Nanbeige LLM Lab's Nanbeige4.3-33B Beats Qwen on Coding Agents With 65.5 DeepSWE Score

Nanbeige4.3-33B is a new open-weight reasoning model from Nanbeige LLM Lab, the foundation-model group inside Chinese recruitment platform BOSS Zhipin. The dense 33-billion-parameter model uses the Apache 2.0 license and focuses on coding, terminal control, tool use, planning, and other agentic workloads.

The lab describes the release as a teacher model for a planned Nanbeige4.3-3B. In model distillation, a smaller student learns from outputs generated by a more capable teacher, potentially carrying some of its reasoning and tool-use behavior onto cheaper hardware.

Inside the 33B release

Dense models activate all parameters for each token. Mixture-of-experts models route each token through only part of a much larger parameter pool, which can reduce computation while retaining a large memory footprint. That distinction matters when comparing model size, inference cost, and hardware requirements.

Agent tests lead the case

Artificial Analysis gives Nanbeige an Intelligence Index score of 20, compared with a median of 13 for open-weight models in its size range. The score ties Qwen3.8-27B and creates what the evaluator identifies as a Pareto-optimal point among dense models: no dense model in the comparison achieves a higher score with the same or fewer total parameters.

DeepSeek V4.1 Flash provides a useful architectural comparison. It scores 33 on the same index and contains 552 billion total parameters, although its mixture-of-experts design activates 16 billion for each token. Nanbeige activates its full 33 billion parameters, while requiring far less storage for the complete weight set.

The Nanbeige model card reports results against Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B at their highest thinking settings. The agent-oriented suites cover command-line work, software engineering, multi-step automation, and planning; AIME measures mathematical problem solving.

These agent results come from the model developer and should be treated as first-party measurements. Differences in prompts, tool environments, sampling settings, inference frameworks, and output limits can materially change agent benchmarks, so teams need evaluations that reproduce their own repositories and tool chains.

Fewer hallucinations, more abstentions

Nanbeige records a net score of -4 on the AA-Omniscience knowledge test, compared with -6 for Qwen3.8-27B. Its hallucination rate is 18%, below Qwen’s 29%, while its correct-answer rate is also lower at 12% versus 17%. The model reaches its result by declining more questions, indicating conservative calibration rather than broader factual coverage.

Frequent abstention can reduce unsupported actions inside an agent loop, although it can also interrupt workflows that require a best-effort answer. Application-level tests should therefore measure refusals, factual accuracy, and task completion together.

Nanbeige also generates substantially more text during evaluation. Artificial Analysis reports roughly 158,000 output tokens per Intelligence Index task, about 44% more than Qwen3.8-27B at the same final score. That difference affects latency, throughput, and inference cost even on self-hosted hardware. Nanbeige trails Qwen on SWE-Bench Pro, FrontierCode 2.0, and HLE-Diamond.

Serving needs memory headroom

The raw BF16 weights require about 66GB before accounting for the key-value cache, framework overhead, batching, or long prompts. Full-precision serving therefore usually requires multiple high-memory GPUs. Quantized builds can fit on a single 48GB card, with available context length and concurrency determined by quantization level and runtime memory use.

The published SGLang example distributes inference across four GPUs:

python -m sglang.launch_server \
  --model-path Nanbeige/Nanbeige4.3-33B \
  --tp-size 4 \
  --mem-fraction-static 0.8 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder

The parser flags tell SGLang how to interpret reasoning output and tool calls. They do not change the underlying model architecture. At publication, the Hugging Face model page lists about 99 community quantizations for runtimes including llama.cpp, Ollama, and LM Studio.

The chat template exposes three controls that affect agent behavior:

• enable_thinking enables model-generated reasoning tokens.
• preserve_thinking retains reasoning from earlier turns. The lab recommends True for multi-turn tool use and coding agents, although retained reasoning consumes context.
• tools supplies callable tool definitions. The lab recommends XML formatting for its strongest reported performance.

The developer’s recommended sampling settings vary by workload:

The advertised 524,288-token context window describes the model’s supported limit, not the capacity of every deployment. Long prompts, preserved reasoning, large output budgets, and concurrent requests expand the key-value cache and can raise memory requirements far beyond the weight size.

The practical fit is self-hosting

Nanbeige offers the clearest advantages to teams running coding or automation agents on controlled infrastructure. Self-hosting keeps repository contents and tool outputs within the operator’s environment, removes per-token API billing, and allows direct control over quantization, batching, prompts, and serving software. The tradeoffs include substantial hardware requirements, high token generation, and uneven results across coding benchmarks.

A deployment evaluation should measure the model on representative repositories, tool schemas, terminal environments, and failure-recovery loops. Useful operating metrics include task completion, unsupported tool calls, refusal rate, output tokens, end-to-end latency, peak memory, and performance after quantization.

The planned 3B student remains a roadmap item. Successful distillation could bring similar agent behaviors to laptop-class hardware, but its capability, release date, context length, and license will depend on the eventual release.