Qwen3-Coder-Next

Qwen Team · 2026-02-28 · analysed 2026-08-10 · skill v0.4.0

hybrid hybrid text-only open-weights large domain-specific

Design overview

Qwen3-Coder-Next is an 80-billion-parameter mixture-of-experts model that activates 3 billion parameters per token, built on the Qwen3-Next base with a hybrid-attention backbone — and that is the report's entire architectural disclosure. There is no configuration table: layers, hidden width, head counts, expert counts, vocabulary, normalization, and activation are all inherited from the base model and left unspecified. What the report documents instead is a complete agentic training recipe. Mid-training expands the context from 32,768 to 262,144 tokens so the model can carry multi-turn tool-use trajectories, mixing natural GitHub data (370 languages, roughly 600B repository-level tokens) with synthetic PR-derived tasks (~800K verifiable instances across 9+ languages), text–code grounding data rewritten by the larger Qwen3-Coder-480B-A35B-Instruct teacher, and FIM data in chat and search-and-replace forms. After continued pretraining, SFT on verified agentic trajectories is followed by specialization into four domain experts — web development, user experience/tool-format adherence, single-turn QA, software engineering — and distillation back into a single unified model. Execution-verifiable RL then runs in two regimes: single-turn, unit-test-verified coding, and multi-turn agentic SWE interaction with trajectory-level rewards, token-level tool-format penalties, and a reinforced reward-hacking blocker. The result is a small-footprint coding agent that matches models with an order of magnitude more active compute on the SWE-Bench suites, arguing that scaling agentic training, not model size, drives coding-agent capability. Architecturally this record is the thinnest in the atlas; as a training-recipe disclosure it is among the most detailed.

What the report claims is novel

Atlas assessment: Every claimed contribution is in the training recipe and data infrastructure — none is architectural: the hybrid-attention MoE backbone is inherited from the Qwen3-Next base and the report discloses no config details at all. vs atlas peers, the mid-training → SFT → multi-expert → distillation → execution-RL pipeline extends the Qwen2.5-Coder/Qwen3-Coder lineage (repository-level pretraining with special concatenation tokens is explicitly carried over from Qwen2.5-Coder), and tool-template diversity plus the reward-hacking blocker are practical engineering contributions of the same family as the agentic-RL recipes seen in Kimi K3, though K3's report is far more architecture-heavy. The genuinely distinctive claims are (1) the demonstration that a 3B-active model can match models with ~10× active compute via scaled agentic training, and (2) the first security-benchmark comparison for a coding agent. Architecturally this is the thinnest record in the atlas: layers, experts, attention internals, and normalization are simply not discussed.

Relation to the atlas

Lineage in

Influence out

No descendants recorded yet.

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyQwen3-Coder-Next[1]
OrganisationQwen Team[2]
Release date2026-02-28[3]
Report URLhttps://arxiv.org/abs/2603.00729[4]
Licensenot disclosed
Open weightsyes[5]

Variants

Scale

Total parameters80B[6]
Active parameters / token3B[7]
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 typehybrid
Attention varianthybrid[8]
Attention layer patternnot disclosed
Depth mixingsequential-residual
Activationnot disclosed
Expert countnot disclosed
Experts per tokennot disclosed
Shared expertsnot disclosed
Routingnot disclosed
Load balancingnot disclosed
Expert granularitynot disclosed
Methodnot disclosed
RoPE base frequencynot disclosed
Partial RoPEnot disclosed
Typenot disclosed
Placementnot disclosed
QK-normno
Attention sinksno
Softcappingno

Context

Trained context length262,144[9]
Deployed context lengthnot disclosed
Extension methodnot disclosed

Tokenizer

Algorithmnot disclosed
Notesnot disclosed

Training

Training tokensnot disclosed
Data compositionmid-training corpus primarily natural data with a smaller synthetic portion: (1) GitHub source code with language support expanded from 92 to 370 languages, file- and repository-level data (~600B repository-level tokens; repository concatenation via special tokens), (2) text–code grounding data from Common Crawl and math/programming/education domains, rewritten into normalized Markdown by Qwen3-Coder-480B-A35B-Instruct, (3) PR-based structured SWE tasks mined from GitHub, (4) synthetic single-turn QA and multi-turn agentic trajectories generated by multiple agent frameworks (SWE-agent, Mini-SWE-agent, OpenHands, Claude-Code, Qwen-Code, Terminus) with Qwen3-Coder-480B-A35B-Instruct as teacher, (5) a small instruction-following mix, (6) FIM data from Stack-V2 (chat-FIM and search-and-replace FIM); pretraining corpus updated through Sep 30, 2025[10]
Curriculum / stagingstaged pipeline: continued pretraining (mid-training) on code/agent data with context 32,768 → 262,144 tokens → SFT → specialization of four domain experts → expert distillation into one unified model → single-turn and multi-turn execution-verifiable RL; best-fit packing (BFP) for sample packing; masking of highly repetitive segments[11]
Optimizernot disclosed
LR schedulenot disclosed
Batch schedulenot disclosed
Precisionnot disclosed
Parallelismnot disclosed
Hardwarenot disclosed
Disclosed computenot disclosed

Post-training

SFTyes[12]
Preference optimisationnone disclosed[13]
Reasoning trainingexecution-verifiable RL in two regimes: single-turn RL on unit-test-verified coding tasks (competitive programming, library-usage, multilingual, secure coding) with majority-vote unit tests; multi-turn agentic RL on SWE tasks with trajectory-level completion rewards plus unfinished-trajectory and turn-level tool-format penalties and a reinforced reward-hacking blocker; long-horizon behavior emerged (average agent turns 50 → 130)[14]
Distillationexpert distillation: knowledge from domain-specialized experts (Web Development, User Experience, Single-turn RL/QA, Software Engineering) distilled into the unified SFT model[15]

Modality

Typetext-only
Attachmentn/a[16]

Inference efficiency

KV-cache designnot disclosed
Quantisation shippednone disclosed[17]
Speculative decodingnone disclosed[17]
Serving optimisationsnone disclosed[17]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
SWE-Bench Verified (SWE-Agent / MiniSWE-Agent / OpenHands)70.6 / 71.1 / 71.3Table 3
SWE-Bench Multilingual (SWE-Agent / MiniSWE-Agent / OpenHands)62.8 / 56.2 / 64.3Table 4
SWE-Bench Pro (SWE-Agent / MiniSWE-Agent)42.7 / 38.7Table 4
Terminal-Bench 2.0 (Terminus2-xml / Terminus2-json / ClaudeCode / QwenCode)34.2 / 36.2 / 30.9 / 25.8Table 5
EvalPlus86.56Table 6
MultiPL-E88.23Table 6
CRUXEval95.88Table 6
LiveCodeBench (v6)58.93Table 6
OJBench23.01Table 6
Codeforces2100Table 6
FullStackBench-en60.58Table 7
FullStackBench-zh57.38Table 7
Spider83.66Table 7
BIRD-SQL63.56Table 7
Aider-Polyglot66.2Table 7
MMLU87.73Table 8
MMLU-Redux91.18Table 8
MMLU-Pro80.52Table 8
GPQA74.49Table 8
SuperGPQA57.45Table 8
HMMT25 Feb70.21Table 9
HMMT25 Nov75.57Table 9
AIME2489.01Table 9
AIME2583.07Table 9
AthenaBench-Mini (CTK / RCM / VSP / TAA / RMS / ATE)85.00 / 44.00 / 58.50 / 5.50 / 24.50 / 8.00Table 14
PrimeVul-Paired (Accuracy)48.33Table 15
SecCodeBench (Gen w/o Hint)61.2Table 16
SecCodeBench (Gen w/ Hint)69.5Table 16
SecCodeBench (Fix w/o Hint)76.4Table 16
SecCodeBench (Fix w/ Hint)83.7Table 16
CWEval (func@1)80.17Table 16
CWEval (func-sec@1)56.32Table 16

Provenance references

  1. stated · Title ('Qwen3-Coder-Next Technical Report'); gallery card names it 'Qwen3 Next 80B-A3B'
  2. stated · Title page byline (line: 'Qwen Team')
  3. stated · Title page date line ('arXiv:2603.00729v1 [cs.CL] 28 Feb 2026')
  4. stated · pipeline
  5. stated · Abstract ('We release both base and instruction-tuned open-weight versions'); §1 ('an open-weight language model')
  6. stated · Abstract, §1 ('80 billion total parameters'); Table 3 (Size 80A3)
  7. stated · Abstract, §1 ('activates only 3 billion per forward pass'); Table 3 (Size 80A3)
  8. stated · §1 ('based on Qwen3-Next with hybrid attention and Mixture-of-Experts'); §6 ('hybrid mixture-of-experts architecture')
  9. stated · §3.1.1, §3.2 ('we expand the training context length from 32,768 tokens to 262,144 tokens'; 'we extend the context length beyond typical pretraining settings to 262,144 tokens')
  10. stated · §3.1, §3.1.1, §3.1.2, §3.1.4
  11. stated · §1, §3.2, §4
  12. stated · §4.1 ('we first perform supervised fine-tuning (SFT)'); data from in-house corpora, verified agentic trajectories, and documentation-grounded QA, filtered by a Mini-SWE-agent-based user-simulator verifier and pairwise judging
  13. stated · §4.1
  14. stated · §4.2.3, §4.2.4
  15. stated · §4.2.5
  16. n/a · §6 (text-only; visual capability planned only for future models)
  17. stated · report-wide absence