A lot of "expensive" fine-tunes are just a run that died at hour 2 while the pod kept billing till morning 💀
On Tensorant every run gets a time limit and an hourly price cap, and the GPU shuts down the moment it ends, fails or gets cancelled.
Hard caps should be the default.
IBM's CEO says open-weight models now do 90% of the job at 10% of the cost.
The training bill should follow the same math. On Tensorant your GPUs run in your own RunPod, Lambda or AWS account at their price, with no markup from us.
Most bad fine-tunes I've seen trace back to the dataset, not the training config.
On Tensorant every generated example cites the passage it came from. Missing evidence, repeats and personal data get flagged, and you approve each one beside its source before a GPU starts.
GRPO is only as good as the checker you give it.
That's why RL on Tensorant uses rule-based rewards you can actually read, and the tuned model gets scored against the base on the same held-out set before you trust it.
A vague reward just teaches the model to game it.
A lot of fine-tuned models get a handful of calls a day. Keeping a GPU on 24/7 for that is how a small part of me dies.
On Tensorant you can publish a model as wake-on-request. It costs nothing while asleep and wakes up when a call comes in.
Same OpenAI-compatible API.
BOOM! I trained my own judge on the Jev datasets. It beat Jev 1.14 🔥
Open 4B, 22 min on one NVIDIA H200 on Hugging Face Jobs, $1.87
4,035 held-out RAG + safety calls:
Jev 1.14: 82.2%
my 4B: 89.4%
Twist: on PubMed answers it never saw, Jev wins. Unless it learns from Jev 👀
Every fine-tuning platform should let you walk away with your weights.
On Tensorant the adapter downloads as a standard LoRA file that loads with PEFT, or goes to your own Hugging Face account, private by default.
If you can't export it, you're renting it.
A lot of "expensive fine-tunes" are just a run that crashed at hour 2 while the pod kept billing till morning.
On Tensorant the GPU shuts down the moment a run ends, fails or gets cancelled. Every run also gets a time limit and an hourly price cap.
Hard spending caps should be the default.
We are releasing Ultramarine 1 Flash, a frontier-class model built for knowledge work. It is open-source and achieves similar performance to Fable 5, while being 60x cheaper.
It was post-trained by a team of 5 for ~$300k all-in. Available on API now, with weights releasing in 2
Sometimes the model already knows the answer, it just says it wrong. too long, too rude, wrong format.
That's where preference training helps. Same prompt, preferred vs rejected reply.
Tensorant now supports DPO.
tensorant.io
972 Followers 2K FollowingData Scientist | ML Engineer | Data Analyst . Port Harcourt, Nigeria. Building fintech AI: fraud detection, NLP, churn. Open to remote/permanent roles.
370 Followers 960 FollowingFounder of ToanKhonTech. I build websites, apps and software. Sharing AI experiments, open-source tools and lessons from real projects.
98 Followers 842 FollowingSelf-taught AI builder. Made ClaimCheck: a public eval of whether agents' "tests pass" is true, plus a fine-tune pilot. Austin
9K Followers 5K FollowingFounder & CEO of AI Research & Engineering Lab | Weekly @twimlai GenAI meetup | @Cohere_Labs Community Lead & Researcher | Prev @TCS @Infobeans | No dupes ever
296 Followers 186 FollowingWant to help build a more regenerative and ethical AI future?
@openbmb ambassador
Projects:
https://t.co/tFb9aFmUFQ
https://t.co/SU1BqYzSIC
86 Followers 265 FollowingSenior Tech Lead @HCLTech | Testing AI coding agents, harnesses and useful GitHub repos | Practical AI engineering without the hype | Educator
799 Followers 381 FollowingI talk to LLMs for a living. CTO @ Scribie AI Formerly Built Intro AI , FormerCTO @ LongShot AI. @manutd Part time shitposting on X
2K Followers 353 FollowingCS PhD @Stanford, building systems for llm data processing, including LOTUS (https://t.co/ZfeKaeFdnf), ACORN and DeepScholar (https://t.co/iJvm1aMdNM)
20K Followers 349 Followingthe on-device clinical AI guy | building @OpenMed_AI | making open models @arcee_ai | AI infra & research @CNRS | follow for local clinical AI demos
1.6M Followers 1K FollowingTechnology’s daily show. Hosted by @johncoogan & @jordihays. Streaming live 11a–2p PT weekdays. Sign up for our free newsletter at https://t.co/Nhf5ohjInO.
1.7M Followers 2 FollowingGet the latest from Hacker News! Top 5 stories on the hour, every hour. By @riklomas from @superhi_ – unofficial and not affiliated with Y Combinator