multimodal LLM
Qwen3.5-397B-A17B
Alibaba's 397B MoE model for text, image and video; teacher for SEA-LION v4.5 distillation. Adapters 18, finetunes 41, quants 77.
Levelmodel
Classmultimodal LLM
Parameters397B
Context262,144
ReleasedFeb 2026
Opennessopen weights permissive
Runs onmulti-GPU node
Statusactive
Benchmark Results
vendor-reported
GPQAknowledge and reasoning
vendor-reported
88.4
050100
Score in %. Higher is better.
MMLU-Proknowledge and reasoning
vendor-reported
87.8
050100
Score in %. Higher is better.
SWE-bench Verifiedcoding
vendor-reported
76.4
050100
Score in %. Higher is better.
Where this comes from
Every score names who ran it: vendor-reported, maintainer-run or independent. A rank is shown with the population it was ranked in (gap rule 3).
Lineage
distillfinetune
Qwen-SEA-LION-v4.5-27B-IT is a child (distill)high confidencemodel cardWhere this comes from
Parent and child relations from model cards, Hugging Face base_model fields and the genealogy tables of the research notes. The tree shows two generations each way, 3 models. Confidence and evidence are the notes' own.
R&D Topics
Identity
- Developer
- Alibaba (Qwen)
- Family
- Qwen3.5
- Released
- Feb 2026 sourcesnotes
- Weights published
- –
- Record checked
- 10 Oct 2026
Where this comes from
Developer and dates from the research notes. Release date is the vendor's public announcement; weights date is when open weights were published.
Architecture and Size
- Architecture
- hybrid Gated DeltaNet + Gated Attention, MoE (512 experts, 10 + 1 active), MTP
- Total parameters
- 397B sources
- Active parameters
- 17B
- Checkpoint parameters
- –
- Note
- From model name
- Context length
- 262,144 tokens
- Input
- text, image, video
- Output
- text
Where this comes from
Parameter counts as the developer states them. Where the checkpoint holds a different total, both are shown (register A16).
Openness and Access
- Openness
- open weights permissive
- Licence
- Apache License 2.0 sources
- Access routes
- open weights download
Where this comes from
The licence file governs over registry tags and aggregator labels.
Deployment
- Runs on
- multi-GPU node
- Footprint
- FP8 and GPTQ-Int4 variants
- Runtimes
- not in the notes
Where this comes from
Deployment tiers and runtimes as named in the notes. Footprint is the notes' figure, not a measurement by AURION.
Indonesia
- Indonesian support
- not stated
- Regional languages
- not in the notes
- Jakarta hosting
- unknown
- Languages
- Card claims 201 languages; Indonesian not individually confirmed
Where this comes from
Indonesia fields are carried on every model; unknown is an allowed value.
Industry Use Cases
No industry use case links this model.
Links
DeveloperAlibaba (Qwen)big tech, China
Alibaba (Qwen)big tech, ChinaModelQwen3.5-397B-A17BAlibaba (Qwen), multimodal LLM
Qwen3.5-397B-A17BAlibaba (Qwen), multimodal LLMR&D TopicsEfficiency Architectures (MoE, Sparse Attention, Long Context)Low-Resource and Regional Language ResourcesOpen-Weight Frontier Releases as Research SubstrateRegional or National Foundation Model Training (Continued Pretraining at 7B to 70B)
Use CasesIndonesian and Regional-Language LLMsPublic SectorIndonesian NLU (Classification, NER)Public SectorBanking Customer Service AssistantsFinanceCitizen Service AssistantsPublic Sector+8 moreUse cases of models in the same R&D topic; none is recorded for this model itself.
Who Runs It
PT Amman Mineral Nusa Tenggara (AMMAN)In production, 8 Jul 2026
PT Bank Tabungan Negara (BTN)In production, 1 Jul 2026
PT PegadaianIn production, 5 Jun 2026
Bank Central Asia (BCA)In production, 16 Mar 2026+19 moreDeployments of these use cases; none names this model.
PT Amman Mineral Nusa Tenggara (AMMAN)In production, 8 Jul 2026PT Bank Tabungan Negara (BTN)In production, 1 Jul 2026PT PegadaianIn production, 5 Jun 2026Bank Central Asia (BCA)In production, 16 Mar 2026+19 moreDeployments of these use cases; none names this model.Organisations 1
R&D topics 4
Open-Weight Frontier Releases as Research SubstrateUsing released weights as the base for downstream research.used byEfficiency Architectures (MoE, Sparse Attention, Long Context)Reducing training and inference cost per capability.produced byRegional or National Foundation Model Training (Continued Pretraining at 7B to 70B)Pretraining or continued pretraining a mid-size model to serve a national or regional lan…used byLow-Resource and Regional Language ResourcesDatasets, benchmarks and adaptation for under-served languages.used by
Sources
- Text Arena: LLM Rankings & AI Model Benchmarks arena.ai https://arena.ai/leaderboard/chat/text
- Qwen/Qwen3.5-397B-A17B , Hugging Face huggingface.co https://huggingface.co/Qwen/Qwen3.5-397B-A17B
- Qwen3.5 - a Qwen Collection huggingface.co https://huggingface.co/collections/Qwen/qwen35
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Qwen3.5-397B-A17B", https://aurion.semai.id/models/alibaba-qwen3-5-397b-a17b/, records of October 2026.