Spotlight: Forgelet — a model factory for tasks, not chatbots
Most teams that want a custom model end up with two bad options: call a hosted LLM API and pay per token forever, or hire someone who knows what a training run actually is. Forgelet is built for the gap in between. You describe a task (function calling, classification, or pulling structured fields out of messy text), optionally hand it a dataset or let it generate one, and it fine-tunes a small model on managed GPUs. What comes back is a file, in GGUF, LiteRT, or CoreML, that you run wherever your app already lives: phone, server, Raspberry Pi.
The interesting decision here is what they chose not to build. This isn't a general-purpose model or a hosted inference API. It's scoped tightly to tasks that don't need a giant model to begin with: routing a support ticket, extracting a name and a date from a paragraph, matching a user's phrase to the right function call. That's the actual insight behind Forgelet's pitch: a small model tuned tightly on one job can match or beat a much bigger general model at that job, and it doesn't need a network call or a token bill to run. Once it's on the device, there's no per-request pricing and no data leaving the phone to explain in a privacy policy.
The pipeline they're replacing is laid out plainly on their own site: provisioning GPUs, formatting JSONL datasets, configuring training runs, debugging exports, validating on target hardware. That's a real job today, often a multi-week one even for people who know what they're doing. Forgelet's bet is that most developers don't want to become ML engineers to get a model that does one specific thing well. They want to describe the job, get a file, and ship it.
Who should try this: app developers with a narrow, well-defined task who currently either pay per-token to a cloud model for something that doesn't need general intelligence, or who have been putting off fine-tuning because the tooling is a swamp. If you're building offline-first mobile apps, embedded devices, or anything where sending user data to a third-party API is a liability rather than a convenience, this is worth a look. The three use cases they show, function calling, classification/routing, and structured extraction, are all things a small tuned model handles fine without needing frontier-model reasoning.
Who should skip it: anyone whose task is open-ended (general chat, creative writing, broad reasoning across unpredictable inputs) is not the target here, and the site doesn't claim otherwise. If your task changes shape often, retraining a downloaded model file every time you tweak requirements is a different workflow than hitting an API with a new prompt, and that tradeoff needs to be weighed honestly. There's also no pricing published yet beyond the framing that running the model is free once trained, and no public detail on how the actual fine-tuning step is priced or gated. Right now Forgelet is in early access, onboarding in waves, so there's not yet a track record of exported models performing at scale in production, and I'd treat any specific accuracy claims as something to verify on your own task before committing to it.
Where this goes next, as I read it: the wedge (define a task, skip the ML pipeline, own the output file) is a genuinely different pitch than most AI tooling right now, which is mostly about renting bigger and bigger models. If Forgelet nails the data-generation step, meaning it can synthesize decent training data from just a task description for teams with no dataset at all, that's the part that would separate it from someone spinning up their own fine-tuning job on a rented GPU. That's also the hardest part to get right, and the part I'd want to see evidence on before betting a production feature on it.
Try Forgelet: forgelet.tech
See the launch: Forgelet on welaunch.sh