Fireworks
@FireworksAI_HQ
The frontier platform for training and inference on open-weights models at scale.
304 Following 32.8K Followers
Proud to be a launch training partner for @LangChain's LangSmith Fine-Tuning and smithtune.
LangSmith traces → managed SFT on Fireworks → eval → deploy. One CLI, no training infra.
Get a Fireworks API key:
Then run smithtune:
Show more
We ran 18 models across 113 real coding tasks on DeepSWE, then went back and asked a simple question: what if every task had routed to the model that handled it best?
Answer: 97.6% solve rate at $1.88 per task, versus the best model at 74.1% at $6.52.
The next frontier is a router.
Full analysis:
Show more
We’ve just added three new speakers to our incredible Fireworks Forge lineup!
We are proud to welcome @BrendanFoody @pirroh and @jefftangney to the stage as we bring together the people, teams, and companies building their own frontier on open models.
Show more
Many assume that agent spend goes toward output tokens.
When we ran DeepSWE on Astra vs. DeepSeek V4.1-Flash, input tokens outnumbered output 174 to 1. 99.6% were cache hits. Those hits are 60% of the bill.
Net result? Same quality. $0.43/task vs $6.52.
Show more
Speed is key for @juicebox_work. They need to search hundreds of thousands of talent profiles in seconds with multiple real-time search agents.
We helped them cut latency 80% and drop inference costs from $4M to $800K/yr, using specialized models tailored to their use case.
Show more
Excited to continue our partnership with the @alibaba_cloud team on this one, bringing the latest version of Qwen's flaship model to the Fireworks community. Start building with Qwen3p8 today:
Show more
Developers need speed and uncompromising production quality. 🚀
We’ve partnered with @FireworksAI_HQ to deliver exactly that for Qwen 3.8 Max. Build faster, scale effortlessly. 🤝
#Qwen# #FireworksAI# #LLM# #OpenSource#
Show more
Reinforcement learning at @cognition's scale is a hard infrastructure problem. We are proud to be part of the stack behind it.
Congrats to the team on SWE-2!
Read more about how we think about RL at Fireworks:
Show more
Introducing SWE-2, our closest model yet to the frontier.
On leading evals, it scores on par with recent frontier models – at up to 70% lower cost.
We scaled RL to multiple trillions of parameters, with a refined recipe that pushes the Pareto curve on both capabilities & cost.
Show more
Recently, we hosted our 12th Nerd Meetup, this time at @lightfld 's office in SF.
Builders came out from Lightfield, @qdrant_engine, and Fireworks.
Building in the AI/infra space and want to join us for the next meetup?
Follow along. We'll announce the next one soon!
Show more
GLM-5.3 is now available for training on Fireworks Dedicated Training API and Managed Training surfaces.
Built for complex coding and long-horizon agents and now available to train on our dedicated infrastructure.
Get started today:
Show more
Deepseek-V4.1-Flash is available now on Fireworks!
It is a 552B Parameter MoE built for coding, cybersecurity, and agents. It is the ideal workhorse model that outperforms Opus 5 and GPT 5.6 Sol at 1/40th the cost on DeepSWE,CyberGym, and Automation Bench!
Interested in higher quality? Reach out, as we’re bringing this model to Fireworks Training soon.
Try Deepseek-V4.1-Flash now on Fireworks:
Show more
The durable advantage is Genspark’s definition of a good deck: it lives in their evaluation, so they can retrain against it each time a stronger base model lands.
Own the intelligence at the core of your product and the quality ceiling is yours to raise.
Full write-up:
Show more
The hardest bugs were numerical. Long-horizon RL drifts when the engine generating rollouts and the one scoring them stop assigning the same probabilities to the same tokens. Aligning tokenization across both kept the training signal trustworthy.
Show more
Fireworks Lab co-developed the RL algorithm with Genspark and ran the training on frontier-grade infrastructure.
Our embedded researchers engineered the reward, ran 100+ experiments, and read trajectories to catch the model gaming the score.
Show more
Genspark owned the product judgment.
They brought their real production environment into training and shaped what the model optimized for: the standard for a good deck, the design principles behind it, and how to sharpen that standard over time.
Show more
Introducing Gen-1 Slides: an open model that matches Claude Opus 5 on slide generation, at ~1/17 of its input-token price.
@genspark_ai post-trained it from a @MiniMax_AI M3 base with Fireworks Lab using long-horizon RL. Live today as the default in Genspark AI Slides. 🧵
Show more
Today, 10am PDT: we'll cover the journey from renting closed frontier models to owning your specialized intelligence, with a practical framework for knowing when to make the leap.
Includes open Q&A w/ Head of AI Developer Education @Prof_OZ
Register:
Show more
Renting a closed model + a brittle, overfit harness only gets you so far.
Routing to open models helps quality, cost, and latency, but you hit a ceiling.
Most ambitious AI teams train their taxonomy, style, and judgment into models they own.
Read more:
Show more
We invite you to Forge.
At Forge, we will challenge you to make your own frontier.
You'll spend the day with the builders, researchers, and technical leaders pushing AI forward.
Lin Qiao. Jensen Huang. Jay Parikh. And more.
Apply to attend:
Show more
Congrats to one of our Kimi K3 Fire Pass Community Hackathon winners @sethsaler for his winning entry: Dayprint!
Our Head of AI Developer Education, the esteemed @Prof_OZ, walks through Seth's entry with live commentary.
#kimik3onfireworks#
Show more
Inspired by all the cool projects I've seen lately on X, dayprint brings your personal context together to plan your day, review your performance, and synthesize how your week went.
My entry for #KimiK3onFireworks# ↓ @FireworksAI_HQ
Show more
Our partners at @CosineAI train their own coding model on real production code, for languages like Fortran and Verilog that generic models fumble.
On cost per successful task, their Lumen Outpost model beats GPT-5.5 by more than 3x. The whole pipeline runs on Fireworks Training.
Build your own frontier:
Show more
