Fastino launches Tinker-compatible training API
Fastino has launched a hosted training API for fine-tuning and reinforcement learning on open-weight models, and it works with the Tinker SDK at per-token prices the company says are 25% to 85% below Tinker's.
Fastino announced the Fastino Training API, a low-level service that runs customer-written supervised fine-tuning (SFT) and reinforcement learning (RL) loops on hosted GPUs. It supports NVIDIA Nemotron 3.5 Super VL from launch, along with Nemotron 3.5 Lightning and Qwen3.6-35B-A3B.
Customers keep writing their own Python training loops and supply their own data, reward functions, environments and evals. The API handles sampling, gradient computation, weight updates and checkpoints. The trainer and sampler run in a single managed GPU container. SFT supports LoRA ranks from 1 to 32 with AdamW. For RL, the API offers grouped sampling with token log probabilities and training with importance sampling, PPO, CISPO or DRO.
According to Fastino, the API is compatible with the open-source Tinker SDK, cookbook, CLI and REST client. Existing users only need to change the server address and API key to switch.
Billing is per token rather than per GPU hour. Prices are per million tokens as of October 15, 2026:
- Nemotron 3.5 Super VL: $0.20 prefill, $0.72 sampling, $0.96 training. Fastino says this is 25% to 65% cheaper than Tinker.
- Nemotron 3.5 Lightning: $0.07, $0.35 and $0.30. Fastino says 29% to 73% cheaper.
- Qwen3.6-35B-A3B: $0.08, $0.45 and $0.45. Fastino says 63% to 89% cheaper.
Nemotron 3.5 Super VL is a 120B-parameter hybrid Mamba-Transformer mixture-of-experts model with 12B active parameters. Fastino fine-tunes it with a 75.2M-parameter LoRA adapter, which updates 0.06% of the model's weights. The company also released finance and healthcare adapters built on an early-access checkpoint using text-only inputs. It says they improved all 15 held-out benchmarks it measured, with full training runs taking between two and 5.5 hours. The excerpt does not name the benchmarks or give individual scores.