Nanbeige4.3-33B: Apache 2.0 open agent model, 512K context
Nanbeige4.3-33B is an Apache 2.0 open agent model with 512K context and big Terminal-Bench claims. The numbers are self-reported. Run your own evals first.
Nanbeige published Nanbeige4.3-33B on Hugging Face: an open agent model under Apache 2.0, with a 512K-token context window and benchmark claims that put it ahead of every other open model near its size. A 33B model you can self-host and use commercially, with no license fine print, is worth a look. But every number on the card is Nanbeige's own, and nobody outside has reproduced them yet.
What actually happened
From the Hugging Face model card:
- Size and license: 33B parameters, BF16 weights, Apache 2.0. It is a fine-tune of Nanbeige4.3-33B-Base.
- Context: up to 524,288 tokens.
- Claimed scores against Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B: 80.9 on Terminal-Bench 2.1 (Qwen: 73.0), 65.5 on DeepSWE (Qwen: 42.5) and 70.4 on Toolathlon-Verified (Qwen: 68.5).
- Where it loses: Qwen3.8-27B is ahead on SWE-Bench Pro (61.7 vs. 60.3) and GDPval (86.4 vs. 84.8). Gemma4-31B is ahead on GPQA Diamond.
- Tool calling: the chat template supports tools in XML or JSON format, and the card gives vLLM serving flags for automatic tool choice.
- Serving: the card's examples use 4-way tensor parallelism. Community quantizations already list llama.cpp, Ollama and LM Studio support.
The card also says evaluation harnesses and settings differ by benchmark. Read that line twice.
Why it matters for your business
Apache 2.0 is the real feature. No user caps, no "acceptable use" addendum, no surprise license change on the next release. If you build an internal agent on it, you own that deployment.
Do the hardware math first. At BF16, 33B parameters is about 66GB of weights before any context. A 512K window adds a lot of KV cache on top. Full precision means multiple GPUs. A quantized build can fit on one workstation card, but test the quality loss on your own tasks.
Self-reported benchmarks are a shortlist, not a decision. Terminal-Bench and DeepSWE test coding agents in sandboxes. Your agent books jobs, reads invoices or updates a CRM. Build 50 to 100 test cases from your real workflow, run Nanbeige next to the model you already use, and compare the cost per correct task.
Key takeaways
- Nanbeige4.3-33B is a 33B open model under Apache 2.0 with up to 524,288 tokens of context
- It claims 80.9 on Terminal-Bench 2.1, but all scores are Nanbeige's own and not yet reproduced
- Qwen3.8-27B still leads on SWE-Bench Pro and GDPval in Nanbeige's own table
- BF16 weights are about 66GB; plan multiple GPUs or test a quantized build
- Run your own task-based evals before you switch any agent to it
Thinking about self-hosting an open model? We build eval sets from your real workflows, test open models next to your current API, and deploy the one that wins on your tasks. Compare self-hosted vs. API costs, or tell us which agent you want to bring in-house.
Sources: Hugging Face: Nanbeige4.3-33B model card.
- #open weights
- #nanbeige
- #ai agents
- #local llm
- #model evals
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
NVIDIA Nemotron 3.5 Bolt E4B: a 4GB local model to post-train
NVIDIA's Nemotron 3.5 Bolt E4B runs in about 4GB of VRAM with tool calling and 64K context. NVIDIA says post-train it before production. Here's why.
Read itArc browser shuts down December 22: plan your team's move
The Browser Company ends Arc on December 22. Arc Search loses AI and sync November 23, and data is deleted in 30 days. How to move a team off a dead browser.
Read it