Fireworks AI

Building upon our past work with Trilogy AI Center of Excellence, it's clear that the unit economics of AI matter deeply. Trilogy is not just a day-zero adopter of frontier open models, but built a playbook for cyber defense using Kimi K3 on Fireworks. Audit every repo, triage every scan, review every PR. Get the full details: https://lnkd.in/gf-9winp

I'm super excited to share we launched Fireworks Nexus! When I give talks, I listen to the audience, trying to understand and track their reaction in the moment and over time. When I raised "scaling AI into bankruptcy", the room was laughing two years ago. Now the room goes silent. What used to feel like someone else’s problem has become their acute pain. Historically, most teams adopted the best frontier models because that was the right decision. Today, open models have caught up, and most routine development tasks no longer require the most expensive frontier model. But now you have a different problem: which model should run which task? Switching models, routing requests, tuning caching, managing enterprise controls, and figuring out where the AI budget is actually going quickly becomes an infrastructure and operation problem of its own. That’s why we built Fireworks Nexus. Nexus is an intelligence layer that sits underneath the agentic harnesses developers already use. It gives engineering teams access to the latest open models, including GLM 5.2 and Kimi K3, automatically routes requests to the right model, maximizes cache efficiency, enforces zero data retention, and provides a single view into how AI is being used across the organization. We recently published a study on oracle routing showing how much this can matter in practice [https://lnkd.in/erifJ24V] Teams should measure whether an open model actually delivers the same outcome as a frontier model on their own codebase and business workflows, rather than relying on generic benchmarks or intuition. I don’t think the future is every developer doing tedious work in model selection. The future is an intelligence layer underneath every organizations – continuously incorporating the best open models as they emerge while keeping quality high, operational complexity invisible, and AI spend sustainable. Two years ago, the problem was whether AI could create enough value. Today, the problem is making sure that value doesn’t get eaten by the cost of running it.

Notion's webinar (tomorrow morning at 10am PT!) will feature our recently launched Fireworks Nexus as a way to help teams build with AI in a way that feels useful, sustainable, and ready to scale. Sarah Sachs, AI Engineering Lead at Notion, will share how to think about AI systems, budgets, and frameworks in the real world, including what teams should consider to avoid going from AI-pilled to AI-poor.

#FireworksAI offers the best performance for Kimi K3 by a huge margin!!! https://lnkd.in/gWCv_vaf

"I believe that the future will not be a small number of AGI models dominating the world... it will be millions of specialized models. One per application, per use case." - Lin Qiao Thanks again for hosting Harry Stebbings and 20VC!

ICYMI yesterday, Kimi K3 is live and Fast on Fireworks. Things to know: - Performance: same quality as top models from Open AI and Anthropic on coding and agent tasks, at a fraction of the cost. - Privacy: Fully US-hosted endpoints and Zero Data Retention, making it day-zero compliant for large enterprises. - Serverless + Training: no GPUs to manage, simply scale instantly with Fireworks Serverless (Fast, Priority tiers available) or fine-tune K3 on your own data using our Serverless LoRA training supported. Stop renting intelligence. We are excited to see what you build. Get started with Kimi K3 now: https://lnkd.in/gRw_Hg4V

We got Day-0 access to Kimi K3 via Fireworks AI, and it’s now the highest scoring open weights model on APEX-Agents. Mean score: 55.4% Pass@1: 39.3% Rank: #7 overall K3’s low token usage paired with their lower pricing makes it a compelling combination. It uses just over half the input tokens as Opus 5 while costing 40% less per token. That represents a significant potential cost savings. We first evaluated Kimi K2.5 in February at 14.4% Pass@1. Since then the family has climbed steadily: K2.6 reached 18.9%, K2.7 Code reached 27.6%, and K3 now sits at 39.3%. Pass@1 nearly tripled in six months, and the frontier is in sight: the current APEX-Agents leader, Opus 5 (Max), scores 43.5% Pass@1 and 60.6% mean. APEX-Agents includes 480 long-horizon tasks across professional services domains like investment banking, management consulting, and corporate law. Across the three APEX-Agents domains, K3 scored: Management consulting: 44.4% Pass@1 (#6) Investment banking: 40.5% Pass@1 (tied #5 with GPT-5.5 xHigh) Corporate law: 33.1% Pass@1 (#6) In corporate law, K3's mean score runs 25+ points ahead of its Pass@1. That means K3 does much of a legal task correctly. It can find the right documents and draft usable language, but is not yet ready to hand over a finished, client-ready document. That profile makes it a strong first-pass assistant for legal work, with a lawyer still closing the last mile. Running APEX-Agents on an open weights model at release requires an inference partner like Fireworks AI. Their zero data retention endpoint lets us run evals with no risk of task data leaving the environment, and their serverless API scales to what long-horizon agentic evals demand without any of the headache of managing the underlying serving infrastructure. Congratulations to the Kimi (Moonshot AI) team!

Heard ya’ll are into open weight models? Kimi K3 is now available in Notion. Hosted by Fireworks AI, a US-based provider.

What a week of frontier intelligence with Kimi K3 and Opus 5 launch! We compared these two great models on a few benchmarks of the quality and cost, based on our most common use cases. Important measurement is **task-base cost**, not token-base cost. In general, open models tend to be more verbose than close models. Hence the quick study. Across SWE (480), Algorithmic(100), Terminal (83), the per-task quality is close with Opus 5 being on par or better, and K3's per-task cost is 2x to 4.6x cheaper, based on the serverless pricing. We aim to increase per-task serving efficiency (via Fireworks Inference) and quality (via Fireworks Training). Share what you find out for the tasks you care about. 👇 More details of our study -- https://lnkd.in/eYtHQR2D

Today, we’re introducing Fireworks Nexus. Fireworks Nexus gives engineering organizations control over the intelligence powering the AI tools they already use. Instead of being locked into a single provider’s pricing, models, and roadmap, you can choose the best model for every task, manage spend centrally, and continuously measure quality as new models emerge. Now you can get the same or better engineering outcomes while spending less. Learn more at the link in the comments.