🌏 中文版
Model Info
| Item | Value |
|---|---|
| Model ID | Nanbeige/Nanbeige4.3-33B |
| Vendor | Nanbeige (AI team under Kanzhun Limited, parent of job-search app BOSS直聘) |
| Parameters | 33.1B (dense, no MoE) |
| Context Window | 524,288 tokens (512K) |
| Input Pricing (USD/1M tokens) | Not available (open weights only, no official hosted API) |
| Output Pricing (USD/1M tokens) | Not available (same as above) |
| Open Source | Yes (Apache 2.0) |
| Release Date | 2026-10-24 |
| Official Announcement | Nanbeige/Nanbeige4.3-33B (Hugging Face model card) |
| Hugging Face | Nanbeige/Nanbeige4.3-33B |
| Family | Nanbeige 4.x (previously 3B-class models like Nanbeige4.2-3B; 4.3-33B is the series' first 30B-class flagship) |
Key Capabilities
- Large lead on agentic coding benchmarks: Terminal-Bench 2.1 hits 80.9 (vs. Qwen3.8-27B 73.0, Muse-Glimmer-30B 61.8, Gemma4-31B 47.2), and DeepSWE hits 65.5 (vs. Qwen3.8-27B's 42.5)
- Ahead across 35 occupations on cowork tasks: on the internal JobBench suite spanning 35 occupations, it beats the baseline in most of them, scoring 44.8 overall vs. Qwen3.8-27B's 34.5
- Native 512K context: 524,288 tokens; Nanbeige recommends
preserve_thinking=truefor agentic and tool-use scenarios to retain multi-turn reasoning traces - Explicitly positioned as a "teacher model": Nanbeige states the Nanbeige4.3 strategy is to first train a high-capacity teacher, then distill it into smaller models (such as the upcoming Nanbeige4.3-3B) — this 33B release is not meant to be an end product on its own
Benchmark Results
| Benchmark | Score | Previous Gen | Strongest ~30B-class Competitor |
|---|---|---|---|
| Terminal-Bench 2.1 (agentic coding) | 80.9 | —(first 30B-class flagship) | Qwen3.8-27B 73.0 |
| DeepSWE | 65.5 | —(first 30B-class flagship) | Qwen3.8-27B 42.5 |
| SWE-Bench Pro | 60.3 | —(first 30B-class flagship) | Qwen3.8-27B 61.7 (Nanbeige trails) |
| JobBench (cowork, 35 occupations) | 44.8 | —(first 30B-class flagship) | Qwen3.8-27B 34.5 |
| GPQA Diamond (reasoning) | 89.5 | —(first 30B-class flagship) | Gemma4-31B 91.3 (Nanbeige trails) |
Source: Nanbeige's official self-reported scores on its Hugging Face model card; see the card's footnotes for judge models and tooling used per benchmark. Not yet independently reproduced. Y-axis 0–100, baseline at y=300, plot top at y=60, 2.4 units per point (y = 300 − score × 2.4).
⚠️ All scores above are Nanbeige's own self-reported results. The model card's footnotes note that some benchmarks used GPT-6 Sol or Claude Code as judge/harness; none have been independently reproduced yet.
Comparison with Prior Gen / Competitors
Nanbeige4.3-33B is the series' first 30B-class model — Nanbeige's earlier public releases topped out at 3B (like Nanbeige4.2-3B), so there's no real "previous generation" to compare against. The only meaningful comparison is with similarly-sized competitors: Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B.
Against Qwen3.8-27B, Nanbeige4.3-33B wins clearly on agent-leaning tasks: Terminal-Bench 2.1 (80.9 vs. 73.0), DeepSWE (65.5 vs. 42.5), and JobBench (44.8 vs. 34.5) are all large leads, suggesting it's tuned more deeply for long-horizon coding and multi-step, multi-tool cowork tasks. But it trails slightly on SWE-Bench Pro (60.3 vs. 61.7) and GDPval (84.8 vs. 86.4), and isn't the strongest on GPQA Diamond either (89.5 vs. Gemma4-31B's 91.3) — suggesting it traded some general-knowledge breadth for depth in agentic coding.
There's no mystery on pricing: it's open weights only, with no official hosted API — the opposite order from most closed-source vendors, who typically price a model first and open-source older versions later. Nanbeige frames this release as a "teacher model" rather than an end product; the real commercial intent is to use it to distill smaller models like Nanbeige4.3-3B down the line. The 33B model itself is more of a stepping stone than a bid to take market share from Qwen3.8-27B.
What This Means for Agent Development
The biggest signal from Nanbeige4.3-33B isn't the scores themselves — it's the explicitly stated "train a big teacher first, then distill smaller students" size-ladder strategy. If you're tracking the capability ceiling for small (3B-class) agent models, what's actually worth watching next isn't the 33B model itself, but the Nanbeige4.3-3B distilled from it: a stronger teacher raises the plausible ceiling for the distilled student.
- If you're building a self-hosted, single-GPU-friendly coding agent: a 33.1B dense model deploys more easily than 70B+ models, and the large leads on Terminal-Bench 2.1 and DeepSWE suggest it may be more reliable than Qwen3.8-27B for multi-step debugging and terminal operations — worth benchmarking directly against your own workload.
- If you're building a general assistant that needs broad world knowledge: GPQA Diamond and SWE-Bench Pro aren't its strongest suits, so Qwen3.8-27B or Gemma4-31B may be steadier choices there.
- If you're waiting on a small (3B-class) distilled version: watch this 33B teacher's scores now, then evaluate how much capability the eventual Nanbeige4.3-3B inherits — distillation quality is capped by the teacher, and this teacher's scores are unusually well-documented as a reference point.
- Not a fit: production environments that need an official SLA or hosted API — right now there's only open weights, no official hosted service at all.
Today's Takeaway
I used to assume small-model distillation always meant "borrow the best existing large model and extract what you need." Nanbeige4.3 flips that: train a brand-new, purpose-built high-capacity teacher first, then distill the later small-model series from it. That means the ceiling for a distillation chain isn't capped by "whichever large model you can borrow" — it's capped by whether a lab is willing and able to invest in building that teacher layer themselves. The 33B flagship's whole reason for existing is to let the later 3B model distill to a higher ceiling.