Pinned Tweet
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 weeks.
excited to see what everyone builds for the ultramarine track!
Hacktoberfest Week 3: Slow and Steady is live! 🦥
🐢 Build something with open-source AI that rewards laziness
🐢 $2,850 in prizes across 21 winners
🐢 Submissions due Oct 25
Join today: dev.to/devteam/join-the-hack…
ALT Hacktoberfest Open-Source AI Challenge: Week 3
Daanish Khazi retweeted
Hacktoberfest Week 3: Slow and Steady is live! 🦥
🐢 Build something with open-source AI that rewards laziness
🐢 $2,850 in prizes across 21 winners
🐢 Submissions due Oct 25
Join today: dev.to/devteam/join-the-hack…
ALT Hacktoberfest Open-Source AI Challenge: Week 3
Robust internal evals by Spenser and team @ Amplitude
+ very happy to see Ultramarine trade a slightly lower eval score than glm base for better qualitative output and style
Ultramarine performed strongly on our internal agent evals, with correctness close to GLM 5.3 Flash at ~1/10 the cost of Opus and ~1/60 that of Astra in the runs we compared. In our review, its answers were better organized, more explanatory and more engaging than GLM's.
I'm very excited for Paper Instruments, we're looking forward to working more with their models!
Daanish Khazi retweeted
Ultramarine 1 Flash by @paperinstr is now available on Design Arena!
Built for professional knowledge work, Ultramarine 1 Flash improves on its GLM-5.3 Flash base model across six benchmarks while using 6.5% fewer tokens. It delivers strong performance on complex tasks at just $0.13 per task, offering a cost-efficient option for agentic workflows.
Congrats to the @paperinstr team on the launch!
Daanish Khazi retweeted
Models are commoditizing much faster than expected.
The price/performance improvement curve on AI is nuts. There really are no permanent AI moats. I've been expecting 10x price/performance improvement per year, but this is 60x(!) cheaper in 6 months from the launch of Fable 5. We'll be trying Ultramarine out at Amplitude right away.
The spartan background in the launch video is a nice on brand touch too.
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 weeks.
Daanish Khazi retweeted
check out github.com/paper-instruments… to see how we use verifiers as a portable rollout engine across training frameworks.
the repo includes a reference tinker integration, but standardizing on verifiers made it very easy to integrate with miles, other tinker compatible apis, and prime-rl for experiments.
77%->98% on @patelnav's internal eval!
Daanish Khazi retweeted
post-trained glm 5.3 flash to understand american knowledge work.. give it a try!
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 weeks.
Daanish Khazi retweeted
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 weeks.
we were in pivot hell 6 months ago around when cursor announced composer 2.5, a high quality post-train of an oss base
made this diagram internally in May, earmarking open space in what will be a much larger category than coding
lots of fun late nights with the incredible team at @FireworksAI_HQ 🤝
Huge congrats to @bertgodel and the exceptional @paperinstr team on Ultramarine 1 Flash.
It’s also a proud @FireworksAI_HQ moment and a proof point for why we built Fireworks Training:
A five-person team used our training infrastructure to post-train an open model to Fable-class performance on knowledge work with ~$300K in post-training spend and 60× lower serving cost on their internal benchmark.
The key lesson: they kept the RL recipe simple and clean, leveraged SOTA training infra, and put the complexity where it mattered ->the learning signal:
→ 88 realistic work environments
→ 19 professional practice areas
→ Agent tasks of up to 250 turns
→ Task-specific rubrics that reward multiple valid answers
The result wasn’t just better quality. Ultramarine improved over its base model across all six knowledge-work benchmarks while generating 6.5% fewer tokens.
The next frontier won’t belong only to closed labs with the biggest budgets. Small, exceptional teams can compete with:
strong open models × great environments × exceptional talent × reliable training infrastructure
The best training infrastructure should disappear beneath the research, letting teams concentrate on environments, rewards, and model quality and not operating a distributed training stack.
Stable infrastructure multiplies what a talented team can do. That’s what we built Fireworks Training for. Don’t waste time paying the infra tax.
If you’re a small team with an ambitious post-training idea, we want to work with you.
Daanish Khazi retweeted
Huge congrats to @bertgodel and the exceptional @paperinstr team on Ultramarine 1 Flash.
It’s also a proud @FireworksAI_HQ moment and a proof point for why we built Fireworks Training:
A five-person team used our training infrastructure to post-train an open model to Fable-class performance on knowledge work with ~$300K in post-training spend and 60× lower serving cost on their internal benchmark.
The key lesson: they kept the RL recipe simple and clean, leveraged SOTA training infra, and put the complexity where it mattered ->the learning signal:
→ 88 realistic work environments
→ 19 professional practice areas
→ Agent tasks of up to 250 turns
→ Task-specific rubrics that reward multiple valid answers
The result wasn’t just better quality. Ultramarine improved over its base model across all six knowledge-work benchmarks while generating 6.5% fewer tokens.
The next frontier won’t belong only to closed labs with the biggest budgets. Small, exceptional teams can compete with:
strong open models × great environments × exceptional talent × reliable training infrastructure
The best training infrastructure should disappear beneath the research, letting teams concentrate on environments, rewards, and model quality and not operating a distributed training stack.
Stable infrastructure multiplies what a talented team can do. That’s what we built Fireworks Training for. Don’t waste time paying the infra tax.
If you’re a small team with an ambitious post-training idea, we want to work with you.
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 weeks.
Daanish Khazi retweeted
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 weeks.
Daanish Khazi retweeted
read the blog and this stuff is actually impressive!
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 weeks.
Daanish Khazi retweeted
Another win for UltraMarine: Classification
UltraMarine beats Luna and Haiku on accuracy, cost, and run-to-run consistency.
I run an email classification eval with Jev as primary and a fallback model when confidence drops below 85%. Jev clears that bar in only 78% of cases. For the other 22%, UltraMarine wins hands down.
So loving this model @bertgodel @joeybesgen
& the @paperinstr team.
More Evals (Fable is the arbitrator):
- 1/80 the cost of Opus,
- 'Accuracy: every number traced (335 of 335), and no factual errors'
- ' Paper is more detailed and practical'
Only I'm not going to call it "ultramarine-1-flash" :-)
@bertgodel @joeybesgen and the @paperinstr team.
I'm going to run a couple more evals - lmk if you want to see them here.
Daanish Khazi retweeted
More Evals (Fable is the arbitrator):
- 1/80 the cost of Opus,
- 'Accuracy: every number traced (335 of 335), and no factual errors'
- ' Paper is more detailed and practical'
Only I'm not going to call it "ultramarine-1-flash" :-)
@bertgodel @joeybesgen and the @paperinstr team.
I'm going to run a couple more evals - lmk if you want to see them here.
Early testing looking VV promising.
Kudos to @bertgodel and the @paperinstr team
It's the first test, tonight I'm going to run my full suite to compare it to Opus, Sol, Luna and Haiku 5.5
Daanish Khazi retweeted
Holy narrative violation
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 weeks.
Daanish Khazi retweeted
wow - lots of businesses will use this
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 weeks.