Oumi
Verified account This account is verified because it's an official organization on X. Learn more@oumi_ai
Let’s build better AI - open is the path forward
17 Following 683 Followers 147 Posts
Latest posts 5
Building an agent is easy. Training the model behind it is the hard part.
Today we’re launching Oumi’s new agent workflow. FAiE connects evaluation, data, training, deployment, and production feedback.
Build in Oumi or bring your own harness.
x.com/i/broadcasts/2…
Today we’re launching Oumi’s new agent workflow. FAiE connects evaluation, data, training, deployment, and production feedback.
Build in Oumi or bring your own harness.
x.com/i/broadcasts/2…
Oumi Fall Launch: AI to Build Agents and Models oumiai.substack.com/p/oumi-fall-la…
oumiai.substack.com Oumi Fall Launch: AI to Build Agents and Models Building an agent is easy. Making it reliable is hard. If every company rents the same frontier model, who gets the compound interest on the AI? For #OpenSourceAIWeek , @raffi (CTO, @mozilla) and @Koukoumidis (CEO, @oumi_ai) tackled that in AI Masterminds.
The questions they tackled:
• What happens when your provider changes pricing or access?
• Cost per token, or cost per completed task?
• Is the model the product, or the harness around it?
What they agreed on: the goal isn’t to pick one model forever. Keep your options open. Start with one critical workload. Evaluate it. Test alternatives. Write a runbook. Then decide where an open or specialized model wins.
Reply to your own thread (link goes here, not in the opening post):
Full conversation: oumi.ai/blog/ai-master…
On Nov 4, we’ll show how to build agents and train the models behind them, and you keep ownership of both. Register: oumi.ai/events/fall26
The questions they tackled:
• What happens when your provider changes pricing or access?
• Cost per token, or cost per completed task?
• Is the model the product, or the harness around it?
What they agreed on: the goal isn’t to pick one model forever. Keep your options open. Start with one critical workload. Evaluate it. Test alternatives. Write a runbook. Then decide where an open or specialized model wins.
Reply to your own thread (link goes here, not in the opening post):
Full conversation: oumi.ai/blog/ai-master…
On Nov 4, we’ll show how to build agents and train the models behind them, and you keep ownership of both. Register: oumi.ai/events/fall26
Jeremy Greer, Oumi's head of research, thinks he figured out how Jev was made. His guesses/analysis + how you can too:
- Qwen 3.5 or 3.6, based on obscure & famous knowledge comparisons, name probing, and sensitive question responses
- 35B A3B, based on obscure & famous knowledge comparisons, benchmark performance, and serving profile
- Tokenizer is a merge of Qwen and gpt-oss
- RLCD is likely just standard soft-label cross entropy from synthetic data
- Synthetic data is a mix of teacher-taught from reasoning models and programmatic answers to code-generated questions
- Serving is likely done with tree batching rather than standard vLLM
- Based on long document performance, they're likely using a form of sparse attention or serving on two GPUs and splitting the prefill between them. Most likely B200s or above.
No insider knowledge here, so these are best guesses based on experiments. All details in the article: oumi.ai/blog/how-to-bu…
- Qwen 3.5 or 3.6, based on obscure & famous knowledge comparisons, name probing, and sensitive question responses
- 35B A3B, based on obscure & famous knowledge comparisons, benchmark performance, and serving profile
- Tokenizer is a merge of Qwen and gpt-oss
- RLCD is likely just standard soft-label cross entropy from synthetic data
- Synthetic data is a mix of teacher-taught from reasoning models and programmatic answers to code-generated questions
- Serving is likely done with tree batching rather than standard vLLM
- Based on long document performance, they're likely using a form of sparse attention or serving on two GPUs and splitting the prefill between them. Most likely B200s or above.
No insider knowledge here, so these are best guesses based on experiments. All details in the article: oumi.ai/blog/how-to-bu…
oumi.ai How to Build a Decision Model Like Jev – Oumi Blog How TypeSafe's Jev likely works: one-token answers, calibrated probabilities, fast serving, and the clues that it's built on Qwen 3.5.