/training
Adapt a model to your domain, outputs and tools. Choose how your team wants to work, define the evaluation, then select the checkpoint and method. Training access is by invitation. Custom engagements are quoted.
/paths
Work with our team
Agree on the task, baseline and evaluation. Work with us to choose the checkpoint, method and serving plan.
Scoped and quoted
What to bring
A task, sample data and a success criterion. Data readiness can be assessed together.
Keep your training loop
Your ML team owns the loop, loss or reward. Scope serverless or dedicated trainers and rollout inference with Grid.
Access by invitation
What to bring
Your framework, checkpoint, training method and evaluation. Add the capacity window for a dedicated run.
Use a built-in method
Bring prepared examples, preferences or a reward. Scope a managed SFT, DPO, ORPO or reinforcement fine-tuning job.
Access by invitation
What to bring
Dataset readiness, an eligible base model and a held-out evaluation. Checkpoint and method support are confirmed together.
Prepare before the job
Choose a path and a candidate, or leave the choice open. Describe the task, data and result your team needs to accept.
0 of 4 requirements drafted. Unanswered requirements are marked “To confirm” in the download.
Answers stay here while you explore the site during this visit. Refreshing clears them.
Continue in another visit
Save an editable brief to reopen these answers later. Files are read on this page. Submit an inquiry through the request form when you are ready.
Agree how the model earns its place.
Compare the untuned baseline with the trained artifact on held-out work. Review quality, exceptions, human effort and cost per accepted result.
Checkpoint and method eligibility, training costs and the serving release are confirmed in the proposal.
Training models · quoted jobs
Start with Qwen 27B adapters, or explore Inkling, Nemotron, GLM and Kimi for a custom job. Choose the model, agree on the evaluation, then confirm the method and compute.
14 training options · 5 model families · access by invitation
6 of 14 options · Highlights · eligibility confirmed per job
Qwen · BF16 weights
By invitationA starting point for domain language, structured answers and tool use.
Adapt the original post-trained checkpoint with your examples. Start with supervised fine-tuning and compare the result with the untuned model.
Published checkpoint and adapter training guide. Grid eligibility is confirmed per job.
Read the primary source for Qwen 3.8 27B · LoRA, opens in a new tabQwen · 4-bit training weights
Custom quoteAn adapter approach when training memory is a constraint.
Scope a custom job using Unsloth’s published 4-bit checkpoint. We confirm trainer compatibility, sequence length and the deployment format before the run.
Published QLoRA recipe and RL entry point. The Grid training run still needs validation.
Read the primary source for Qwen 3.8 27B · QLoRA, opens in a new tabInkling · BF16 weights
Custom quoteExplore multimodal adaptation for text, images and audio.
276B total parameters, 12B active. Scope the modalities, dataset and trainer before selecting adapters or a full-parameter job.
Open weights and vendor fine-tuning on Tinker. Grid compatibility is a separate review.
Read the primary source for Inkling Small, opens in a new tabNemotron · BF16 weights
Custom quoteConsider a smaller model for a focused assistant or agent.
A 4B hybrid Mamba and attention model. Scope support for its architecture, training method and target serving environment.
Published BF16 checkpoint. Method and trainer compatibility require scoping.
Read the primary source for Nemotron 3 Nano 4B, opens in a new tabGLM · BF16 weights
Custom quoteExplore long-horizon coding and verifiable agent tasks.
744B total parameters, 40B active. Published RL work uses SAO and slime. Scope distributed training, environments, verifiers and the checkpoint licence.
Version-specific post-training report. The SAO paper evaluates GLM 5.2; the 5.3 report describes its continued use.
Read the primary source for GLM 5.3, opens in a new tabKimi · MXFP4 weights
Custom quoteExplore long-horizon coding and native vision for a custom engagement.
2.8T total parameters, 104B active. Published weights use quantization-aware MXFP4. Confirm a viable training checkpoint, quantization support and the Kimi K3 licence before quoting a run.
Open-weight release with MXFP4. A standard BF16 or QLoRA training path has not been validated for Grid.
Read the primary source for Kimi K3, opens in a new tabTraining is by invitation. Published weights identify a candidate for scoping. The quote confirms checkpoint licensing, trainer compatibility, dataset volume, sequence length, epochs, method, evaluation and compute.
Inference prices do not price training. Training, storage and deployment are agreed separately. The model’s native context window does not promise a training sequence length.
Explore RL research and pilot requirementsReinforcement learning · research and scoping
There is published RL work for GLM 5.3 and Qwen 3.8. Match the evidence to the checkpoint, then design a pilot around a measurable task and a reliable reward. Sources reviewed 1 October 2026.
27B recipe · 2.4T research
The 2.4T results do not demonstrate a 27B training outcome. For 27B, confirm the recipe, trainer, rollout path and reward with a small pilot.
Unsloth documents a Qwen 3.8 27B RL entry point using its Qwen 3.5 training path. QwenGyre separately reports long-horizon agent RL on Qwen 3.8 2.4T.
Version-specific post-training report
The SAO paper evaluates GLM 5.2. The 5.3 release report describes its continued use. Scope the BF16 checkpoint, licence, distributed trainer and your task environments before a run.
Z.ai describes continued long-horizon RL with single-rollout asynchronous optimization (SAO), compaction, task verifiers and the slime framework. GLM 5.3 keeps the same base as 5.2.
Bring the rollout environment, reward function or verifier, dataset, untuned baseline and success criterion. Pilot eligibility and training cost are confirmed in a custom quote.
Read the evidence and pilot guide/api
The Training API is two trainers. Serverless for LoRA on shared capacity. Dedicated when you need full-parameter, distillation, or a cluster of your own. RL rollouts are the sampler, not the trainer.
Custom loop · LoRA · per-token
Write the loop in Python. Grid attaches to shared trainers. LoRA only. You pay for training tokens, not idle GPUs. Use this when a standard managed job is not enough and you do not need a dedicated cluster.
Custom loop · LoRA or full-parameter
Your trainers, held for the job. Custom loss, distillation, DPO, ORPO, and full-parameter updates run here. Time-based capacity, quoted. Tear it down when the run ends.
/methods
Supervised, preference, reinforcement, distillation and full-parameter training. Eligibility is per checkpoint and method. A model in the scoping list does not establish support for every method below. Confirm the trainer and evaluation in your training scope.
Managed or Training API
Supervised fine-tuning
Train on labeled examples of the output you want. Text and vision. The default first job. LoRA on managed. Full-parameter on the dedicated API.
Managed or Dedicated API
Direct preference optimization
Train on preferred and non-preferred pairs for the same prompt. DPO compares the policy to a reference model. Use it for tone, format, and answers that have no single gold string.
Managed or Dedicated API
Odds ratio preference optimization
The same preference pairs as DPO. ORPO folds a supervised objective in, so you do not hold a separate reference model. Same dataset shape: input, preferred_output, non_preferred_output.
Managed
Reinforcement fine-tuning
Train against a reward function, not only paired text. Grid runs the loop: rollouts, scoring, checkpoint, promote. Bring an evaluator. For teams that already have a trainer, use RL rollouts instead.
Dedicated API
Teacher to student
A custom loop that transfers behavior from a larger teacher into a smaller student. Not a managed preset. Quoted with the dedicated API.
Dedicated API
Update the base weights
LoRA is the fast start. Full-parameter is the ceiling when adapters cannot move the behavior. Managed jobs stay LoRA. Full-parameter is dedicated only.
Custom training
Work with our team to define the task, choose a compatible checkpoint and method, and agree the evaluation. The proposal defines the training run, output artifact, costs and serving plan. An accepted artifact receives account access through the agreed release process.
If the work is a whole workload on the factory, not one training job, that is a custom project. Talk to us there.
Keep your trainer
Point your own RL loop at Grid for sampling. We run the rollout inference. You keep the trainer. OpenAI-compatible. Reserved capacity, quoted. This is the sampler for the Training API path. It is not managed RFT, which runs the whole loop for you.
RL rolloutsWhat you train is what you serve
The proposal defines how an accepted checkpoint is deployed on Grid Serverless, Dedicated or Private. Account access, format and rollout are confirmed before release. Smart Router can include an enabled model you select. Training is quoted as jobs; inference token rates price serving.
Training docsAgree the checkpoint, data, evaluation and serving plan before the run. Training access is by invitation; custom engagements are quoted.