Nanbeige uploaded Nanbeige4.3-33B weights to Hugging Face on October 24, 2026, followed by model-card updates on October 25. The release provides downloadable coding-agent weights for teams evaluating controlled deployment. Hardware capacity, serving compatibility, and the interpretation of vendor benchmarks remain important constraints.
October 24 brought weights, not just a preview
The primary repository history records an initial commit and folder upload on October 24, configuration and README changes on October 25, evaluation-result additions on October 27, and another README update on October 28. Those are repository events, not interchangeable announcement dates. We could not independently inspect the linked X launch post, so we do not assign it a separate announcement timestamp.
The current file listing contains fourteen Safetensors shards and totals approximately 66.3 GB. That establishes a downloadable release, not merely an announcement or gated preview. A separate Nanbeige4.3-33B-Base repository is also available; its card positions it as a foundation for custom post-training, rather than the same ready-to-evaluate agent checkpoint.
API availability is a different question. As checked November 1, the main Hugging Face page lists no Inference Provider deployment. Its local serving examples are not evidence of a Nanbeige-operated hosted API, service-level agreement or published API price. This check does not establish that no third-party endpoint exists anywhere.
The hardware story is larger than “33B”
The published configuration identifies a dense LlamaForCausalLM architecture with BF16 weights, 52 layers, eight key-value heads and a maximum position count of 524,288. Nanbeige supplies Transformers, SGLang and vLLM instructions; the two server recipes use four-way tensor parallelism. These are example configurations, not a verified minimum GPU count or a guarantee for a particular accelerator.
A useful capacity distinction follows from the configuration. Assuming an ordinary BF16 key-value cache, 52 layers × two cache tensors × eight KV heads × 128 dimensions × two bytes gives 212,992 bytes per cached token. At 524,288 tokens, that is about 104 GiB for one sequence's cache alone, before weights and runtime overhead. This is our arithmetic, not a measured deployment result; cache precision and serving implementation can change it.
Consequently, fitting the weight files is not equivalent to serving the advertised context window. Our recommendation is to size an evaluation around realistic repository context and concurrency first. Do not procure hardware on the assumption that “open 33B” means a full-context workstation deployment.
Coding results are promising, but remain vendor evidence
Nanbeige's published comparison reports 80.9 on Terminal-Bench 2.1, 23.8 on Terminal-Bench 4.0 and 65.5 on DeepSWE. MarkTechPost's October 25 coverage independently documents the release and reproduces these vendor results. It also notes a counterexample to blanket superiority: Nanbeige scores 60.3 on SWE-Bench Pro against 61.7 for Qwen3.8-27B in the lab's table.
These are different benchmarks, not successive measurements of one task. The model card specifies Terminus 2 for Terminal-Bench 2.1 and mini-swe-agent v2.4.6 for Terminal-Bench 4 and DeepSWE. MarkTechPost also identifies discrepancies between some competitor figures and their own cards. Neither coverage nor a repository score table substitutes for an independent rerun. Nerova has not run firsthand performance tests.
What the license and teacher-model plan change
Hugging Face labels both checkpoints Apache-2.0. The main file listing does not show a standalone LICENSE file; organizations should retain the relevant revision and verify the governing terms before deployment or redistribution. Downloadability should not be mistaken for a complete disclosure of training materials.
The Base card describes the 33B model as the teacher for future smaller agents, naming Nanbeige4.3-3B as an upcoming example. That is a roadmap statement, not proof of downloadable 3B weights.
For business adoption, our conclusion is narrow: evaluate the released 33B checkpoint on a bounded coding workflow. Keep repository writes reviewable, record completed-task latency and peak memory, and compare accepted patches rather than headline scores. The October release creates a concrete self-hosting option; it does not establish production reliability, inexpensive full-context inference or a managed API service.