Nanbeige4.3-33B
Released Oct 25, 2026 · knowledge cutoff unpublished
- Status
- Superseded
- Location
- China
- Type
- LLM
- Modality
- Text
- Context window
- 524K
- Max output
- —
- Speed
- —
- Price ($/MTok in / out)
- — / —
- Cost per task
- —
- Open weights
- Yes
News1 story
Strengths & weaknesses
Strengths
- Built to sit inside agent harnesses — XML/JSON tool calling and an option to preserve its thinking across turns
- Runs on a single 24 GB consumer GPU or a Mac at 4-bit, around 19 GB, so local agent work doesn't need a server
- Apache 2.0 with the base checkpoint published alongside, so it can be fine-tuned and shipped commercially without licence friction
Weaknesses
- Thinks at length by default — quantisers advise an output budget of 32K tokens or more, thinking included, or answers get cut off
- Every headline score comes from Nanbeige's own harness runs, and no hosted API exists yet, so trying it means self-hosting a 66 GB BF16 checkpoint
Released 2026-10-25 (Nanbeige's X announcement; some outlets date the Hugging Face upload to 10-24 or report 10-26 in UTC+8) by BOSS Zhipin's Nanbeige LLM Lab as a ~33.1B dense Llama-architecture model with a 524,288-token context, Apache 2.0 weights and a separate Base checkpoint. All scores are self-reported from the model card with preserve_thinking on: Terminal-Bench 2.1 80.9% (Terminus 2 harness), GPQA Diamond 89.5, SWE-bench Pro 60.3, AIME 2026 95.2; it also reports DeepSWE 65.5 (mini-swe-agent) and a GDPval figure of 84.8 that is not Artificial Analysis' Elo, so gdpvalAA is null. No SWE-bench Verified or HLE row was seen. The card recommends max new tokens of 65,536 for agentic work and 262,144 for reasoning, which are settings rather than a documented cap, so maxOutput is null. No hosted inference provider was listed at publication, so availability is self-host.
Benchmarks4 of 10 reported
GPQA Diamond
SWE-bench Pro
Terminal-Bench 2.1
AIME (latest)retired
MMLU-Proretired
SWE-bench Verified
Humanity's Last Exam
LMArena Elo
GDPval-AA v2
ARC-AGI-2
For developers
API model strings
Not researched
Licence
Apache License 2.0 · Permissive · commercial use permitted
Retirement
No retirement announced
Lineage
No recorded predecessor