Nanbeige4.1-3B: Versatile 3B LLM for Reasoning & Coding

• Nanbeige4.1-3B is a 3B-parameter open-source LLM unifying general reasoning, code generation, preference-aligned dialogue, and deep-search behavior in one dense model.
• It employs an advanced post-training recipe with extended long-context SFT and sequential point-wise and pair-wise reinforcement learning to refine both reasoning and coding capabilities.
• The model shows significant benchmark improvements in coding efficiency, multi-turn tool use, and agentic search, demonstrating a unified approach to diverse, real-world tasks.

Nanbeige4.1-3B is a 3B-parameter open-source generalist LLM from the Nanbeige LLM Lab (Boss Zhipin), presented as a compact successor within the Nanbeige 3B line and optimized to unify general reasoning, preference-aligned dialogue, code generation, tool use, and deep-search behavior in a single dense model. The designation “reasons, aligns, and acts” summarizes its intended scope: stronger math and science problem solving, better human preference fit and conversational quality, and sustained multi-step interaction with tools and retrieval systems. The released checkpoint is hosted at https://huggingface.co/Nanbeige/Nanbeige4.1-3B (Yang et al., 13 Feb 2026).

1. Lineage and conceptual scope

Nanbeige4.1-3B is introduced as a unified generalist successor to prior Nanbeige 3B models. It is initialized from Nanbeige4-3B-Base and post-trained with a broader recipe than Nanbeige4-3B-2511, with explicit emphasis on three capability clusters: general reasoning, code generation, and long-horizon agentic / deep-search behavior. Its intended use cases include general reasoning, mathematical reasoning, scientific reasoning, coding and competitive programming, instruction following, dialogue, preference-aligned responses, tool use and function calling, deep search and browsing, research-agent-style tasks, and multi-turn long-horizon interaction (Yang et al., 13 Feb 2026).

This positioning is best understood against the earlier Nanbeige4-3B family, which the preceding technical report describes as a family of small-scale autoregressive LLMs with explicit Base and Thinking variants. That family was framed around 23T high-quality pretraining tokens, over 30 million diverse instructions, 64K context, Fine-Grained Warmup-Stable-Decay (FG-WSD) scheduling, Solution Refinement, CoT Reconstruction, Dual-Level Preference Distillation, and multi-stage reinforcement learning (Yang et al., 6 Dec 2025). Nanbeige4.1-3B inherits the base-model substrate from that line, but its defining identity is not merely as another reasoning-specialized checkpoint; it is presented as a single 3B model intended to combine reasoning, alignment, coding, and agentic search within one deployment target (Yang et al., 13 Feb 2026).

The authors further characterize Nanbeige4.1-3B, to the best of their knowledge, as the first open-source small LLM to achieve this level of versatility in a single model. That formulation is best treated as a positioning claim rather than a formally provable property, although the benchmark suite does support unusually broad capability concentration at 3B scale (Yang et al., 13 Feb 2026).

2. Model specification and the boundary of disclosure

The explicit model specification is sparse. The report gives a parameter count of 3B, states that the system is a single dense 3B model, identifies Nanbeige4-3B-Base as the initialization point, and states that the context length after post-training reaches 256K. It also makes clear that the model is open-source (Yang et al., 13 Feb 2026).

By contrast, most conventional architectural descriptors are omitted. The report does not specify the backbone family, number of layers, hidden size, attention head count, tokenizer or vocabulary, positional encoding type, grouped-query or multi-query attention, RoPE or ALiBi usage, MLP variant, activation function, normalization type, KV-cache design, quantization guidance, serving throughput, or memory footprint. It also provides no full architectural equations for the transformer itself (Yang et al., 13 Feb 2026).

This omission matters because Nanbeige4.1-3B is presented as a strong empirical system rather than as an architecture paper. A plausible implication is that the paper’s main scientific contribution lies in post-training design, reward construction, and agentic data synthesis, not in a newly disclosed backbone. For readers interested in the inherited training philosophy beneath the base checkpoint, the earlier Nanbeige4 report is materially more informative, especially on the prior 23T-token pretraining recipe and the preceding 64K long-context setup (Yang et al., 6 Dec 2025).

3. Supervised post-training and the 256K curriculum

The post-training recipe begins from Nanbeige4-3B-Base and first applies an expanded supervised fine-tuning stage. Relative to the earlier Nanbeige4-3B-2511 line, the changes are explicit: the SFT mixture is rebalanced toward more code, more challenging math, and more hard general-domain problems; the context curriculum is extended from 32K → 64K to 32K → 64K → 256K; and the synthetic reasoning pipeline is strengthened through scaled-up Solution Refinement iterations and a stronger CoT Reconstruction model for cleaner reasoning traces (Yang et al., 13 Feb 2026).

The final 256K-stage SFT mixture is deliberately multi-domain: 27% Code, 26% Deep-Search, 23% STEM, 13% Tool-use, and 10% General. This distribution is important because it shows that long-context post-training is not reserved for only one capability family; the model is shaped as a unified generalist even at the longest context stage (Yang et al., 13 Feb 2026).

The terms Solution Refinement and CoT Reconstruction are inherited from the Nanbeige4 framework. In the earlier report, Solution Refinement is described as a checklist-conditioned process in which a teacher-selection mechanism chooses the best teacher for a task, candidate answers are co-generated, and an evaluator performs cross-evaluation against criteria such as correctness, completeness, consistency, executability, and safety, followed by iterative generate-review-revise cycles. CoT Reconstruction then uses a chain-completion model to generate a brief summary chain of thought and a fuller explicit chain of thought consistent with the refined final solution (Yang et al., 6 Dec 2025). Nanbeige4.1-3B does not redefine these ideas; rather, it strengthens their use in synthetic reasoning data generation (Yang et al., 13 Feb 2026).

The SFT ablation confirms that these adjustments are consequential. Against the previous Nanbeige SFT baseline, LCB V6 rises from 45.5 to 62.0, LCB Pro Medium from 1.8 to 22.8, HMMT Nov from 60.7 to 74.3, IMO-Answer-Bench from 34.8 to 48.9, Arena-Hard V2 from 45.5 to 60.2, and Multi-Challenge from 42.6 to 44.4 (Yang et al., 13 Feb 2026).

4. Reward modeling, alignment, and coding optimization

After SFT, the training pipeline moves into general reinforcement learning because the authors observe repetition, redundant thinking, and formatting issues. The first RL stage uses GRPO, with 8 rollouts per prompt, and a point-wise general reward model trained on curated large-scale human preference data. Because the point-wise reward model scores each sampled response independently, it is intended to suppress repetitive outputs, low-readability outputs, and over-redundant reasoning (Yang et al., 13 Feb 2026).

A second general RL stage then introduces pair-wise reward modeling. The pair-wise reward model is trained on paired comparison data from code generation and LMArena-style conversations, both single-turn and multi-turn. Response pairs are formed from a strong model and a weak model, then filtered with the same checklist strategy used in Nanbeige4; for multi-turn data, the full dialogue history is concatenated into the reward-model input. The report also mentions a swap-consistency regularizer to mitigate position bias, defined as an MSE-style consistency constraint under response-order swapping, though the exact formula is not provided (Yang et al., 13 Feb 2026).

The pairing of these RL stages is sequential rather than collapsed into one objective: SFT → point-wise RL → pair-wise RL. Their empirical roles are differentiated. On Arena-Hard V2, scores increase from 60.2 at SFT to 66.6 after point-wise RL and 73.8 after pair-wise RL; on Multi-Challenge, the sequence is 44.4 → 47.7 → 55.1. On LCB V6, the corresponding progression is 62.0 → 66.0 → 65.6, which suggests that pair-wise RL contributes most clearly to alignment-style benchmarks rather than to coding performance (Yang et al., 13 Feb 2026).

Coding receives its own dedicated RL pipeline. The system uses a unified judge system with two components: an execution-based sandbox for correctness and an instruct judge model for time-complexity comparison. In the first code-RL stage, reward is based on pass rate, and on-policy filtering keeps only problems of moderate difficulty: from 8 rollouts, a problem is retained when the number of successful solves satisfies $k \in [1,5]$. In the second stage, the reward becomes gated by correctness: if a solution is not fully correct, the objective optimizes formatting + correctness only; if it passes all tests, a time-complexity reward is added. Each sample includes the problem statement, test cases, a time-complexity-optimal solution, and an optimal complexity label (Yang et al., 13 Feb 2026).

This correctness-first, efficiency-second structure is meant to avoid a common coding-RL failure mode in which a model passes tests through brute force or suboptimal algorithms. The appendix examples support that framing qualitatively, showing improvements such as $O(N \log N)$ to $O(N)$, $O(N + M\log N)$ to $O(N)$, and $O(N^2 \log N)$ to $O(N \log N)$, although the report does not provide a final benchmark table isolating the marginal effect of the time-complexity reward alone (Yang et al., 13 Feb 2026).

5. Agentic behavior, deep search, and long-horizon tool use

In this report, deep search denotes retrieval-centric tasks involving complex multi-hop reasoning, extremely long-context settings, and iterative interaction with the environment to gather information. The model is trained for this setting with a synthetic data pipeline built around a Wikipedia entity-relation graph. Head entities are chosen from those updated in the past six months; the system performs conditional random walks over the graph; and the resulting relational chains, augmented with temporal context, are fed to a strong LLM to synthesize hard questions (Yang et al., 13 Feb 2026).

Those questions are then solved using multiple agent frameworks, which generate diverse reasoning paths and tool trajectories. The resulting traces are mapped into unified multi-turn tool-invocation sequences, after which turn-level judgment is applied. Each turn is evaluated for logical soundness, tool-call accuracy, and informational gain. Failed turns are either excluded from training or assigned negative reward (Yang et al., 13 Feb 2026).

The paper argues that this turn-level supervision improves long-horizon stability by sharpening credit assignment: bad intermediate steps can be penalized without waiting for final task completion, and good intermediate steps can be reinforced before the final answer is reached. There is no direct ablation isolating turn-level supervision alone, so that specific causal claim remains less directly demonstrated than the effect of synthetic search data itself (Yang et al., 13 Feb 2026).

The evaluation environment is also explicitly specified. Search and tool-use experiments are run in the Mindflow framework with Serper for search, Jina for webpage content extraction, and E2B Sandbox as the execution environment; HuggingFace is explicitly disabled during search evaluation. The report further claims that Nanbeige4.1-3B can reliably execute up to 600 tool-call turns for complex problem solving. That claim is important, but it is only partially substantiated: the paper does not provide a benchmark table by turn count, a success-versus-depth curve, or trace-level statistics for very long trajectories (Yang et al., 13 Feb 2026).

6. Empirical profile, comparative standing, and caveats

The evaluation suite spans code, math, science, alignment, tool use, and deep-search benchmarks. Relative to its predecessor and to several Qwen baselines, Nanbeige4.1-3B shows broad gains, with especially large improvements in coding and agentic search (Yang et al., 13 Feb 2026).

Benchmark	Nanbeige4-3B-2511	Nanbeige4.1-3B
LCB-V6	46.0	76.9
AIME 2026 I	84.10	87.40
GPQA	82.2	83.8
Arena-Hard-V2	60.0	73.2
BFCL-V4	53.80	56.50
GAIA (text-only)	19.42	69.90
xBench-DeepSearch-10	11.00	39.00

The coding results are among the clearest. Nanbeige4.1-3B scores 76.9 on LCB-V6, 81.4 on LCB-Pro-Easy, and 28.1 on LCB-Pro-Medium, beating all listed Qwen baselines in that table, including much larger models. On a real-world stress test covering LeetCode Weekly Contests 484–488, it achieves an 85.0% pass rate (17/20), compared with 65.0% for Qwen3-30B-A3B-2507, 55.0% for Qwen3-4B-2507, and 50.0% for Qwen3-32B (Yang et al., 13 Feb 2026).

Its reasoning profile is strong but not uniformly dominant. In mathematics, it reaches 87.40 on AIME 2026 I and 77.92 on HMMT Nov, beating Qwen3-4B-2507, Qwen3-32B, and Qwen3-30B-A3B-2507 on those two benchmarks, while on IMO-Answer-Bench it scores 53.38, slightly below Qwen3-30B-A3B-2507 at 54.34 and below Qwen3-Next-80B-A3B at 58.00. In science, it posts 83.8 on GPQA and 12.60 on HLE (Text-only), improving over small-model baselines and the predecessor but still trailing the 80B model on HLE (Yang et al., 13 Feb 2026).

Alignment and tool use are also materially improved. On Arena-Hard-V2, Nanbeige4.1-3B reaches 73.2, above Qwen3-30B-A3B-2507 at 60.2, Qwen3-32B at 56.0, and Nanbeige4-3B-2511 at 60.0. On Multi-Challenge, it scores 52.21, above Qwen3-30B-A3B-2507 at 49.40 but below Qwen3-Next-80B-A3B at 56.52. For tool use, it attains 56.50 on BFCL-V4, the best score in the reported table, and 48.57 on Tau2-Bench, which is competitive but below Qwen3-Next-80B-A3B at 57.40 (Yang et al., 13 Feb 2026).

The most striking gains appear in deep-search and agentic evaluation. Against the reported baselines, Nanbeige4.1-3B reaches 69.90 on GAIA (text-only), 19.12 on BrowseComp, 31.83 on BrowseComp-ZH, 22.29 on HLE (text-only), 41.44 on SEAL-0, 75.00 on xBench-DeepSearch-05, and 39.00 on xBench-DeepSearch-10. The synthetic deep-search data ablation is especially informative: starting from Nanbeige4-3B-2511, adding Synthetic QA raises GAIA from 19.4 to 58.3, BrowseComp from 0.8 to 14.4, BrowseComp-ZH from 3.1 to 30.1, HLE (text-only) from 13.9 to 22.4, SEAL-0 from 12.6 to 36.0, xBench-DeepSearch-05 from 33.0 to 76.0, and xBench-DeepSearch-10 from 11.0 to 30.0. This strongly suggests that synthetic search-data construction is one of the principal drivers of the model’s agentic performance (Yang et al., 13 Feb 2026).

The paper’s caveats are substantial. It is much stronger on capability claims than on safety analysis. There is no dedicated section on harmful content refusal, jailbreak robustness, tool misuse prevention, hallucination under search, benchmark contamination, or privacy and security risks in agent settings. Architectural and compute transparency are limited: the report does not provide optimizer settings, batch sizes, hardware, total compute, or training duration. The 600-turn tool-call claim remains only partially evidenced. Finally, several search results depend on a specific tool stack—Mindflow, Serper, Jina, and E2B—and some evaluations, notably GAIA and HLE, are reported only on text-only subsets (Yang et al., 13 Feb 2026).

Taken together, Nanbeige4.1-3B is best understood as a 3B-scale demonstration that broad competence and specialization need not be separated into different small models. Its distinctive contribution is not a newly disclosed backbone, but a stacked post-training recipe that combines long-context SFT, sequential point-wise and pair-wise reward modeling, correctness-first and complexity-aware code RL, and synthetic deep-search trajectories with turn-level supervision. Within the evidence reported, that recipe yields an unusually broad empirical profile for a compact open model (Yang et al., 13 Feb 2026).