Nanbeige4.2

Nanbeige LLM Lab, Boss Zhipin · analysed 2026-08-10 · skill v0.4.0

dense text-only open-weights medium frontier-generalist

Design overview

Nanbeige4.2-3B is a 3B non-embedding-parameter (4B total) dense model whose one structural idea is depth on a budget: a Looped Transformer that passes hidden states through the same Transformer stack twice, roughly doubling effective computational depth without adding parameters. The report is explicit that the loop is trained from scratch (upcycling a standard transformer into a loop tested worse), that two passes is the sweet spot (retaining about 75% of a standard transformer's token efficiency for a significant capacity gain), and that KV-cache sharing across passes was rejected: it halves inference KV memory but consistently underperforms the full loop, so the shipped model keeps separate caches per pass. Beyond that single architectural decision, the report discloses almost no block details — no layer count, hidden width, heads, vocabulary, normalization, activation, or positional encoding — and concentrates instead on the post-training recipe that makes a 3B model agentic: a three-stage SFT curriculum that grows context 64K → 128K → 256K while shifting the target-token mixture from 82.7% STEM reasoning to 68.9% agentic data; turn-level loss masking so the model learns recovery from failed turns; two-stage RLHF over Think and Non-Think responses with a point-wise reward model; reasoning RL with problem-dependent length budgets and a difficulty-aware penalty; and agentic RL whose action-centric rubrics feed turn-level process rewards alongside outcome rewards, cutting the action error rate by about 20%. The result is positioned as a compact general agent, deployable locally (OpenClaw), that beats Qwen3.5-9B and Gemma4-12B on the reported agentic suites.

What the report claims is novel

Atlas assessment: The block-level differentiator is the two-pass Looped Transformer: no other atlas record reuses its own layer stack for a second pass to gain effective depth without parameters (the report cites the Loopy line of work, not in the atlas). Everything else — SFT curriculum, two-stage RLHF on Think/Non-Think, length-controlled reasoning RL, rubric-based agentic RL — is recipe work similar in spirit to the mini-activation agentic post-training of minimax-m2 but at 3B scale and with the length-control twist. The report is honest about its scope: it discloses almost no architecture hyperparameters, so the loop's exact depth/width and attention configuration cannot be compared against atlas peers. The claimed 'first at this scale' positioning (code + office + tool-use agents with retained reasoning) is plausible but not independently verifiable from the report.

Relation to the atlas

Lineage in

Influence out

No descendants recorded yet.

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyNanbeige4.2[1]
OrganisationNanbeige LLM Lab, Boss Zhipin[2]
Release datenot disclosed
Report URLhttps://huggingface.co/Nanbeige/Nanbeige4.2-3B/raw/main/Nanbeige42_report.pdf[3]
Licensenot disclosed
Open weightsyes[4]

Scale

Total parameters3B[5]
Active parameters / token3B
Layersnot disclosed
Hidden dimensionnot disclosed
FFN inner dimensionnot disclosed
FFN rationot disclosed
Query headsnot disclosed
KV headsnot disclosed
Head dimensionnot disclosed
Vocabulary sizenot disclosed
Embedding tyingnot disclosed

Core block

Block typedense
Attention variantnot disclosed
Attention layer patternnot disclosed
Depth mixingsequential-residual
Activationnot disclosed
Expert countn/a[6]
Experts per tokenn/a[6]
Shared expertsn/a[6]
Routingn/a[6]
Load balancingn/a[6]
Expert granularityn/a[6]
Methodnot disclosed
RoPE base frequencynot disclosed
Partial RoPEnot disclosed
Typenot disclosed
Placementnot disclosed
QK-normnot disclosed
Attention sinksnot disclosed
Softcappingnot disclosed
Othernot disclosed

Context

Trained context length262,144[9]
Deployed context length262,144[10]
Extension methodthree-stage SFT curriculum extending the maximum training context 64K → 128K → 256K (training at the deployed length); no positional-encoding extension method described[11]

Tokenizer

Algorithmnot disclosed
Notesnot disclosed

Training

Training tokens28T[12]
Data compositioncorpus exceeding Nanbeige 4.1 in scale and quality; increased sampling weights for mathematics, code, and synthetic QA data (found particularly beneficial for compact models); a small proportion of agentic trajectory data mixed into pre-training ('an initial step toward agentic pre-training')[13]
Curriculum / stagingsingle-stage from-scratch pre-training on 28T tokens (no staging disclosed); post-pre-training, the SFT recipe runs a three-stage curriculum (64K → 128K → 256K context) that shifts the supervised target-token mixture from STEM-centered reasoning (82.7% at 64K) through a transition stage (47.8% STEM / 29.5% agentic at 128K) to agentic-dominated (68.9% agentic / 22.4% STEM / 8.7% general at 256K)[14]
Optimizernot disclosed
LR schedulenot disclosed
Batch schedulenot disclosed
Precisionnot disclosed
Parallelismnot disclosed
Hardwarenot disclosed
Disclosed computenot disclosed

Post-training

SFTyes[15]
Preference optimisationRLHF[16]
Reasoning trainingThink/Non-Think hybrid modes with two-stage RLHF; reasoning RL with a problem-dependent length-control objective (offline median-length budgets per problem, difficulty-aware penalty alternating constrained and free-expansion phases); agentic RL with action-centric rubrics providing turn-level process rewards plus outcome rewards, restricted to easy tasks (short trajectories, high pass@8) at this scale[17]
Distillationnot disclosed

Modality

Typetext-only
Attachmentn/a[18]

Inference efficiency

KV-cache designfull (non-sharing) two-pass loop: each pass through the shared layer stack keeps its own KV cache, so inference KV memory is roughly doubled versus a single-pass stack; the report explicitly rejected KV-cache sharing across loop passes because it halves the cache but reduces performance
Quantisation shippednone disclosed
Speculative decodingnone disclosed
Serving optimisationsnone disclosed

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
GDPval Rubrics74.3Table 3
AgentIF-Oneday67.5Table 3
OfficeQA-Pro21.1Table 3
PinchBench-V274.7Table 3
ClawGym65.0Table 3
Claw-Eval Pass^352.2Table 3
MCP-Atlas57.8Table 3
SWE-Bench Verified63.6Table 3
SWE-Bench Pro46.9Table 3
Terminal-Bench 2.044.1Table 3
HLE (w/o Search)17.8Table 3
SciCode35.6Table 3
GPQA Diamond87.4Table 3
HMMT-Feb-202682.8Table 3
IMO-Answer-Bench67.3Table 3
LiveCodeBench-V672.5Table 3
AA-LCR58.7Table 3
IF-Bench54.6Table 3
Recruit-Bench63.3Table 3

Provenance references

  1. stated · Title ('Nanbeige4.2-3B: Unlocking Agentic Capabilities in a Compact Model')
  2. stated · Title page byline
  3. stated · pipeline
  4. stated · Abstract ('The model checkpoint is available at https://huggingface.co/Nanbeige/Nanbeige4.2-3B'); §1 ('We open-source Nanbeige4.2-3B')
  5. stated · Abstract ('a compact general agentic model with 3B non-embedding parameters'); Table 1 and Table 3 list total 4B / non-embedding 3B
  6. n/a · §2.1
  7. n/a · §3.2.1
  8. stated · §3.2.1 ('progressively extends the maximum training context from 64K to 128K and 256K tokens'); Appendix B.1 ('Context Window: 256k tokens')
  9. stated · §3.2.1 ('a three-stage curriculum that progressively extends the maximum training context from 64K to 128K and 256K tokens')
  10. stated · Appendix B.1 ('Context Window: 256k tokens')
  11. stated · §3.2.1
  12. stated · §2.2 ('Our pre-training corpus comprises 28T tokens')
  13. stated · §2.2
  14. stated · §2.1, §3.2.1
  15. stated · §3.2.1 ('Starting from the pretrained checkpoint, we conduct supervised fine-tuning with a three-stage curriculum'); turn-level loss masking via execution feedback
  16. stated · §3.2.2 ('we first apply a two-stage RLHF procedure covering both think and non-think responses'; point-wise reward model over response quality)
  17. stated · §3.2.2, §3.2.3, §3.2.4
  18. n/a · §1