Nanbeige/Nanbeige4.3-33B

One-line: Nanbeige’s flagship 33B dense “teacher” agent — lab-reported 80.9 Terminal-Bench 2.1, 65.5 DeepSWE and 70.4 Toolathlon-Verified, Apache-2.0 with a 512K context, in a plain Llama architecture that loads in stock transformers/llama.cpp-style stacks.

⚠ All scores below are lab-reported (the model card’s HTML table); no independent reproduction exists yet — Artificial Analysis has no page, BenchLM 404s, and both third-party write-ups (explainx.ai, startupfortune) explicitly note every number is self-reported. The comparison columns are the lab’s own re-runs of Qwen3.8-27B / Muse-Glimmer-30B / Gemma4-31B under its own harness, so cross-column comparisons are indicative, not controlled.

Released 2026-10-24 (repo createdAt 2026-10-24T02:49Z; card lastModified 2026-10-25T11:11Z). Surfaced outside the strict 36h window (41.2h at run time) via the per-org sweep + unfiltered trendingScore and promoted under the soft-boundary rule: 297 likes in 41h, 20+ same-day community quant repos, Reddit >140 points, X + press coverage. It was missed by prior runs only because Stream A’s downloads >= 50 floor excludes brand-new large repos (HF shows 0 downloads at capture).

Coding benchmarks

Benchmark	Score	Source
HumanEval	—	not published
HumanEval+	—	not published
MBPP	—	not published
LiveCodeBench	—	not published
Aider polyglot	—	not published
SWE-bench Verified	—	not published (card uses Terminal-Bench 2.1/4.0 + DeepSWE + SWE-Bench Pro instead)
SWE-bench Multilingual	—	not published
SWE-Bench Pro	60.3	lab — card (vs Qwen3.8-27B 61.7, Muse-Glimmer-30B 36.9, Gemma4-31B 51.2)
Terminal-Bench 2.1	80.9	lab — card, Terminus 2 harness (vs Qwen3.8-27B 73.0)
Terminal-Bench 4.0	23.8	lab — card, mini-swe-agent v2.4.6 (vs Qwen 6.3)
DeepSWE	65.5	lab — card, mini-swe-agent v2.4.6 (vs Qwen 42.5)
SWE-Atlas-QnA	50.8	lab — card, mini-swe-agent v2.4.6; 1,304 rubric tasks judged by GPT-6 Sol
NL2Repo	53.0	lab — card, Claude Code harness (vs Qwen 48.3)
SciCode	52.6	lab — card (vs Qwen 46.7)
SkillsBench	61.1	lab — card (vs Qwen 56.4)

Agentic / tool-use benchmarks

Benchmark	Score	Source
BFCL	—	not published
τ-Bench	—	not published (Tau3-Bench instead, below)
ToolACE	—	not published
GAIA	—	not published
Tau3-Bench	48.8	lab — card, official repo (vs Qwen 48.5)
Toolathlon-Verified	70.4	lab — card, no max_tokens cap (vs Qwen 68.5)
Automation Bench	54.2	lab — card, v1.0.6, 600 public tasks (vs Qwen 38.3)
JobBench	44.8	lab — card, Claude Code harness + search tool (vs Qwen 34.5)
GDPval	84.8	lab — card, revised rubrics (vs Qwen 86.4)
Claw-Eval (pass³)	78.3	lab — card, Claw-Eval General subset, 157 single-turn tasks
DeepPlanning	63.3	lab — card, official repo
EnterpriseOps-Gym	41.0	lab — card, official repo
YC-bench	$1,025,437	lab — card, long-horizon business sim scored in dollars (vs Qwen $641,290)
Tool-use contract	XML (recommended) or JSON; enable_thinking / preserve_thinking chat-template switches; qwen3_coder tool parser, qwen3 reasoning parser	card (no BFCL/τ number published)

Reasoning (context only)

Benchmark	Score	Source
MMLU / MMLU-Pro	—	not published
GPQA Diamond	89.5	lab — card (trails Gemma4-31B’s 91.3)
MATH	—	not published
AIME 2026	95.2	lab — card (vs Qwen 94.5 — effectively a tie)
AA-LCR	84.0	lab — card (vs Qwen 82.0)

Efficiency

• Active params: 33.1B dense — all layers active (52 layers, hidden 7168, 56 query / 8 KV heads GQA, ffn 22016, vocab 166912). No MoE routing.
• Context length: 524,288 tokens (512K) native.
• VRAM (fp16 rough): ~66 GB (14 BF16 shards, 66.2 GB measured). Comfortable on 2×48 GB or 1×80 GB; the card’s own SGLang/vLLM examples use --tp-size 4. A 4-bit quant (~17 GB) fits a single 24 GB consumer card — community MLX/GGUF/oQ*/EXL3 quants appeared within hours of the drop.
• Runtime: plain Llama architecture (LlamaForCausalLM, standard model_type: llama) — unlike the custom-arch Nanbeige4.2-3B, this needs no repo model code; the card’s trust_remote_code=True is vestigial (no modeling_*.py in the repo). Recommended sampling: agentic/tool use temp 1.0 / top_p 0.95 / 65,536 new tokens; reasoning/chat temp 0.6 / top_p 0.95 / 262,144 new tokens. Evaluated with preserve_thinking=true.

What makes it notable

It is the strongest published agentic dense model in the ~30B class on the harnesses labs actually use for agents in late 2026: 80.9 on Terminal-Bench 2.1 (Terminus 2), 65.5 on DeepSWE and 70.4 on Toolathlon-Verified — margins over Qwen3.8-27B of +7.9 / +23.0 / +1.9 points — paired with a 512K context and an XML/JSON tool contract that is explicitly designed for multi-turn agent loops. It is candid where it loses (SWE-Bench Pro 60.3 vs Qwen 61.7; GDPval 84.8 vs 86.4; GPQA Diamond 89.5 vs Gemma4-31B 91.3), and its reasoning numbers are essentially ties rather than wins — the differentiation is specifically long-horizon agent work, not raw knowledge. Two caveats matter for local use: everything is lab-reported with zero third-party verification as of 2026-10-26, and the harnesses (Terminus 2, mini-swe-agent v2.4.6, Claude Code) do not transfer automatically to Claude Code/Codex-style scaffolds. Deployment friction is low — stock Llama arch, Apache-2.0, no gate — but 33B dense at BF16 is a 66 GB footprint, so local runs mean a quant. It is also positioned as the teacher of a planned Nanbeige4.3-3B distillation (not yet released), which is the variant that would matter most for on-device agents.

See also

• welcome
• Source: https://huggingface.co/Nanbeige/Nanbeige4.3-33B
• Base variant (skip — pretrain): https://huggingface.co/Nanbeige/Nanbeige4.3-33B-Base
• Prior family entry: Nanbeige—Nanbeige4.2-3B
• Coverage: https://explainx.ai/blog/nanbeige-4-3-33b-open-agentic-model-guide-2026 · https://startupfortune.com/nanbeiges-new-33b-open-model-takes-aim-at-qwen-and-gemma-on-coding/
• Community quants (created 2026-10-24/25, corroborate the drop date): MLX (pratikgx, kapitsaprime, fcmeyer), GGUF (Razor3D, anik-jha), oQ* (pyros-vault, a-gordo), EXL3 (lxe), W3A16 (stuchapin)