QWEN MODEL SELECTION GUIDE · 2026
EN

Tongyi Qwen Model Selection Guide

All Qwen models analyzed · hybrid thinking · omni-modal · visual generation · full speech pipeline · open-source deployment — Sources: Alibaba Cloud Bailian official docs Model catalog · Model pricing · Classification docs (synced 2026-08-30)

138Models
12Categories
10MMax context
Tier
FlagshipBalancedBudget
Speed
FastestFastStandardSlowSlowest
Modality
TextImageAudioVideoCode
Reasoning
BestDeepStandardBasicFast
Price
Premium¥20+ / ¥72+Expensive¥8-20 / ¥36-72Costly¥2-8 / ¥8-36Moderate¥0.5-2 / ¥2-8Affordable¥0.1-0.5 / ¥0.4-2Cheapest<¥0.1 / <¥0.4Unit: CNY / 1M tokens (input / output), North China 2 (Beijing) list price

Quick Picks

Direct recommendation by task type

Chat / Writingqwen3.7-plus
Deep thinking / reasoningqwen3.8-max
Top flagshipqwen3.8-max-prime
Cheapestqwen3.7-flash
Expert codingqwen3-coder-plus
Long-document processingqwen-long
Image / video understandingqwen3-vl-plus
Omni-modal chatqwen3.5-omni-plus
Realtime voice chatqwen-audio-3.0-realtime-plus
Speech transcriptionqwen-audio-3.0-asr-flash-filetrans
Speech synthesisqwen-audio-3.0-tts-plus
Image generationqwen-image-3.0-pro
Video generationhappyhorse-1.1-t2v
Semantic search / RAGtext-embedding-v4
Realtime speech translationqwen3.5-livetranslate-flash-realtime

Naming Conventions

Master these patterns and instantly know any model name

SuffixMeaningExample
qwen3.X version number Higher number = newer and stronger qwen3 → qwen3.5 → qwen3.8
-max Flagship, strongest capability qwen3.8-max
-prime Prime mode, faster, stronger and pricier qwen3.8-max-prime
-plus Balanced enhanced qwen3.7-plus
-flash Lightweight and fast, cheapest qwen3.7-flash
-turbo Speed-optimized (legacy) qwen-turbo
qwq- Deep thinking only (thinking mode) qwq-plus
-coder Coding qwen3-coder-plus
-long Ultra-long text only qwen-long
-vl Vision-language model qwen3-vl-plus
-omni Omni-modal (text/image/audio/video) qwen3.5-omni-plus
-ocr Document text extraction qwen3.5-ocr
-asr / -tts Speech recognition / synthesis qwen-audio-3.0-asr-flash
-realtime WebSocket realtime qwen3.5-omni-plus-realtime
-filetrans / -streaming File transcription / streaming realtime qwen-audio-3.0-asr-flash-filetrans
-livetranslate Realtime speech translation qwen3.5-livetranslate-flash
-instruct / -vc / -vd Instruct / voice clone / voice design qwen3-tts-instruct-flash
-embedding / -rerank Embedding / reranking text-embedding-v4, qwen3-rerank
wan2.X- Wanxiang visual generation (image/video) wan2.7-image, wan2.6-t2v
-t2i/-t2v/-i2v/-r2v/-s2v T2I/T2V/I2V/reference-to-video/audio-to-video wan2.6-t2i, wan2.6-r2v
-latest / -YYYYMMDD Latest snapshot / dated snapshot qwen3.7-plus-2026-05-26
cosyvoice- / fun-asr- Speech synthesis / recognition series cosyvoice-v3.5-plus, fun-asr
z-image- / happyhorse- Fast image / third-party video generation z-image-turbo, happyhorse-1.1-t2v

Text Generation Models

Qwen's core text models; the whole family supports hybrid thinking mode (the enable_thinking switch) and 1M context; vision input is covered in the "Vision Understanding" section

Model IDTierPriceModalityThinkingSpeedContextMax outputNotes
qwen3.8-max-primeNEW Flagship
Best Slow 1M 64k Prime mode, currently the strongest, ¥24/¥72
qwen3.8-maxNEW Flagship
Best Slow 1M 64k Latest flagship, hybrid thinking on by default, ¥12/¥36
qwen3.7-max Flagship
Best Slow 1M 64k Previous-gen flagship, 50% off for a limited time
qwen3.7-plusRec Balanced
Deep Standard 1M 64k Flagship balanced choice, ¥2/¥8 (20% off, limited time), built-in tools + web
qwen3.8-27bNEW Balanced
Deep Standard 1M 64k Dense 27B (1M context), ¥3/¥12
qwen3.6-plus Balanced
Deep Standard 1M 64k Previous-gen balanced
qwen3.5-flash Budget
Basic Fast 1M 64k Previous-gen lightweight, from ¥0.2
qwen3.5-plus Balanced
Standard Standard 1M 64k Mature and stable
qwen3.8-flashNEW Budget
Standard Fast 1M 64k Latest lightweight, hybrid thinking on by default, ¥0.8/¥2.7
qwen3.7-flashRec Budget
Basic Fast 1M 64k Cheapest, ¥0.2/¥0.8
qwen3.6-flash Budget
Standard Fast 1M 64k Previous-gen lightweight
qwen3-coder-plus Flagship
Deep Standard 1M 64k Coding-specialized, agentic coding optimized, ¥4/¥16
qwen-long Budget
Basic Standard 10M 8k Ultra-long text (tens of millions of chars), ¥0.5/¥2
qwen-mt-plus Balanced
Basic Fast Machine translation, multilingual
qwen3-max Balanced
Deep Slow 128K 32k Qwen3-generation flagship (legacy), from ¥2.5

QwQ Deep Thinking Models

Thinking-only mode (always reasons before answering; cannot be disabled); for a controllable switch, use the hybrid thinking mode of text generation models

Model IDTierPriceModalityThinkingSpeedContextMax outputNotes
qwq-plusRec Balanced
Best Slow 128K 32k Thinking-only mode, for logic/math reasoning, ¥1.6/¥4

Omni-Modal Models (Omni)

Understands text, audio, images and video at once; outputs text and speech; for realtime voice chat see the "Realtime voice chat" row

Model IDTierPriceInputOutputAPIThinkingFeaturesNotes
qwen3.5-omni-plus Flagship
HTTP Fast FC+web Omni-modal flagship, 3h audio / 1h video
qwen3.5-omni-plus-realtime Flagship
WebSocket Fast FC+web Realtime audio/video chat
qwen3.5-omni-flash Balanced
HTTP Fast FC+web Lightweight omni-modal, from ¥2.2
qwen3.5-omni-flash-realtime Balanced
WebSocket Fast FC+web Realtime
qwen3-omni-flash Budget
HTTP Standard FC Supports thinking mode; single input limited to 150s
qwen3-omni-flash-realtime Budget
WebSocket Fast Realtime; no FC/web/thinking
qwen-audio-3.0-realtime-plusNEW Flagship
WebSocket Fast Semantic VAD+FC S2S voice chat flagship; meaningless backchannels won't interrupt
qwen-audio-3.0-realtime-flashNEW Budget
WebSocket Fast Semantic VAD+FC Cost-sensitive voice chat
qwen-omni-turbo Budget
HTTP Fast Legacy omni-modal; migrate to qwen3.5-omni

Realtime Speech Translation (Livetranslate)

~3s-latency simultaneous interpretation, works out of the box; file mode translates audio/video files

Model IDTierPriceInputAPILanguagesNotes
qwen3.5-livetranslate-flash-realtimeNEWRec Balanced Audio ¥40/1M
WebSocket 60 Top choice for realtime interpreting; 29 speech + 31 text outputs
qwen3.5-livetranslate-flash Balanced Audio ¥40/1M
HTTP 60 Audio/video file translation
qwen3-livetranslate-flash-realtime Budget Audio ¥64/1M
WebSocket 18 Previous-gen realtime, incl. 5 Chinese dialects
qwen3-livetranslate-flash Budget Audio ¥10/1M
HTTP 18 File mode, video-context aware

Vision Understanding Models (VL / OCR)

Image analysis, video understanding, OCR document extraction. Note: all qwen3.8/3.7/3.6/3.5 models in the Text Generation section also support vision input (1M context, up to 2h video); the VL series adds vision-specialized enhancements

Model IDTierPriceModalityThinkingSpeedContextMax outputFeaturesNotes
qwen3-vl-plusRec Flagship
Deep Standard 256K 64k 1h video Vision flagship, from ¥1/¥10
qwen3-vl-flash Budget
Standard Fast 256K 64k 1h video Lightweight vision
qwen3.5-ocr Balanced
Basic Standard Docs/tables/handwriting OCR-specialized, optimized text extraction accuracy
qwen-vl-ocr Budget
Fast Standard Legacy OCR; migrate to qwen3.5-ocr
qvq-max Balanced
Deep Slow Visual reasoning Legacy visual reasoning; capability merged into Qwen3-VL
qvq-plus Budget
Standard Slow Visual reasoning Legacy visual reasoning
qwen-vl-max Balanced
Basic Standard Legacy VL
qwen-vl-plus Budget
Fast Fast Legacy VL

Image Generation Models

Text-to-image and image editing; the qwen-image-3.0 family supports agent prompt rewriting and Chinese text rendering

Model IDTierPriceText-to-imageEditingMax outputMax resolutionNotes
qwen-image-3.0-proNEWRec Flagship Yes Yes 6 2048x2048 Image 3.0 flagship, agent prompt rewriting, small-text/multilingual rendering
qwen-image-3.0NEW Balanced Yes Yes 6 2048x2048 Same as above, faster generation
wan2.7-image-proNEW Flagship Yes Yes 4 (12 consecutive) 4096x4096 Brand color palette, multi-image reference (9 images), character consistency
wan2.7-imageNEW Balanced Yes Yes 4 (12 consecutive) 2048x2048 Same as above, faster generation
z-image-turbo Budget Yes No 1 2048x2048 10x faster, ~1/5 the price, photorealistic portraits/product shots
qwen-image-2.0-pro Balanced Yes Yes 6 2048x2048 Previous-gen pro
qwen-image-2.0 Balanced Yes Yes 6 2048x2048 Previous-gen standard
qwen-image-max Balanced Yes No 1 1664x928 Legacy flagship
qwen-image-plus Budget Yes No 1 1664x928 Legacy balanced
qwen-image-edit-max Balanced No Yes 6 2048x2048 Image editing only
qwen-image-edit-plus Budget No Yes 6 2048x2048 Lightweight image editing
wan2.6-t2i Balanced Yes Yes 4 1440x1440 Wanxiang text-to-image
wan2.6-image Balanced Yes Yes 4 1440x1440 Wanxiang image generation & editing

Video Generation Models (Wanxiang / HappyHorse)

Billed by output video seconds (input free); some models tiered by resolution

Model IDTierTypeBillingResolutionNotes
wan3.0-video-primeNEWRec Flagship Text-to-video ¥0.45-1.8/sec 1080P Latest Wanxiang video flagship
wan3.0-videoNEW Flagship Text-to-video ¥0.3-1.2/sec 1080P 30% off for a limited time
wan2.7-t2vNEW Flagship Text-to-video ¥0.6-1/sec 1080P
wan2.7-i2vNEW Flagship Image-to-video (with audio) ¥0.6-1/sec 1080P
wan2.7-r2vNEW Flagship Reference-to-video (with audio) ¥0.6-1/sec 1080P
happyhorse-1.1-t2vNEWRec Flagship Text-to-video ¥0.45-1.2/sec 1080P 60% off for a limited time, latest third-party video generation
happyhorse-1.1-i2vNEW Flagship Image-to-video ¥0.45-1.2/sec 1080P Image-driven video
happyhorse-1.1-r2vNEW Flagship Reference-to-video ¥0.45-1.2/sec 1080P Reference image/character driven
happyhorse-1.0-video-editNEW Flagship Video editing ¥0.9-1.6/sec 1080P 20% off for a limited time; input + output both billed
wan2.6-t2v Flagship Text-to-video Per-second billing 1080P Wanxiang text-to-video
wan2.6-i2v Flagship Image-to-video Per-second billing 1080P Wanxiang image-to-video
wan2.6-r2v Flagship Reference-to-video Per-second billing 1080P Wanxiang reference-to-video
wan2.2-t2v-plus Balanced Text-to-video Per-second billing 1080P Previous-gen pro
wan2.2-i2v-plus Balanced Image-to-video Per-second billing 1080P Previous-gen pro
wan2.2-kf2v-flash Balanced First/last-frame video Per-second billing 720P Generated by interpolating first + last frames
wan2.2-s2v Balanced Audio-to-video Per-second billing 720P Audio-driven digital human/animation
wan2.2-animate-move Balanced Motion transfer Per-second billing 720P Transfers video motion to an image character
wan2.2-animate-mix Balanced Character swap Per-second billing 720P Character replacement/swap in video
wanx2.1-t2v-turbo Budget Text-to-video Per-second billing 720P Economy
wanx2.1-i2v-turbo Budget Image-to-video Per-second billing 720P Economy
wanx2.1-kf2v-plus Budget First/last-frame video Per-second billing 720P Economy
wanx2.1-vace-plus Budget Video editing Per-second billing 720P Wanxiang video editing

Speech Recognition Models (ASR)

Two paths: realtime (WebSocket streaming) and non-realtime (HTTP file transcription); the Qwen-Audio-3.0 family supports hotwords and Prompt context injection

Model IDTierModeAPIAccuracy boostSpeaker diarizationDuration/SizeNotes
qwen-audio-3.0-asr-flash-streamingNEWRec Flagship Realtime WebSocket Hotwords+Prompt Unlimited Realtime recognition flagship, multilingual incl. dialects
qwen-audio-3.0-asr-flash-filetransNEWRec Flagship Non-realtime HTTP Hotwords+Prompt Yes 12h/2GB File transcription flagship, speaker diarization
qwen-audio-3.0-asr-flashNEW Balanced Non-realtime HTTP Hotwords+Prompt 5 min/2GB Fast short-audio transcription
fun-asr-realtime Balanced Realtime WebSocket Hotwords Unlimited FunASR realtime
fun-asr-mtl-realtime Balanced Realtime WebSocket Hotwords Unlimited FunASR multilingual realtime
fun-asr Balanced Non-realtime HTTP Hotwords Yes 12h/2GB FunASR file version, speaker diarization
fun-asr-mtl Balanced Non-realtime HTTP Hotwords Yes 12h/2GB FunASR multilingual
qwen3-asr-flash-realtime Balanced Realtime WebSocket Unlimited Supports emotion recognition
qwen3-asr-flash-filetrans Balanced Non-realtime HTTP 12h/2GB Supports emotion recognition
qwen3-asr-flash Budget Non-realtime HTTP (OpenAI-compatible) 5 min/10MB Supports emotion recognition
paraformer-realtime-v2 Budget Realtime WebSocket Legacy; migrate to Fun-ASR / Qwen-ASR
paraformer-v2 Budget Non-realtime HTTP Yes Legacy
sensevoice-v1Deprecated Budget Non-realtime HTTP Retiring soon; migrate ASAP
gummy-realtime-v1Deprecated Budget Realtime WebSocket Retiring soon; migrate ASAP
gummy-chat-v1Deprecated Budget Realtime WebSocket 1 min Retiring soon; migrate ASAP

Speech Synthesis Models (TTS)

Standard synthesis / voice cloning (audio sample cloning) / voice design (created from text description); instruct mode tunes tone, emotion and speaking rate in natural language

Model IDTierPriceSeriesAPIVoice cloneVoice designInstructNotes
qwen-audio-3.0-tts-plusNEWRec Flagship Qwen-Audio-TTS WS / HTTP Yes Yes Yes Latest TTS flagship, full capability
qwen-audio-3.0-tts-flashNEW Balanced Qwen-Audio-TTS WS / HTTP Yes Yes Yes Lightweight
cosyvoice-v3.5-plus Flagship CosyVoice WS / HTTP Yes Yes Yes Supports SSML and LaTeX formula reading
cosyvoice-v3.5-flash Balanced CosyVoice WS / HTTP Yes Yes Yes Lightweight
cosyvoice-v3-flash Budget CosyVoice WS / HTTP Yes Yes Yes Lightweight, ¥1/10K chars
cosyvoice-v3-plus Balanced CosyVoice WS / HTTP Yes Yes Previous-gen pro
MiniMax/speech-2.8-hd Balanced MiniMax HTTP Yes Third-party high-fidelity voices
qwen3-tts-flash Budget Qwen3-TTS HTTP Economy standard synthesis
qwen3-tts-instruct-flash Budget Qwen3-TTS HTTP Yes Instructs tone/emotion/style
qwen3-tts-vc Balanced Qwen3-TTS-VC WS / HTTP Yes Voice cloning only
qwen3-tts-vd Balanced Qwen3-TTS-VD WS / HTTP Yes Voice design only
qwen-tts Budget Qwen-TTS (Legacy) HTTP / WS Legacy token-billed version; migrate to Qwen3-TTS
qwen-voice-enrollment Balanced ¥0.01/voice Voice management Voice registration and management (prerequisite for voice cloning)
qwen-voice-design Balanced ¥0.2/voice Voice design service Creates new voices from text descriptions, billed per voice

Embedding & Rerank Models

Semantic search, RAG retrieval and accuracy; rerank is recommended after Embedding retrieval to re-rank Top-N results

Model IDTierPriceTypeDimensionsMax tokensNotes
text-embedding-v4Rec Flagship Text embedding 64~2048 8,192 Latest text embedding, v3-compatible dimensions, MTEB leader
qwen3.7-text-embeddingNEW Flagship Text embedding 64~2048 8,192 New-gen text embedding (Qwen3.7), ¥0.5
text-embedding-v3 Budget Text embedding 512~1024 8,192 Migrate existing v3 indexes
qwen3-vl-embedding Balanced Multimodal embedding 256~2560 32,000 Image-text hybrid retrieval (fused + independent vectors)
qwen2.5-vl-embedding Balanced Multimodal embedding 512~2048 32,000 Image-text hybrid retrieval (fused vectors only)
tongyi-embedding-vision-plus Balanced Multimodal embedding 64~1152 1,024 Cross-modal search (independent vectors)
tongyi-embedding-vision-flash Budget Multimodal embedding 64~768 1,024 Low-cost cross-modal
qwen3-rerankRec Flagship Rerank 4,000/passage 100+ languages, up to 500 docs
qwen3-vl-rerank Balanced Multimodal rerank 8,000/passage Text/image/video hybrid ranking
gte-rerank-v2 Budget Rerank 4,000/passage Text semantic retrieval

Specialized Models

Vertical-use models for translation, document parsing, deep research and text analysis

Model IDTierPriceUseNotes
qwen-mt-plus Balanced Machine translation Multilingual translation, with terminology injection
qwen-doc-turbo Balanced Document parsing Document content extraction and structuring
qwen-deep-research Flagship Deep research Web search + multi-step reasoning + long reports, ¥54/¥163
tongyi-xiaomi-analysis-pro Balanced Conversation analysis Tongyi Xiaomi Pro, ¥1.0/¥2.7
tongyi-xiaomi-analysis-flash Budget Conversation analysis Tongyi Xiaomi Lite, ¥0.2/¥0.4
qwen-math-plus Flagship Math reasoning Math/contest problem solving, ¥4/¥12
qwen-math-turbo Balanced Math reasoning Lightweight math reasoning, ¥2/¥6
qwen-mt-flash Budget Machine translation Lightweight, ¥0.7/¥1.95
qwen-mt-lite Budget Machine translation Economy, ¥0.6/¥1.6
qwen-mt-turbo Budget Machine translation Speed-optimized, ¥0.7/¥1.95
qwen-mt-image-2.0 Balanced Image translation Translates text in images into multiple languages, ¥0.004/image
fun-music-v1 Budget ¥0.002/sec Music generation Generates music from lyrics/prompts, billed per second

Open-Source Models (Self-Deployable)

Open weights, self-deployable, no API call limits, only hardware costs; Bailian also offers hosted calling of the open-source versions

Model IDTierPriceThinkingContextNotes
qwen3.8-2.4t-a95bNEWOSS Flagship Self-hosted Best Latest open-source flagship, hybrid thinking on by default
qwen3.6-35b-a3bOSS Balanced Self-hosted Deep 256K Open-source balanced (MoE 35B-A3B)
qwen3.6-27bOSS Balanced Deep Open-source dense 27B (Qwen3.6-gen), ¥3/¥18
qwen3.5-397b-a17bOSS Flagship Self-hosted Deep 32K Open-source flagship (MoE 397B-A17B)
qwen3.5-122b-a10bOSS Balanced Self-hosted Standard 32K Open-source mid-size (MoE 122B-A10B)
qwen3.5-35b-a3bOSS Budget Self-hosted Standard 32K Open-source small (MoE 35B-A3B)
qwen3.5-27bOSS Budget Self-hosted Standard 32K Open-source dense 27B
qwen3-next-80b-a3b-instructOSS Flagship Fast 256K Qwen3-Next new-gen open-source MoE 80B-A3B, ¥1/¥4
qwen3-next-80b-a3b-thinkingOSS Flagship Deep 256K Qwen3-Next thinking edition (thinking-only), ¥1/¥10
qwen3-coder-nextOSS Flagship Fast 256K New open-source coding model, ¥1/¥4
qwen3-235b-a22b-instruct-2507OSS Flagship Fast 256K Qwen3 open-source flagship MoE 235B-A22B
qwen3-235b-a22b-thinking-2507OSS Flagship Deep 256K Qwen3 open-source thinking flagship, ¥2/¥20
qwen3-vl-235b-a22b-instructOSS Flagship Fast 256K Open-source vision flagship MoE 235B-A22B
qwen3-vl-235b-a22b-thinkingOSS Flagship Deep 256K Open-source vision thinking flagship
qwen3-omni-30b-a3b-captionerOSS Balanced Fast Open-source omni-modal video captioning/description

Legacy & Retiring Models

Legacy/retiring info for the models covered on this page. Migrate as recommended and don't use them in new projects. Source: Alibaba Cloud Bailian official docs (synced 2026-08-30)

Model IDLifecycleReplacementMigration advice
sensevoice-v1Deprecated Deprecated fun-asr Retiring soon; migrate to Fun-ASR / Qwen-ASR
gummy-realtime-v1Deprecated Deprecated fun-asr-realtime Retiring soon; migrate ASAP
gummy-chat-v1Deprecated Deprecated fun-asr-realtime Retiring soon; migrate ASAP
qwen-omni-turbo Legacy qwen3.5-omni-plus Legacy omni-modal; capabilities fully upgraded
qwen-tts Legacy qwen3-tts-flash / qwen-audio-3.0-tts-plus Legacy token-billed TTS
paraformer-realtime-v2 Legacy fun-asr-realtime / qwen3-asr-flash-realtime Legacy ASR, migration advised
paraformer-v2 Legacy fun-asr / qwen3-asr-flash-filetrans Legacy ASR
qwen-vl-ocr Legacy qwen3.5-ocr Legacy OCR
qvq-max Legacy qwen3-vl-plus Visual reasoning merged into Qwen3-VL
qvq-plus Legacy qwen3-vl-flash Legacy visual reasoning
qwen-vl-max Legacy qwen3-vl-plus Legacy VL
qwen-vl-plus Legacy qwen3-vl-flash Legacy VL

Capability Matrix

Quickly match the best model to your needs

ScenarioRecommendedAlternativesKey capability
Chat / Writing qwen3.7-plus qwen3.6-plus Deep + 1M context
Deep thinking / reasoning qwen3.8-max qwq-plus Best Hybrid thinking
Top flagship qwen3.8-max-prime qwen3.8-max ¥24/¥72 Prime mode
Ultra-low cost qwen3.7-flash qwen3.8-flash ¥0.2/¥0.8
Agentic coding workflows qwen3-coder-plus qwen3.7-plus + 1M context
Long-document processingTens of millions of chars qwen-long qwen3.7-plus 10M context
Image / video understanding qwen3-vl-plus qwen3.7-plus 2h video · 1M context
Omni-modal understanding (image+audio+video) qwen3.5-omni-plus qwen3-omni-flash
Realtime voice chat qwen-audio-3.0-realtime-plus qwen3.5-omni-plus-realtime Semantic VAD + WebSocket
Speech transcription / dictation qwen-audio-3.0-asr-flash-filetrans fun-asr 12h + speaker diarization
Speech synthesis qwen-audio-3.0-tts-plus cosyvoice-v3.5-plus Clone / Design / Instruct
Realtime speech translation qwen3.5-livetranslate-flash-realtime qwen3-livetranslate-flash-realtime 60 languages · 3s latency
High-quality image generation qwen-image-3.0-pro wan2.7-image-pro 2048x2048 / 4096x4096
Video generation happyhorse-1.1-t2v wan2.6-t2v 1080P · per-second billing
Semantic search / RAG text-embedding-v4 qwen3-vl-embedding 2048-dim + qwen3-rerank
Self-hosted deployment qwen3.8-2.4t-a95b qwen3.6-35b-a3b Open weights