The weights are yours
Every checkpoint lands in your workspace as standard files. Serve them anywhere, and keep them if you leave.
Model training by Maincode, the team behind MatildaBy Maincode, the team behind Matilda
Train models you own.
Keep the data in Australia.
Ask Matilda Code, or any MCP-capable coding agent or IDE, to run the training. We run it on AMD Instinct MI355X and NVIDIA B300 GPUs hosted in Australia, bill by the GPU-hour, and put the weights in your workspace.
- A$9.35 per GPU-hour, including GST
- Weights are yours
- Hosted in Australia
- One MCP endpoint
matilda-trainMCP serverconnected
›Fine-tune Qwen 2.5 on our support tickets so it replies like our team. One GPU, under $5.
research_write_textsupport_tickets.jsonl
research_quote1 × MI355X · 15 min · A$2.34 max
research_submitbluegum-support-lora
└ run 7c1e4b · RUNNING
research_metricsmetrics.jsonl
0%epoch 1/8loss 2.41
└ run 7c1e4b · COMPLETED
research_artifactsadapter/
└ 4.4 MB adapter → weights/
Done in 2m 31s for A$0.39 including GST. Loss 2.41 → 0.27. The tickets stayed in Australia, and the adapter is yours to keep.
As used by
Fine-tuning that ends with a model you keep
Other clouds keep your tuned model and charge you to serve it. Here you pay for the training run, and everything it produces is yours.
Training stays in Australia
Each run is isolated on Australian infrastructure that Maincode operates.
- Region
- Australia
- Outbound internet
- Blocked
- File access
- Your account only
A$9.35per GPU-hour, including GST
Priced per run, not per token
Data transfer and a data scientist included. A finished model costs nothing to keep.
Run it from your coding agent
One MCP endpoint for Matilda Code or any MCP client. SSH, Slurm and a REST API too.
From a prompt to your own weights, in three steps
No console to learn and no cluster to run. If you can ask your coding agent for a change, you can ask it for a model.
- Step 01
Connect your agent
Add one remote MCP server to Matilda Code, or any MCP-capable coding agent or IDE, then approve the connection in your browser. Any client that supports remote MCP over HTTP with OAuth works.
https://train.matilda.com/mcp/research - Step 02
Ask for the model you want
Describe the run in plain language. Your agent stages the data, checks the request against your limits with research_validate, then submits a durable job that keeps running if you close the laptop.
Fine-tune Qwen3-8B on data/tickets.jsonl. Hold out 5% for eval, use 8 GPUs, and stop when eval loss stops improving. - Step 03
Take the weights
Checkpoints and adapters land in your workspace. Download them through your agent, SSH or the HTTP API, and serve them anywhere you like.
runs/4f19a2/ ├─ model/ │ ├─ config.json │ └─ model-00001-of-00004.safetensors ├─ metrics.jsonl └─ run-record.json
One rate per GPU-hour. Everything else is included.
Pay by the second on demand, or reserve whole 8-GPU nodes for less, with the reserved term paid in advance.
| Booking | Per GPU-hour | Per node, monthly | Reserved term total | Good for |
|---|---|---|---|---|
| On demand | A$9.35 incl GST A$8.50 ex GST | A$54,604.00 | – | Experiments and bursts, billed per second |
| 1 month | A$9.075 incl GST A$8.25 ex GST | A$52,998.00 | A$52,998.00 | A single training run |
| 3 months | A$8.58 incl GST A$7.80 ex GST | A$50,107.20 | A$150,321.60 | A project |
| 6 months | A$8.25 incl GST A$7.50 ex GST | A$48,180.00 | A$289,080.00 | A research program |
| 12 months | A$7.48 incl GST A$6.80 ex GST | A$43,683.20 | A$524,198.40 | Ongoing training |
| 24 months | A$6.545 incl GST A$5.95 ex GST | A$38,222.80 | A$917,347.20 | A model roadmap |
| 36 months | A$4.95 incl GST A$4.50 ex GST | A$28,908.00 | A$1,040,688.00 | The lowest rate |
| Preemptible | A$4.675 incl GST A$4.25 ex GST | – | – | Jobs that checkpoint and can be requeued |
Every booking includes
- The MCP endpoint for your coding agents
- Managed Slurm, SSH and a REST API
- Data transfer in and out
- A persistent workspace for code, data and checkpoints
- PyTorch, Transformers, PEFT, Accelerate and Datasets, ready to go
- A Maincode data scientist to port and tune your first runs
CPU-only allocations: A$0.066 per vCPU-hour including GST (A$0.06 ex GST), with up to 4 GiB RAM per vCPU included. Billed per second of allocated time. GPU allocations already include their CPU and memory.
NVIDIA B300: pricing on request. Tell us the run you have in mind and we’ll quote it.
One 70B fine-tune, priced in Australia
The GPU-hour price is half the story.
The other half is how long the run takes.
Vendor-submitted MLPerf Training results, times each provider’s on-demand GPU-hour price in its Australian region. Your workload will differ.
2.7× faster
than the best published 8× H100 result: 10.18 against 27.54 minutes
4× cheaper
per run than AWS Sydney on demand: A$11.54 against A$47.06, excluding GST
How it compares with hyperscaler fine-tuning
What an Australian team gets from each provider’s managed fine-tuning today, side by side.
Feature
Matilda TrainHosted in AustraliaAmazon BedrockModel customisation
Matilda TrainHosted in AustraliaAmazon BedrockModel customisationWhere training runsIn AustraliaUS regions only: us-east-1 for Nova, us-west-2 for Llama, Claude 3 Haiku, Qwen3 and gpt-oss
Models you can trainAny open-weight model your code can load, with LoRA or full-parameter trainingA fixed list: Nova, Llama up to 3.3, Claude 3 Haiku, Qwen3 32B, gpt-oss-20B and Titan
Download the weightsYes, every checkpointNo documented way to download a customised model
Keeping the model onlineNothing to pay. It’s a file you own.Most custom models need Provisioned Throughput: US$13.08 to US$24.00 per model unit-hour for a custom Llama 3.1 70B
Newest training GPUs in AustraliaAMD MI355X and NVIDIA B300H100 on EC2 in Sydney; H200 only through Capacity Blocks
Coding-agent workflowOne MCP endpoint, with runs in AustraliaAWS’s SageMaker agent workflow (May 2026) runs in us-east-1, us-west-2, eu-west-1 and ap-northeast-1
Help porting your codeA Maincode data scientist, includedPaid AWS Support plans
When to stay put. If you want a customised Nova or Claude 3 Haiku and will only ever serve it inside Bedrock, use Bedrock’s customisation. Choose Matilda Train when you want to own the model, keep training in Australia, or train something their menus don’t offer.
FeatureMatilda TrainHosted in AustraliaGoogle CloudGemini Enterprise Agent Platform, formerly Vertex AI
Matilda TrainHosted in AustraliaGoogle CloudGemini Enterprise Agent Platform, formerly Vertex AIWhere training runsIn AustraliaGemini: us-central1 or europe-west4. Open models: US, EU, Singapore or Jakarta
Where the tuned model is servedWherever you choose to run itTuned Gemini is served from US or EU multi-region endpoints
Download the weightsYes, every checkpointNot for tuned Gemini. Tuned open models, yes
When the base model retiresYour weights keep workingTuned models can’t move to a newer Gemini, and stop working when their base model retires
Keeping the model onlineNothing to pay. It’s a file you own.Tuned Gemini 3 costs 1.5× the base model per token. Tuned open models need an hourly endpoint
Newest training GPUs in AustraliaAMD MI355X and NVIDIA B300H100, in one zone each in Sydney and Melbourne; no TPUs
Help porting your codeA Maincode data scientist, includedPaid Customer Care plans
When to stay put. If you need a tuned Gemini and will only ever serve it on Google Cloud, use Google’s tuning. Choose Matilda Train when you want to own the model, keep training in Australia, or train something their menus don’t offer.
FeatureMatilda TrainHosted in AustraliaMicrosoft FoundryFormerly Azure AI Foundry
Matilda TrainHosted in AustraliaMicrosoft FoundryFormerly Azure AI FoundryWhere training runsIn AustraliaFrom Australia East, Global or Developer training. Microsoft says neither guarantees data residency; regional training is US and EU only
Models you can trainAny open-weight model your code can load, with LoRA or full-parameter trainingGPT-4.1 family, o4-mini (reinforcement fine-tuning) and a short list of open models such as Llama 3.3 70B, Qwen3-32B and gpt-oss-20b
Download the weightsYes, every checkpointNo documented way to download a fine-tuned OpenAI model
Keeping the model onlineNothing to pay. It’s a file you own.US$1.70 an hour per deployed fine-tune, with or without traffic: about US$1,241 a month
Newest training GPUs in AustraliaAMD MI355X and NVIDIA B300GB200, H200, H100 and MI300X VMs; H100 lists at A$25.39 per GPU-hour
How you pay for trainingPer GPU-hour, A$9.35 including GST on demand, data transfer includedPer training token, or per hour for o4-mini reinforcement fine-tuning
Help porting your codeA Maincode data scientist, includedPaid Azure support plans
When to stay put. If you need a fine-tuned GPT-4.1 or o4-mini and will only ever serve it on Azure, use Foundry. Choose Matilda Train when you want to own the model, keep training in Australia, or train something their menus don’t offer.
Agents train within limits you set
Letting an agent near GPUs is a fair worry. Matilda Train enforces the rules on the server, so they hold whichever client connects.
Checked before it runs
Every request is validated against the GPU, time and resource limits your administrators set.
Scoped and revocable
Grant read, submit and file access separately, and disconnect any client from your account at any time.
No duplicate jobs
An agent that retries can’t launch the same run twice.
Isolated runs
Each allocation has its own identity, filesystem and network isolation, with no outbound internet.
A record of every run
The script digest, parameters and input checksums are captured at submission, so you can show how a model was made.
| research:read | Your profile, run state, logs, metrics and artifacts |
|---|---|
| jobs:write | Submit, cancel and resume your runs |
| files:read | Read and download your workspace files |
| files:write | Create files, replace or delete reviewed files, and upload datasets |
Current-generation GPUs, hosted in Australia
Neither of these GPUs is offered in AWS, Google Cloud or Azure’s Australian regions. Both carry 288 GB of memory per GPU, 3.6× an 80 GB H100, so bigger models and longer contexts fit on fewer GPUs.
Instinct MI355X
AMD288 GBHBM3E per GPU
- Per node
- 8 GPUs, 2.3 TB of GPU memory
- Software
- PyTorch on ROCm
- MLPerf
- Llama 2 70B fine-tune in 10.18 min on 8 GPUs
- Price
- A$9.35 per GPU-hour on demand including GST, A$4.95 on a 36-month term
B300
NVIDIA288 GBHBM3E per GPU
- Per node
- 8 GPUs, 2.3 TB of GPU memory
- Software
- The CUDA stack, including FP4 training
- Good for
- CUDA-only code and FP4 training
- Price
- On request
Bring your stack
Your training scripts come with you. On MI355X, a Maincode data scientist helps port anything CUDA-specific in your first run.
- Train
- PyTorchTransformersTRLPEFTAccelerateAxolotlUnslothLLaMA-Factorytorchtitan
- RL and kernels
- verlFlashAttentionLiger Kernelbitsandbytes
- Evaluate and serve
- vLLMSGLang
- Models
- QwenLlamaGemmagpt-ossMistralDeepSeek
Built by the team that trains Matilda
Maincode built this platform to train its own models. The scheduler, the isolation, the run records and the agent tools are the ones our researchers use every day, so the problems that slow training down have already been dealt with.
When you start, a Maincode data scientist works with your team to port your code and tune the first runs. You work with people who train models for a living.
- Company
- Maincode, an Australian AI company and the maker of Matilda
- Where your data lives
- Australian infrastructure that Maincode operates
- Who can see your files
- Not other users. Being in the same organisation doesn’t give anyone access to your runs or files.
- Support
- A data scientist from day one, included in the price
A few useful answers.
More detail, examples and troubleshooting are in the researcher documentation.
Do I need to know Slurm or run a cluster?
No. Your coding agent handles submission, monitoring and results through the MCP endpoint, and Maincode runs the hardware and scheduler. If you already know HPC, native Slurm is there too. See the job guide.
Which models can I train?
Any open-weight model your code can load, including the Qwen, Llama, Gemma, gpt-oss, Mistral and DeepSeek families. Use LoRA or QLoRA for quick adaptation, or full-parameter training when you need it. With 288 GB per GPU, larger models fit on fewer GPUs.
Do I own the model?
Yes. Checkpoints, adapters and merged weights are written to your workspace as standard files. Download them whenever you like and serve them anywhere. There’s no fee to keep them. The workspace guide covers downloads.
Where does my data go?
It stays on Australian infrastructure operated by Maincode. Each run is isolated, with its own identity and filesystem and no outbound internet, so training data can’t be sent elsewhere during a run. Other users, including people in your organisation, can’t see your files.
How is billing worked out?
Per GPU-hour. On-demand use is billed per second at A$9.35 including GST (A$8.50 excluding GST), reserved nodes cost less, and preemptible capacity is A$4.675 including GST (A$4.25 excluding GST) for jobs that checkpoint. CPU-only allocations cost A$0.066 per vCPU-hour including GST (A$0.06 excluding GST), with up to 4 GiB RAM per vCPU included, billed per allocated second; GPU allocations include their CPU and memory. Prices are in Australian dollars. Transfer and support are included, subject to fair-use limits.
AMD or NVIDIA: does my code need to change?
Most PyTorch and Hugging Face training code runs on MI355X as it is. Code that depends on CUDA-only kernels can run on B300 instead, or a Maincode data scientist can help port it when you start.
How is this different from Bedrock, Google Cloud or Foundry fine-tuning?
Those managed services tune a fixed list of models, mostly in US or EU regions, and document no way to download a tuned Gemini, GPT-4.1 or Bedrock custom model. Matilda Train trains any open model in Australia and gives you the weights. If you specifically need a tuned Gemini or GPT-4.1, use their service. Compare each provider.
Which coding agents work?
Any client that supports remote MCP over HTTP with OAuth, including Matilda Code. You can start a client with read-only access and approve more permissions when a task needs them. The connection guide walks through it.
Can I use SSH and a desktop editor?
Yes. Start an interactive allocation, register your public SSH key, and download its SSH configuration. Connect using OpenSSH and the open-source wstunnel client on macOS, Windows or Linux. The SSH setup guide walks through it.
Bring your training.
Own the model.
Create your account, add compute credits and connect your coding agent. You pay by the GPU-hour, and every checkpoint is yours to download.
Amazon Bedrock, Google Cloud, Gemini Enterprise Agent Platform, Vertex AI, Gemini, Microsoft Foundry, Azure and other product names are trademarks of their owners, used here to identify and compare.