01
the workflow gap
It has never seen your schemas, CAD
parts, or your definition of correct.
Teach an open model your job. On your GPUs. You keep the weights.
Train Qwen or Llama on the tasks your business actually runs, in your cloud or on-prem. You keep a compact weights file: version it, test it, and serve it privately.
Free beta with Google sign-in at studio.g-ftech.com. The drawing on the right is the whole pitch. Everything else on this site is the proof, backed by raw data files and commit dates.
The problem
Six gaps between a frontier model and your workflow.
Prompting a commercial API works for drafts and demos. It breaks when the task defines your margin. None of this is about hallucination, and none of it is fixed by a bigger model.
02
the lab gap
SFT + RL pipelines, simulators,
verifiers, GPUs. A handful of labs.
03
the ownership gap
Rented API: your data goes out,
answers come back. Zero weights.
04
the model moves under you
Providers deprecate on their schedule.
Each migration is weeks you did not plan.
05
you cannot prove it got better
No frozen test set, no sign-off, so
nobody can approve it for production.
06
your data never reaches the model
Logs, tickets, CAD parts, git history
sit in storage. The API never sees them.
None of this is about hallucination, and none of it is fixed by a bigger model.
The solution
One run, four stages. Each leaves an artifact you keep.
Frontier APIs are good infrastructure, but a bad place to keep the workflow that defines your margin. Post-training bakes your data, edge cases, and definition of correct into a file you can version, test, and host.
Fine-tune (SFT)
Teach vocabulary, schemas, and tool syntax from your logs and git diffs. The base model stays untouched.
an adapter, ~100 MB
Reinforce (GRPO)
Improve weights through trial and error in simulated gyms. Winning attempts get rewarded.
better adapter + curveslatent reasoning: opt-in
Evaluate
Sealed test set frozen before training starts. Real proof comparing your model against base before production.
a scorecard, signed↑ the only stage that can say no
Deploy
Serve inside your own firewall on private GPUs. 1-click rollback, zero prompts or weights leave your network.
a private endpoint
BYOC
Every stage runs in your cloud account, on your GPUs.
The gym is sealed: an encrypted in-memory filesystem (/dev/shm), wiped when the run ends. Your data and your environment are unreadable to any operator, including us.
One gym, in full
What does an agent actually train against?
A gym is a flight simulator with an honest judge. The model gets real tools, tries a real task, and a verifier says pass or fail. That verdict is the only reward signal.
Dynamic balance & locomotion (Unitree G1 / Go1)
- Stack
- MuJoCo 3.3 + torque-limited joint controllers + contact physics
- Task
- Stabilize dynamic bipedal and quadrupedal gaits under lateral push disturbances and slippery terrain in MuJoCo. Joint torque limits, ground reaction forces, and zero-fall recovery gate reward.
- Runtime
- OPENENV / MUJOCO
- Tags
- #robotics #mujoco #humanoid-g1 #quadruped-go1
Changelog
What changed this month, with dates.
GFT Studio beta updated: run ledger, gym selection, and live reward curves at studio.g-ftech.com.
Verifier runtime isolation: encrypted in-memory /dev/shm sandboxes with zero host retention.
Platform architecture published: verifiable checkpoints, sealed baselines, and one stage table.
Bimanual stationery manipulation gym added (two Cartesian grippers, MuJoCo 3.3.7).
Qwen 3.8 27B on two H100s: 1,800 measured requests, 0 errors, about $15 of compute for the whole sweep.
Distributed GRPO pipeline: asynchronous rollout workers decoupled from actor weights update.
Engineering notes
Longer reads
- Teaching a 27B Model to Write Trading Alphas: 101 Formulas, 12 Rewards and One Unseen Year16 min read
- Multi-Reward RL, Part 3: GDPO + CISPO + REPO-R at 27B, and the Advantage Floor That Stopped Learning12 min read
- Latent-GRPO on 8× B300, Part 2: Two Silent Adapter Bugs, Parity Gates That Learned, and Rollouts on All Eight GPUs13 min read
- From 36 Minutes to 2.6: Making a Latent-GRPO Training Step on Qwen 3.8 27B 14× Faster16 min read
g factor technologies
We work with a few teams at a time.
g factor technologies inc. is a Delaware C-Corp building infrastructure for specialized AI training. We help engineering teams build custom simulators, design rewards that models can't game, and deploy private checkpoints on their own hardware. And we publish our raw benchmarks along the way.
Built by two world-championship-level competitors, one benchmark author, and an ex-Apple platform lead (ex-Apple, Zendesk, Alibaba, Nebius). All full-time.
Request demo
Show us the workflow you want your agent to master.
A short note on the model, the environment and how you would judge success is enough. We reply from a person, not a sequence.