56 query heads and 8 KV heads (NVIDIA docs). Because the architecture is unmodified, existing Llama tooling should work without changes. - Fine-tuning: NVIDIA's recipes run full SFT on two 8×H100 nodes with 16-way context parallelism, or rank-16 LoRA on one node. Both pack Tulu-3 data to 131K tokens. NVIDIA reports about 27 s per step for full SFT and about 36 s for LoRA (NVIDIA docs).
- Evaluation: the reviewed sources do not document training data, training procedure, evaluation harnesses, scaffolds, sampling settings or thinking-mode settings. Without those, the agentic scores cannot be reproduced.
Claims and evidence
Coverage also cites 23.8 vs 6.3 on Terminal-Bench 4.0 (AI TL;DR). All of these figures are self-reported. The press coverage repeats them, and no independent evaluation was found to corroborate them.
Read the full section
| Claim (all reported by Nanbeige) | Nanbeige4.3 | Comparison |
| DeepSWE | 65.5 | Qwen3.8-27B 42.5 |
| Terminal-Bench 2.1 | 80.9 | Qwen3.8-27B 73.0 |
| Automation Bench | 54.2 | Qwen3.8-27B 38.3 |
| JobBench | 44.8 | Qwen3.8-27B 34.5 |
| SWE-Bench Pro | 60.3 | Qwen3.8-27B 61.7 (Qwen ahead) |
| GPQA Diamond | 89.5 | Gemma4-31B 91.3 (Gemma ahead) |
Sources: Startup Fortune and the model card. Coverage also cites 23.8 vs 6.3 on Terminal-Bench 4.0 (AI TL;DR).
All of these figures are self-reported. The press coverage repeats them, and no independent evaluation was found to corroborate them.
Context and prior work
Earlier Nanbeige releases were small models. Nanbeige4.1-3B was built on Nanbeige4-3B-Base with SFT and RL, and claimed results above Qwen3-4B and sometimes above Qwen3-30B (arXiv 2602.13367). The comparison set of Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B places it in the crowded open-model tier around 30B (model card).
Read the full section
Earlier Nanbeige releases were small models. Nanbeige4.1-3B was built on Nanbeige4-3B-Base with SFT and RL, and claimed results above Qwen3-4B and sometimes above Qwen3-30B (arXiv 2602.13367). The 4.3 release inverts that pattern: the 33B model is positioned as a teacher, and a Nanbeige4.3-3B is planned as a distilled student (AI TL;DR). The comparison set of Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B places it in the crowded open-model tier around 30B (model card).
Limitations, safety and contested findings
- The model card has no limitations, safety or bias section (model card).
- The results are not a clean sweep. The model trails Qwen3.8-27B on SWE-Bench Pro and Gemma4-31B on GPQA Diamond.
- Without a published evaluation methodology, the size of the gaps (for example, 65.5 vs 42.5 on DeepSWE) may reflect differences in harness or scaffold rather than in the models themselves.
Read the full section
- The model card has no limitations, safety or bias section (model card).
- The results are not a clean sweep. The model trails Qwen3.8-27B on SWE-Bench Pro and Gemma4-31B on GPQA Diamond. Startup Fortune describes it as strong at agentic coding specifically and only competitive elsewhere (Startup Fortune).
- Without a published evaluation methodology, the size of the gaps (for example, 65.5 vs 42.5 on DeepSWE) may reflect differences in harness or scaffold rather than in the models themselves.
Business and practitioner implications
- License: Apache 2.0 allows commercial fine-tuning and self-hosting without per-token fees (model card).
- Cost: Startup Fortune reports that H100 rental prices rose about 40% between October 2025 and March 2026.
- Before adopting: teams should benchmark the model on their own repositories and agent scaffolds, and test it against Qwen3.8-27B, which leads on SWE-Bench Pro.
Read the full section
- License: Apache 2.0 allows commercial fine-tuning and self-hosting without per-token fees (model card).
- Cost: Startup Fortune reports that H100 rental prices rose about 40% between October 2025 and March 2026. It argues that a 33B model that fits on consumer hardware at 4-bit changes the build-vs-API calculation for coding agents (Startup Fortune). Running a 512K context still needs substantial memory for the KV cache.
- Before adopting: teams should benchmark the model on their own repositories and agent scaffolds, and test it against Qwen3.8-27B, which leads on SWE-Bench Pro.
- Distillation: teams that need small on-device agents may get more value from the planned 3B student than from the 33B teacher.