Training
Explore Qwen 27B adapters and custom training for Inkling, Nemotron, GLM and Kimi. Models, methods and compute are scoped by invitation.
Fine-tuning adapts a base model to your task. Agree the training scope and evaluation before the run. An accepted artifact can join your account catalog after its serving format, access and release are confirmed. You can then opt into Smart Router for the enabled model.
Training runs as jobs, not as inference requests. Batch is not the way to do it. The meter is separate, and it is not settled on the billing page yet.
Access is by invitation. Request access on Training or explore the training models. Training receives a separate quote from inference pricing.
Prepare a training brief#
Use the training planning brief to select a requested path and a candidate checkpoint, then describe the task, data readiness, model or method notes and evaluation criterion. You can leave the path or checkpoint open for discussion. A selected checkpoint remains a scoping candidate.
Take the downloaded brief to your team or carry the drafted requirements into a qualified training request. Shared links preserve the selected path and candidate; written requirements stay on the current page. An inquiry prepares a discussion and does not start a job. Keep datasets and credentials in the agreed evaluation environment.
Models for scoping#
The training list covers five families. Search by model, checkpoint or workload, filter by family, then open a model-specific inquiry. Published weights establish a candidate for review. They do not establish an enabled Grid trainer or support for every method.
| Family | Models | Checkpoint and method review |
|---|---|---|
| Qwen | Qwen 3.8 27B, LoRA and QLoRA | Original BF16 and published Unsloth 4-bit checkpoints. See the adapter paths below. |
| Inkling | Inkling, Inkling Small | Inkling and Inkling Small BF16 weights. Custom multimodal training, quoted after compatibility review. Vendor fine-tuning on Tinker does not enable a Grid trainer. |
| NVIDIA Nemotron | Nano 4B and 30B A3B, 3.5 Lightning, Super 120B A12B, Ultra 550B A55B, Nano Omni, 3.5 Content Safety | Published BF16 checkpoints linked from each training card. Scope the architecture, licence, method and modalities separately for each checkpoint. |
| GLM | GLM 5.3, GLM 5.3 Flash | GLM 5.3 BF16 and Flash BF16. Distributed post-training by custom scope. The 5.3 release has RL evidence; Flash requires its own method review. |
| Kimi | Kimi K3 | Published MXFP4 weights. Confirm a viable training checkpoint, quantization support and the Kimi K3 licence. A standard BF16 or QLoRA path has not been validated for Grid. |
The inquiry preserves the selected checkpoint and approach. Add your task, dataset size, format and modalities, success criterion, evaluation examples, timing and budget range. Custom jobs also ask which method you want to explore. The quote confirms licensing, trainer compatibility, dataset volume, sequence length, epochs, compute, evaluation, storage and deployment.
For reward-based training, read the RL evidence and pilot guide. It distinguishes the Qwen 27B implementation guide, Qwen 2.4T research, and GLM 5.3’s version-specific post-training report.
Qwen 3.8 27B#
Qwen 3.8 27B is the smaller Qwen option for specializing on your own examples. It is a dense, native vision-language model under Apache 2.0, with a native context of 262,144 tokens. The checkpoint and training sequence length are agreed for each job; the model's maximum context is not a promised training length.
| Approach | Published checkpoint | Grid path |
|---|---|---|
| LoRA on BF16 weights | Qwen/Qwen3.8-27B | Training inquiry by invitation. Start with SFT and evaluate against the untuned model. |
| QLoRA on 4-bit weights | unsloth/Qwen3.8-27B-unsloth-bnb-4bit | Custom job, quoted after trainer compatibility and deployment scoping. |
The original Qwen release is a post-trained checkpoint, which can be adapted further. It is not a separate pretraining-base release. Unsloth's Qwen 3.8 training guide documents the 4-bit QLoRA approach. A published training recipe does not establish an enabled Grid trainer for your account.
Qwen/Qwen3.8-27B-FP8 is a separate quantized serving format. Evaluate an adapted checkpoint before deciding the serving format and promoting the artifact.
The inquiry carries your selected checkpoint and adapter approach. Add the task, dataset size and format, success criterion, evaluation examples and timing. The quote confirms dataset volume, sequence length, epochs, compute, evaluation, storage and deployment. Inference input, cache and output rates do not price training tokens or a training job.
Three paths#
| Path | Who it is for | What you do |
|---|---|---|
| Custom training | You need ML expertise | Co-design or fully custom-build. We sit with your team. Quoted. |
| Training API | You have an ML team | Write the loop in your environment. Connect to Grid trainers and RL rollouts. Invitation. |
| Managed training | You have ML engineers | Built-in SFT, DPO, or ORPO jobs on your datasets. Invitation. |
| Scoped Agent RL | You have multi-step tool workflows | Trajectory reinforcement learning against deterministic verifiers. See Agent RL. Scoped by invitation. |
Training API shapes#
| Trainer | What it is for | Mode |
|---|---|---|
| Serverless training API | You write the loop. Shared trainers. | LoRA. Per-token. |
| Dedicated training API | You write the loop. Your trainers. | LoRA or full-parameter. Quoted. |
Pick managed when a standard method is enough. Pick the serverless API when you need a custom LoRA loop and do not want idle GPUs. Pick dedicated when you need full-parameter updates, distillation, or a custom loss. Pick custom training when you want us to co-design or run the job.
Methods#
SFT. Supervised fine-tuning on labeled examples. Text and vision. The usual first job.
DPO. Direct preference optimization. Preferred and non-preferred pairs. Compares the policy to a reference model.
ORPO. Odds ratio preference optimization. The same pairs as DPO, with a supervised term, so you do not hold a separate reference.
Agent RL / RFT. Reinforcement fine-tuning on multi-step trajectories against deterministic verifiers and rubrics. Scoped by invitation on Agent RL. If you already have a trainer, use RL rollouts for sampling only.
Distillation. Teacher to student, as a custom dedicated loop. Not a managed preset.
Full-parameter. Update the base weights. Dedicated API only. Managed jobs stay LoRA.
Eligibility is per checkpoint and per method. Qwen 3.8 27B has the adapter options described above. The other model families are custom scoping candidates. Their presence in the training list does not promise managed SFT, DPO, ORPO, RFT, adapters or full-parameter support. The adapted artifact receives its own catalog id when it is accepted and enabled for your account.
Data#
Managed jobs take OpenAI-compatible chat JSONL: messages with role and content. Preference jobs add preferred_output and non_preferred_output. One-turn pairs for DPO and ORPO.
Custom training#
If you need ML expertise to co-design or fully custom-build the model, Forward Intelligence sits with your team. Quoted. See custom training.
If the work is a whole workload on the factory, not one training job, that is a custom project.
What this page does not claim#
- A public rate or a free tier.
- Self-serve GA.
- That Grid inference already runs on the Brazil or US sites in development. Training jobs today run on the factory we operate on AWS.