◆ AsiaGPU

Models

Only open-source models, and only the best current ones from each developer. Every model is tested on real GPUs with exact settings before it's offered, and pinned to the exact version that was tested, so nobody can slip a changed copy in later. When a developer releases something better, we test it, recommend it, and retire the older one with at least 60 days' notice by email. Your chats and files stay in your workspace whatever model you use.

Available now

ModelDeveloperGood forLicenceStage
Qwen3 8BAlibaba (Qwen)Alibaba's 8B general model. Strong in Chinese, English and Southeast Asian languages; can think step by step before answering.Apache-2.0 (commercial use allowed)Available
Qwen3-VL 8BAlibaba (Qwen)Alibaba's 8B vision model: everything Qwen3 8B does for everyday work, and it reads pictures too (invoices, screenshots, charts, photos).Apache-2.0 (commercial use allowed)Recommended
gpt-oss 20BOpenAIOpenAI's open-weight reasoning model (21B, 3.6B active per token, so it answers fast). Thinks before answering; good at tools and coding. Strongest in English.Apache-2.0 (commercial use allowed)Available
Gemma 4 12BGoogleGoogle's 12B model in its 4-bit version (quantised by Google, so quality is kept): reads pictures, 35+ languages. Light on memory: leaves most of a 24 GB card free.Apache-2.0 (commercial use allowed)Available
Ministral 3 14BMistral AIMistral's 14B model (FP8): everyday chat, tools and pictures. Strong in European languages, also Chinese, Japanese, Korean and Arabic.Apache-2.0 (commercial use allowed)Available
Z-Image TurboAlibaba (Tongyi-MAI)Alibaba's 6B picture model: photorealistic images, posters and illustrations from a text description, in a few seconds. Understands Chinese and English prompts, including text in pictures.Apache-2.0 (commercial use allowed)Recommended
FLUX.2 klein 4BBlack Forest LabsBlack Forest Labs' small picture model: creates pictures from a description, and edits photos you give it ("make the background white", "add a red scarf").Apache-2.0 (commercial use allowed)Available
Wan 2.2 5BAlibaba (Wan)Alibaba's video model: 5-second 720p clips from a description, or from a picture you give it. Each takes a few minutes.Apache-2.0 (commercial use allowed)Available

Coming next

Planned, not yet tested. When we add one we use that developer's newest version, which may have a newer name than listed here. All of them allow commercial use. Hardware each one needs.

One 24-32 GB card · RTX 4090 / 5090 class: most models, and the cheapest to run.

ModelDeveloperForWhyLicence
DeepSeek R1 8BDeepSeekHard problems (thinks step by step)DeepSeek's reasoning distilled into Qwen3 8B: maths, logic and code, step by step.MIT
Qwen3 14BAlibaba (Qwen)Everyday chat and writingSmarter than 8B, still fast; strong in Asian languages.Apache-2.0
Qwen3 30B-A3B (4-bit)Alibaba (Qwen)Everyday chat and writingNear-32B quality at the speed of a 3B model.Apache-2.0
Phi-4 reasoningMicrosoftHard problems (thinks step by step)Small model that is very good at maths and logic.MIT
SEA-LION v4AI SingaporeEveryday chat and writingTuned for Southeast Asian languages: Malay, Indonesian, Thai, Vietnamese, Filipino, Tamil, Burmese, Khmer, Lao.Gemma terms (commercial use allowed)
Typhoon 2SCB 10XEveryday chat and writingBest for Thai.Llama / Qwen terms (commercial use allowed)
Sailor2 20BSea AI LabEveryday chat and writingSoutheast Asian languages, including local ones like Javanese and Sundanese.Apache-2.0
Qwen3-Coder 30B-A3B (4-bit)Alibaba (Qwen)CodingCoding assistant that works with coding tools and editors.Apache-2.0
Devstral SmallMistral AICodingCoding agent that edits whole projects.Apache-2.0
Qwen3-VL 32B (4-bit)Alibaba (Qwen)Reading picturesReads charts, forms and screenshots more accurately than the 8B.Apache-2.0
PaddleOCR-VLBaiduScanned documentsTurns scanned PDFs and photos of documents into text and tables, in 100+ languages.Apache-2.0
Qwen3 Reranker 0.6BAlibaba (Qwen)Better document searchFinds the right passages in long documents more reliably.Apache-2.0
Qwen3GuardAlibaba (Qwen)Content safety checksLets a business block unsafe questions and answers.Apache-2.0
CosyVoiceAlibaba (FunAudioLLM)Natural voices (text to speech)Natural voices in Chinese, English, Japanese, Korean and dialects.Apache-2.0
Chatterbox MultilingualResemble AINatural voices (text to speech)Natural voices in 23 languages, including Malay.MIT
Kokoro 82MhexgradNatural voices (text to speech)Tiny and very fast voice for English; fits next to any model.Apache-2.0
SenseVoiceAlibaba (FunAudioLLM)Speech to textFaster speech-to-text for Chinese, Cantonese, English, Japanese and Korean; detects emotion.Apache-2.0 style (commercial use allowed)
ACE-StepACE Studio & StepFunCreating musicSongs and background music from a description.Apache-2.0

One 48 GB card · L40S or RTX 6000: bigger picture models.

ModelDeveloperForWhyLicence
Qwen3.8 27BAlibaba (Qwen)Everyday chat and writingAlibaba's newest mid-size model: reads pictures, thinks step by step; needs two 24 GB cards.Apache-2.0
Qwen-Image-EditAlibaba (Qwen)Editing picturesEdits pictures by instruction: change text, remove objects, restyle.Apache-2.0
Qwen-ImageAlibaba (Qwen)Creating picturesBest open model for posters and pictures with Chinese or English text.Apache-2.0

One 80-141 GB card · A100, H100 or H200: higher quality.

ModelDeveloperForWhyLicence
gpt-oss-120bOpenAIHard problems (thinks step by step)OpenAI's biggest open model, on a single H100.Apache-2.0
Llama 4 Scout (4-bit)MetaEveryday chat and writingMeta's model; reads pictures; remembers very long documents.Llama 4 licence (commercial use allowed)
GLM-4.5-Air (4-bit)Zhipu AI (Z.ai)Everyday chat and writingStrong all-rounder for agents and tool use.MIT
Wan 2.2 A14BAlibaba (Wan)Creating videosHigher-quality video from text or a picture.Apache-2.0
Qwen3-Next 80B-A3BAlibaba (Qwen)Everyday chat and writingVery fast for its quality; long memory.Apache-2.0

Several cards · The best open models available, for teams that need top quality.

ModelDeveloperForWhyLicence
MiniMax-M2MiniMaxCodingTop open model for coding and agents.MIT
Qwen3 235B-A22BAlibaba (Qwen)Everyday chat and writingAlibaba's flagship.Apache-2.0
GLM-4.6Zhipu AI (Z.ai)CodingFlagship for coding and long tasks.MIT
DeepSeek V3.2DeepSeekHard problems (thinks step by step)Among the very best open models; strong reasoning.MIT
Kimi K2Moonshot AIEveryday chat and writing1-trillion-parameter model; excellent at tools and agents.Modified MIT (commercial use allowed)
Mistral Large 3Mistral AIEveryday chat and writingMistral's flagship; reads pictures.Apache-2.0

Hardware each model needs

For planning machines, and for hosts checking what their machine can serve. GPU memory is what one deployment needs, with room for several people at once; a model can share a card with others while the card has memory left. The GPU columns show how many cards of that type it takes in one machine (up to 8). RAM is the computer's own memory we recommend, and disk the model's download.

Measured for models available now; estimated from each model's size for the planned ones until they are tested. Every machine also needs disk space once for the AI engines themselves: 33 GB for the language-model engine and 53 GB for the picture engine.

ModelGPU memoryRAMDiskRTX 4090
24 GB
RTX 5090
32 GB
L40S / RTX 6000
48 GB
A100 / H100
80 GB
RTX PRO 6000
96 GB
H200
141 GB
Qwen3 8B
8B · BF16 · measured
20 GB16 GB15 GB1 card1 card1 card1 card1 card1 card
Qwen3-VL 8B
8B · FP8 · measured
20 GB16 GB9.9 GB1 card1 card1 card1 card1 card1 card
gpt-oss 20B
21B total, 3.6B active · MXFP4 · measured
20 GB16 GB14 GB1 card1 card1 card1 card1 card1 card
Gemma 4 12B
12B · 4-bit · measured
16 GB16 GB10 GB1 card1 card1 card1 card1 card1 card
Ministral 3 14B
14B · FP8 · measured
21 GB16 GB16 GB1 card1 card1 card1 card1 card1 card
Z-Image Turbo
6B · BF16 · measured
20 GB32 GB31 GB1 card1 card1 card1 card1 card1 card
FLUX.2 klein 4B
4B · BF16 · measured
21 GB16 GB15 GB1 card1 card1 card1 card1 card1 card
Wan 2.2 5B
5B · BF16 · measured
27 GB64 GB32 GB—1 card1 card1 card1 card1 card
Kokoro 82M
82M · BF16
1 GB16 GB0.4 GB1 card1 card1 card1 card1 card1 card
SenseVoice
234M · BF16
2 GB16 GB1 GB1 card1 card1 card1 card1 card1 card
Qwen3 Reranker 0.6B
0.6B · BF16
3 GB16 GB1.2 GB1 card1 card1 card1 card1 card1 card
PaddleOCR-VL
0.9B · BF16
6 GB16 GB2 GB1 card1 card1 card1 card1 card1 card
CosyVoice
0.5B · BF16
6 GB16 GB5 GB1 card1 card1 card1 card1 card1 card
Chatterbox Multilingual
0.5B · BF16
6 GB16 GB3 GB1 card1 card1 card1 card1 card1 card
Qwen3Guard
4B · BF16
11 GB16 GB8 GB1 card1 card1 card1 card1 card1 card
ACE-Step
3.5B · BF16
14 GB16 GB8 GB1 card1 card1 card1 card1 card1 card
Typhoon 2
8B · BF16
20 GB16 GB16 GB1 card1 card1 card1 card1 card1 card
Qwen3 14B
14.8B · FP8
21 GB16 GB16 GB1 card1 card1 card1 card1 card1 card
Qwen3 30B-A3B (4-bit)
30B total, 3B active · 4-bit
21 GB32 GB17 GB1 card1 card1 card1 card1 card1 card
Phi-4 reasoning
14B · FP8
21 GB16 GB15 GB1 card1 card1 card1 card1 card1 card
DeepSeek R1 8B
8B · BF16
22 GB32 GB16 GB2 cards1 card1 card1 card1 card1 card
Devstral Small
24B · 4-bit
22 GB16 GB14 GB2 cards1 card1 card1 card1 card1 card
SEA-LION v4
27B · 4-bit
24 GB32 GB17 GB2 cards1 card1 card1 card1 card1 card
Qwen3-Coder 30B-A3B (4-bit)
30B total, 3B active · 4-bit
24 GB32 GB17 GB2 cards1 card1 card1 card1 card1 card
Sailor2 20B
20B · FP8
26 GB32 GB20 GB2 cards1 card1 card1 card1 card1 card
Qwen3-VL 32B (4-bit)
33B · 4-bit
27 GB32 GB20 GB2 cards1 card1 card1 card1 card1 card
Qwen3.8 27B
27B · FP8
38 GB32 GB31 GB2 cards2 cards1 card1 card1 card1 card
Qwen-Image-Edit
20B · FP8
42 GB64 GB37 GB——1 card1 card1 card1 card
Qwen-Image
20B · FP8
42 GB64 GB37 GB——1 card1 card1 card1 card
Llama 4 Scout (4-bit)
109B total, 17B active · 4-bit
68 GB48 GB58 GB4 cards4 cards2 cards1 card1 card1 card
GLM-4.5-Air (4-bit)
106B total, 12B active · 4-bit
68 GB48 GB57 GB4 cards4 cards2 cards1 card1 card1 card
Wan 2.2 A14B
27B total, 14B active · BF16
70 GB128 GB68 GB———1 card1 card1 card
gpt-oss-120b
117B total, 5B active · MXFP4
72 GB48 GB65 GB4 cards4 cards2 cards1 card1 card1 card
Qwen3-Next 80B-A3B
80B total, 3B active · FP8
95 GB64 GB82 GB8 cards4 cards4 cards2 cards2 cards1 card
MiniMax-M2
230B total, 10B active · FP8
260 GB128 GB230 GB——8 cards4 cards4 cards4 cards
Qwen3 235B-A22B
235B total, 22B active · FP8
270 GB128 GB236 GB——8 cards4 cards4 cards4 cards
GLM-4.6
355B total, 32B active · FP8
400 GB192 GB355 GB———8 cards8 cards4 cards
Kimi K2
1T total, 32B active · 4-bit
650 GB320 GB594 GB————8 cards8 cards
Mistral Large 3
675B total, 41B active · FP8
750 GB352 GB680 GB—————8 cards
DeepSeek V3.2
685B total, 37B active · FP8
760 GB368 GB690 GB—————8 cards