Supported Models

MinT serves base models from two lineup pools, separated by access plan, plus a broader technically-compatible set you can request access to:

Community Lineup — available on the shared hosted endpoint at mint.macaron.xin via API-key request.
Enterprise-only Lineup — available via Enterprise plan capacity reservation on a dedicated cluster.

Community Lineup

These are the Qwen3 base models with provisioned capacity on mint.macaron.xin and test coverage in mint-quickstart. The list mirrors the Qwen entries returned by service_client.get_server_capabilities().supported_models from a live preflight. Use one of these for a smoke run if you have no preference.

Qwen3 base models

Variant	Training Type	Architecture	Size	Context	Default	Tested algorithms
`Qwen/Qwen3-0.6B`	Hybrid	Dense	Tiny	32k	yes	SFT, GRPO
`Qwen/Qwen3-4B-Instruct-2507`	Instruction	Dense	Compact	32k	cookbook	SFT, DPO, GRPO
`Qwen/Qwen3-4B-Thinking-2507`	Reasoning	Dense	Compact	32k	no	SFT, GRPO
`Qwen/Qwen3-30B-A3B-Instruct-2507`	Instruction	MoE	Medium	32k	optional	SFT, GRPO
`Qwen/Qwen3-235B-A22B-Instruct-2507`	Instruction	MoE	Large	32k	no	SFT, GRPO

Per-model notes

Qwen3-0.6B — lightweight default. Quickstart, custom_reward, custom_loss, and sampling_log all run on this.
Qwen3-4B-Instruct-2507 — base model for the four maintained cookbook recipes (dapo-aime, chat-dpo, fingpt, lawbench).
Qwen3-4B-Thinking-2507 — reasoning / chain-of-thought variant of the 4B model.
Qwen3-30B-A3B-Instruct-2507 — mid-scale instruction following.
Qwen3-235B-A22B-Instruct-2507 — large-scale instruction tuning. Volcano A800 cluster: inference_tp=16, train_tp=4, train_pp=1, train_ep=8.

Column legend

Training Type — Hybrid supports both thinking and non-thinking modes; Instruction is chat-tuned without chain-of-thought; Reasoning always emits chain-of-thought before the visible response.
Architecture — Dense activates all parameters per token; MoE is a sparse mixture-of-experts.
Size — total parameters, not active. Tiny < 1B; Compact 1B–4B; Medium 30B–32B; Large 70B+.
Context — native context window without YaRN extension.
Default — yes is the quickstart default; cookbook is the default in maintained cookbook recipes; optional means scripts accept it via MINT_BASE_MODEL; no means request-only.

Override the default by setting MINT_BASE_MODEL before running any quickstart script:

export MINT_BASE_MODEL=Qwen/Qwen3-30B-A3B-Instruct-2507
python quickstart/quickstart.py

Embodied / VLA

Variant	Tested algorithms	Notes
`mintx.OPENPI_FAST_MODEL` (constant)	VLA via SDK / HTTP	Embodied-agent track. See VLA.

Enterprise-only Lineup

These models require an Enterprise plan. Capacity is provisioned per customer on a dedicated cluster — they are not available on the Community shared endpoint. Canonical Hugging Face IDs are confirmed at the time of capacity provisioning.

Family	Variant	Tested algorithms	Notes
GLM	GLM-5	SFT, RL	Zhipu GLM-5 family. Provisioned on customer cluster on request.
GLM	GLM-5.1	SFT, RL	Successor to GLM-5; capacity reserved per customer.
Kimi	Kimi-K2	SFT, RL	Moonshot Kimi-K2. Long-context workloads.
Kimi	Kimi-K2.5	SFT, RL	Successor to Kimi-K2.
DeepSeek	DeepSeek-V3	SFT, RL	DeepSeek V3 base model.

To reserve capacity for any of these, email sales@mindlab.ltd or Schedule a Demo and mention the model and workload class.

Technically Compatible

The MinT server endpoint accepts any HuggingFace-hub-style model string that maps to one of the supported transformer architecture families below. Lineup is the explicitly-tested subset. Other families known to work in principle:

Qwen series (Qwen2.5, Qwen3 — Instruct, Thinking, Coder variants)
Llama 3.x family
Gemma 2.x and 3.x
DeepSeek family

There is no in-repo evidence of these being run end-to-end on MinT today. If you need one of them on the lineup with provisioned capacity, request access (below).

Request a Model

If a model you need is not listed:

Email sales@mindlab.ltd with the model identifier and your intended workload (SFT / DPO / RL, batch size, expected weeks of usage).
Or open an issue at the public mint-quickstart repo: github.com/MindLab-Research/mint-quickstart/issues.

VLM (Vision-Language Model) base models are tracked separately as a server capability — see the VLM page for the current state.

How to know what your endpoint actually serves. A successful preflight returns a capabilities.supported_models list. The default quickstart.py prints Auth preflight: OK (N supported models) — call service_client.get_server_capabilities() from your own script to enumerate them.