Models

Open-weight models, served on your terms.

ARK supports model families, not a fixed list. Any model whose weights are built on an architecture ARK runs comes online by conversion and configuration, with no change to the platform. 6 chat models, one embedding model and one speech-to-text model answer on ARK Cloud today.

Models and fine-tunes
29,000+official models and their fine-tunes ARK can serve, counted 20 October 2026
Range
0.6B to 480Bparameters for chat; size is a question of GPUs, not of code
Attention
1M tokensshared by all requests in flight on Qwen3.6-35B-A3B on ARK Cloud, up to 256K per request
Live on ARK Cloud
6 chat modelsplus embeddings and speech to text, ready to call
On ARK Cloud

Call these right now.

Live on the API with the prices shown. GET /api/v1/models always lists what is online.

All prices
On ARK CloudSelf-hosted

Qwen3.6-35B-A3B

Qwen (Alibaba Cloud) · April 2026

A mixture-of-experts vision-language model: fast, multimodal and a strong all-rounder for chat, tools and documents.

Parameters
35B total · 3B active
Context on ARK Cloud
256K tokens
Licence
Apache 2.0
VisionReasoningTools
ARK Cloud price$0.0156 in · $0.4499 out
arklabsai/Qwen3.6-35B-A3B
On ARK CloudSelf-hosted

Qwen3.5-122B-A10B

Qwen (Alibaba Cloud) · February 2026

A mixture-of-experts vision-language model for hard reasoning and long documents, in 201 languages.

Parameters
122B total · 10B active
Context on ARK Cloud
256K tokens
Licence
Apache 2.0
VisionReasoningTools
ARK Cloud price$0.038 in · $1.3902 out
arklabsai/Qwen3.5-122B-A10B
On ARK CloudSelf-hosted

Qwen3.8-Flash-Next

Qwen (Alibaba Cloud) · August 2026

An experimental preview of the architecture behind Qwen4: a 512-expert mixture of experts with Qwen Sparse Attention and n-gram embeddings, and a vision encoder.

Parameters
125B total · 6B active
Context on ARK Cloud
256K tokens
VisionReasoningTools
ARK Cloud price$0.057 in · $2.0853 out
arklabsai/Qwen3.8-Flash-Next
On ARK CloudSelf-hosted

Qwen3.8-27B

Qwen (Alibaba Cloud) · August 2026

The newest dense Qwen vision-language model, for text and images.

Parameters
27B
Context on ARK Cloud
256K tokens
Licence
Apache 2.0
VisionReasoningTools
ARK Cloud price$0.038 in · $1.3902 out
arklabsai/Qwen3.8-27B
On ARK CloudSelf-hosted

Qwen3.5-4B

Qwen (Alibaba Cloud) · February 2026

A small vision-language model for high-volume, low-latency work: classification, extraction, routing.

Parameters
4B
Context on ARK Cloud
128K tokens
Licence
Apache 2.0
VisionTools
ARK Cloud price$0.0097 in · $0.3268 out
arklabsai/Qwen3.5-4B
On ARK CloudSelf-hosted

Bielik-11B-v3.0-Instruct

SpeakLeash & ACK Cyfronet AGH · November 2025

A Polish-first language model trained across 32 European languages, built by Poland's SpeakLeash community and ACK Cyfronet AGH.

Parameters
11B
Context on ARK Cloud
32K tokens
Licence
Apache 2.0
Polish-first
ARK Cloud price$0.10 in · $0.40 out
arklabsai/Bielik-11B-v3.0-Instruct
On ARK CloudSelf-hosted

bge-m3

BAAI · January 2024

Multilingual embeddings for search and retrieval: 1,024 dimensions, more than 100 languages, inputs up to 8,192 tokens.

Parameters
568M
Context on ARK Cloud
8K tokens per input · 1,024 dims
Licence
MIT
Embeddings
ARK Cloud price$0.0002 per 1M tokens
On ARK CloudSelf-hosted

whisper-large-v3-turbo

OpenAI · October 2024

Speech recognition in 99 languages, with text, segment timestamps and SRT or WebVTT subtitles.

Parameters
809M
Context on ARK Cloud
Any length of audio
Licence
MIT
Speech to text
ARK Cloud price$0.0018 per audio hour
arklabsai/whisper-large-v3-turbo
Supported families

Everything ARK can run, by family.

Every model below shares its architecture with models we run. The named models are the official checkpoints of each family. Behind each family stand its fine-tunes on Hugging Face, 29,464 in all, supported the same way.

Qwen (Alibaba Cloud)

Qwen3.5 · Qwen3.6 · Qwen3.8

Hybrid-attention vision-language models with a 256K-token native context, dense and mixture-of-experts.

ChatVisionReasoningTools
Architecture
Qwen3_5 · Qwen3_5Moe (conditional generation)
Chat format
ChatML with <think> and <tool_call>
Licence
Apache 2.0
Fine-tunes on Hugging Face
3,761 with this architecture
Qwen (Alibaba Cloud)

Qwen3.8-Flash-Next

The first open-weight release of the architecture behind Qwen4: Gated DeltaNet with Qwen Sparse Attention, n-gram embeddings and a mixture of experts, with a vision encoder.

ChatVisionReasoningTools
Architecture
Qwen4Exp (conditional generation)
Chat format
ChatML with <think> and <tool_call>
Fine-tunes on Hugging Face
46 with this architecture
Qwen (Alibaba Cloud)

Qwen3 · Qwen3-Coder

Dense and mixture-of-experts text models with switchable thinking, plus the Instruct-2507, Thinking-2507 and Coder lines.

ChatReasoningToolsCode
Architecture
Qwen3ForCausalLM · Qwen3MoeForCausalLM
Chat format
ChatML with <think> and <tool_call>
Licence
Apache 2.0
Fine-tunes on Hugging Face
9,402 with this architecture
Meta

Llama 3 · 3.1 · 3.2 · 3.3

Meta's multilingual instruction models, from 1B to 405B parameters.

ChatTools
Architecture
LlamaForCausalLM
Chat format
Llama 3 chat format with tool tokens
Fine-tunes on Hugging Face
7,966 with this architecture

Llama 3 (the first release) has no tool-call format: it is served for chat.

Mistral AI

Ministral 3 · Mistral Small · Magistral · Devstral

Mistral AI's open models for chat, reasoning and agentic coding. ARK serves their text side; image input is not served for this family.

ChatReasoningToolsCode
Architecture
Mistral3 (conditional generation) · Ministral3ForCausalLM
Chat format
Mistral chat format with [THINK] and [TOOL_CALLS]
Licence
Apache 2.0 (Devstral 2 123B: modified MIT with a revenue limit)
Fine-tunes on Hugging Face
175 with this architecture
Z.ai

GLM-4.7-Flash

The lightweight mixture-of-experts member of the GLM-4.7 series, for English and Chinese.

ChatReasoningTools
Architecture
Glm4MoeLiteForCausalLM
Chat format
GLM chat format with <think> and <tool_call>
Licence
MIT
Fine-tunes on Hugging Face
40 with this architecture
SpeakLeash & ACK Cyfronet AGH

Bielik

Polish-first language models built in Poland, from 1.5B to 11B parameters. Served for chat, without tool calls or separate reasoning.

ChatPolish-first
Architecture
LlamaForCausalLM
Chat format
ChatML
Licence
Apache 2.0
Fine-tunes on Hugging Face
39 with this architecture

Bielik 11B v2.0 to v2.3 and Bielik 7B use the Mistral architecture, which ARK has not served since ARK 2.0: ask us about them.

BAAI

BGE-M3

Multilingual dense embeddings: 1,024 dimensions, more than 100 languages, inputs up to 8,192 tokens.

Embeddings
Architecture
XLMRobertaModel
Chat format
Embeddings
Licence
MIT
Fine-tunes on Hugging Face
459 with this architecture
OpenAI

Whisper

Speech recognition in 99 languages, with segment timestamps and subtitles.

Speech to text
Architecture
WhisperForConditionalGeneration
Chat format
Speech to text
Licence
Apache 2.0 · MIT (large-v3-turbo)
Fine-tunes on Hugging Face
7,405 with this architecture

The multilingual checkpoints, from tiny (39M parameters) to large. The English-only .en variants are not confirmed yet: ask us about them.

What "supported" means

No change to the platform.

A model is supported when ARK can bring it online without any change to the ARK platform: we convert the model with ARK's tooling, publish it, add its configuration and start a compute service, and it answers on the API. That works for every model whose weights are built on an architecture ARK runs today. Conversion changes how the weights are stored, never what they are; the chat template and the model's configuration are configuration, and can be set per model.

  • Benchmarked models run in our release benchmark each time ARK ships.
  • Supported models share their architecture with a benchmarked model but have not been converted yet. We confirm each one when we onboard it.
  • Fine-tunes of these models are supported on the same terms, with whatever chat template they ship. Hugging Face listed 29,464 fine-tunes and merges of the official checkpoints with the same architecture and full-precision weights (counted on 20 October 2026; Qwen3.8-Flash-Next's on 6 November).
  • Tool calls and reasoning are read in the formats of the families above. A fine-tune that answers in one of them gets both; one that invents a format of its own answers chat, and its tool calls need that format added to ARK.
  • Size is a hardware question: the largest models need a correspondingly large GPU fleet. See the supported hardware.

Not supported today

  • Base (pre-trained) checkpoints without chat tuning: ARK serves chat, not plain text completion.
  • Pre-quantized uploads such as GGUF, AWQ or GPTQ: ARK converts from full-precision weights.
  • Vision models outside the Qwen3.5, Qwen3.6 and Qwen3.8 family, such as Qwen3-VL or Llama 3.2 Vision.
  • Architectures ARK does not run yet, such as Qwen3-Next, the full GLM-4.7, Gemma or Phi. These need engineering work first.
  • Specialist models outside ARK's endpoints: rerankers, classifiers and guard models, text-to-speech.

Need one of these, or a model you trained yourself? Talk to us.

Licences and model cards

Each model comes with its maker's terms.

ARK serves open-weight models under the licences and use policies their makers publish. When you use a model on ARK Cloud or in your own deployment, its licence and use policy apply alongside our terms of service. Read them on each model card.

FamilyLicenceModel cards
Qwen3.5 · Qwen3.6 · Qwen3.8Qwen (Alibaba Cloud)Apache 2.0Qwen3.6-35B-A3B, Qwen3.5-122B-A10B, Qwen3.5-4B, Qwen3.8-27B, Qwen3.6-27B, Qwen3.5-0.8B, Qwen3.5-2B, Qwen3.5-9B, Qwen3.5-27B, Qwen3.5-35B-A3B, Qwen3.5-397B-A17B
Qwen3.8-Flash-NextQwen (Alibaba Cloud)Qwen Community License 1.0Qwen3.8-Flash-Next
Qwen3 · Qwen3-CoderQwen (Alibaba Cloud)Apache 2.0Qwen3-8B, Qwen3-14B, Qwen3-32B, Qwen3-Coder-30B-A3B-Instruct, Qwen3-0.6B, Qwen3-1.7B, Qwen3-4B, Qwen3-30B-A3B, Qwen3-235B-A22B, Qwen3-4B-Instruct-2507, Qwen3-4B-Thinking-2507, Qwen3-30B-A3B-Instruct-2507, Qwen3-30B-A3B-Thinking-2507, Qwen3-235B-A22B-Instruct-2507, Qwen3-235B-A22B-Thinking-2507, Qwen3-Coder-480B-A35B-Instruct
Llama 3 · 3.1 · 3.2 · 3.3MetaLlama 3, 3.1, 3.2 and 3.3 Community LicensesLlama-3.1-8B-Instruct, Llama-3.2-1B-Instruct, Llama-3.2-3B-Instruct, Llama-3.3-70B-Instruct, Llama-3.1-70B-Instruct, Llama-3.1-405B-Instruct, Meta-Llama-3-8B-Instruct, Meta-Llama-3-70B-Instruct
Ministral 3 · Mistral Small · Magistral · DevstralMistral AIApache 2.0 (Devstral 2 123B: modified MIT with a revenue limit)Ministral-3-8B-Reasoning-2512, Ministral-3-14B-Reasoning-2512, Devstral-Small-2-24B-Instruct-2512, Ministral-3-3B-Instruct-2512, Ministral-3-3B-Reasoning-2512, Ministral-3-8B-Instruct-2512, Ministral-3-14B-Instruct-2512, Mistral-Small-3.1-24B-Instruct-2503, Mistral-Small-3.2-24B-Instruct-2506, Magistral-Small-2509, Devstral-2-123B-Instruct-2512
gpt-ossOpenAIApache 2.0 with the gpt-oss usage policygpt-oss-20b, gpt-oss-120b, gpt-oss-safeguard-20b, gpt-oss-safeguard-120b
GLM-4.7-FlashZ.aiMITGLM-4.7-Flash
BielikSpeakLeash & ACK Cyfronet AGHApache 2.0Bielik-11B-v3.0-Instruct, Bielik-11B-v2.6-Instruct, Bielik-11B-v2.5-Instruct, Bielik-PL-11B-v3.0-Instruct, Bielik-Minitron-7B-v3.0-Instruct, Bielik-PL-Minitron-7B-v3.0-Instruct, Bielik-4.5B-v3.0-Instruct, Bielik-1.5B-v3.0-Instruct
BGE-M3BAAIMITbge-m3, bge-m3-unsupervised
WhisperOpenAIApache 2.0 · MIT (large-v3-turbo)whisper-large-v3-turbo, whisper-large-v3, whisper-large-v2, whisper-large, whisper-medium, whisper-small, whisper-base, whisper-tiny
Missing a model?

Ask for the one you need.

Tell us which model you need, on ARK Cloud or on your own GPUs, and what you will use it for. If it fits a supported family, it is a conversion away.

Request a model

Send the Hugging Face id, the deployment you have in mind and the GPUs you run. We reply with what it takes to bring it online.