The new pattern in open-weight releases is to ship the teacher first. Nanbeige did not push out a small model to chase leaderboard spots; it released a 33B dense model as the flagship teacher of the 4.3 series, with a 3B student planned to be distilled from it later. That inverts the usual release order, where small models went out first to win on cost-effectiveness.

Here is what the teacher actually is. The model card describes Nanbeige4.3-33B as a dense model under the Apache 2.0 license, with Hugging Face metadata showing roughly 33.1 billion parameters, BF16 weights around 66GB, a context window of 524,288 tokens (512K), and support for English and Chinese. Both the instruct checkpoint and a Base checkpoint were released at the same time, and the Hugging Face page already carries 44 community quantizations and 5 third-party finetunes, with a Llama architecture tag. Nanbeige previously built small models such as 4.1-3B and 4.2-3B. This time the 33B is stated plainly as the flagship teacher, with the small-model distillation pipeline downstream of it.

The leaderboard: big wins on long-horizon work, three losses

The vendor's comparison set is Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, all run at their highest thinking effort with identical inference hyperparameters and harness per benchmark. The key rows:

  • Terminal-Bench 2.1 (Terminus 2 harness): 80.9, against 73.0 / 61.8 / 47.2
  • Terminal-Bench 4.0 (mini-swe-agent v2.4.6): 23.8, against 6.3 / 1.6 / 0.0
  • DeepSWE (mini-swe-agent v2.4.6): 65.5, against 42.5 / 4.4 / 2.7
  • SWE-Bench Pro: 60.3, below Qwen3.8-27B's 61.7
  • GDPval: 84.8, below 86.4
  • GPQA Diamond: 89.5, below Gemma4-31B's 91.3

The distribution of those gaps is more interesting than any single number. On long-horizon agent tasks the margin runs into the teens or twenties, with Terminal-Bench 4.0 landing at 23.8 versus 6.3. On pure reasoning the gap nearly vanishes: AIME 2026 is 95.2 against 94.5, and GPQA Diamond is 89.5 against 89.2. The vendor put all three losses in the table, which is more candid than most launch-day benchmark sheets.

The Agentic Cowork half may matter more

More interesting than the coding leaderboards is the half of the table labeled Agentic Cowork: Automation Bench 54.2 against 38.3, JobBench 44.8 against 34.5 (the model card says it beats the baseline in most of the 35 occupations evaluated), YC-bench priced in dollars at 1,025,437 against 641,290, Toolathlon-Verified 70.4 against 68.5, Claw-Eval 78.3 (counting only the 157 single-turn tasks in the General subset), Tau3-Bench 48.8 against 48.5, DeepPlanning 63.3 against 49.2, and EnterpriseOps-Gym 41.0 against 40.2.

That half tests whether a model can do multi-step work across business tools like an office assistant, rather than solving coding puzzles. It is also the more fragile half. These suites depend heavily on the harness, on which tools get exposed, and on who grades: SWE-Atlas-QnA alone is graded by a rubric judge, GPT-6 Sol, against 1,304 task-specific rubrics (50.8 against 41.9). And the rivals were picked by the lab itself. The card does not compare against larger open MoE models, and it does not compare against closed frontier models either.

The base checkpoint deserves a mention too: Nanbeige4.3-33B-Base beats Gemma4-31B-Base across all eight base-model benchmarks, with MMLU-Pro 66.9 against 59.3 and GPQA 62.0 against 41.5.

Running it: 66GB of weights and a 512K context

At BF16 the weights are about 66GB, which makes a single 80GB GPU tight once you add KV cache for a 512K context, and the official example uses four-way tensor parallelism. Both SGLang and vLLM get explicit commands (--tp-size 4, --reasoning-parser qwen3, --tool-call-parser qwen3_coder). Two chat-template switches matter: enable_thinking is on by default, and preserve_thinking is recommended for multi-turn tool use, office tasks and code-agent workflows but turned off for ordinary chat and question answering. The recommended tool-call format is xml, with json kept for compatibility. Inference settings split into two profiles: for agentic and tool-use tasks, temperature 1.0, top_p 0.95 and max_new_tokens 65,536; for reasoning and chat, temperature 0.6, top_p 0.95 and max_new_tokens 262144. The Transformers example also loads with trust_remote_code=True, so it is worth reading the repository code before running it on a machine holding secrets.

So what

Let us be clear about one thing first: every number in that table is the lab's own test run. There is no third-party reproduction at the time of writing, and the outlets covering it all label the results as vendor-reported. Harness choice matters enormously for agent leaderboards, and the numbers move when you change it.

Is it worth trying? My advice is not to treat a public leaderboard as the conclusion. Pick two or three recurring tasks from your own work (triaging a ticket queue, drafting a report from a spreadsheet), run them through the harness you actually use, and compare the cost per completed task rather than the score. A model that scores five points higher but needs twice the retries is not cheaper.

The deeper question is this: the success of the "build a teacher first, then compress a student" route does not really depend on how strong this 33B is. It depends on whether the not-yet-released Nanbeige4.3-3B can absorb what the teacher taught it. The 33B is the means; the 3B is the point.