Articles by Lin
-
Owning vs. Renting Intelligence
Owning vs. Renting Intelligence
Mythos got shut down this week. Whether you agreed with the decision or not is almost beside the point.
25 Comments
Activity
37K followers
-
Lin Qiao reposted thisLin Qiao reposted thisWe heard you. You wanted GLM 5.2 to go faster. Introducing GLM 5.2 Fast, the same model and the same quality as GLM 5.2 Standard, now running at 140 tok/s on shared serverless. No reserved GPUs, no stack changes. GLM 5.2 is built for agent loops: reading repo context, planning, calling tools, editing files, running tests over long horizons. Speed is what makes those loops feel usable and keeps long context practical. Same weights, same OpenAI- and Anthropic-compatible APIs, full 1M-token context. It's just a faster path to your tokens. Need to go even faster? We've hit 446 tok/s on Artificial Analysis. Reach out and chat with us about a custom deployment, we'll tune one to your workload. Learn more → https://lnkd.in/gBVjvtQf
-
Lin Qiao shared thisHonored to be named to Magazine Forbes AI 50 and Fast Company AI 20. AI is becoming a new factor of production. The key question is no longer how much intelligence you consume, but how much intelligence you own. We're already seeing this happen at scale. Since this interview, daily token volume through Fireworks has doubled as more companies adopt open models and tailor them to their own products, data, and workflows. The future isn't one model serving everyone. It's millions of specialized models, owned and optimized by the companies that depend on them. We're still in the very early innings. As AI evolves from answering questions to executing work, the economics increasingly favor ownership. Every workflow becomes an opportunity to encode expertise, automate decisions, and compound productivity. Open models can be customized and optimized for specific use cases at a fraction of the cost of general-purpose models, turning intelligence into a strategic asset that compounds over time. Congratulations to everyone on the list. We're all building different pieces of the future, and it's still day one. https://lnkd.in/eeDDy_4xFireworks AI’s Lin Qiao wants to make AI agents cheaper to runFireworks AI’s Lin Qiao wants to make AI agents cheaper to run
-
Lin Qiao shared thisMythos didn't just expose a moderation problem. It exposed a dependency problem -- a company that owns its intelligence gets better every week; a company that rents it is one policy change away from starting over.
-
Lin Qiao shared thisYour model, your frontier! The question every team building with AI has been arguing about for three years: can open-source models do frontier-quality work on the hardest domains? We partnered with Harvey to test it on their new Legal Agent Benchmark. Legal work is unforgiving. Long reasoning chains, strict citation discipline, and under all-pass grading a single missed criterion fails the entire task. Two results stood out. First, hybrid legal agents can beat frontier models on quality and cost by routing selectively to a frontier advisor. An open-source GLM 5.1 worker does the bulk of the reasoning and drafting, and pulls in Claude Opus as a callable advisor only on the sub-tasks where a second opinion changes the answer. The advisor is sparsely targeted, which moved a tuned GLM to 18% all-pass, beating Opus 4.7 at 14%. Second, post-training can push open models to frontier-level legal performance. We post-trained Kimi K2.6, a trillion-parameter MoE, directly on LAB trajectories and lifted its score from a 0.863 base to 0.886 at 11x cheaper cost compared with closed models. Both of these ran on one platform. Fireworks trains, evaluates, and serves on the same infrastructure, so the model that comes off the training run is the same model, bit-for-bit, that serves production traffic. No numeric drift, no second deployment pipeline, no research-to-production gap to cross. What the leading AI companies are asking for has changed. It is no longer "how do we get the smartest model on every query." It is "how do we build specialized intelligence with a mixture of models, at a fraction of the cost, and control the ones that don't need the frontier." Open weights at the core. Applications get full control of the model, model performance, speed and cost. That is the pattern that won. See the full writeup with the Harvey team: https://lnkd.in/eduRe927
-
Lin Qiao shared thisI'm super excited to announce we GA-ed our first party product with Microsoft Foundry today! This is after 9 months of hard work to bring Fireworks platform to fully integrated with Microsoft Foundry. We offer the frontier specialized intelligence through the best quality models from Fireworks and the newest MAI models! We are ready to build deeply with Microsoft!
-
Lin Qiao shared thisFireworks is first party on Microsoft Foundry to deliver the best duo of frontier inference and enterprise guardrail. Learn more at https://lnkd.in/eCwVY5-4
-
Lin Qiao shared thisBest team! Join us!Lin Qiao shared this“3X is for losers” The most memorable quote from a very memorable first QBR as a leadership team. “3x 3x 2x 2x” is a rule of thumb that most companies hope to follow. But that rule just doesn’t work for a space as big as inference or a company as fast as Fireworks AI. Scaling a sales organization when you have the wind at your back is amazing. If you’re curious, independent and operate with urgency - you’ll fit in great here. This team is having fun every day. Come join us! Roberto Barroso-Luque Steven Wu Brenna Shubin Patty Li Chan Manchanda Rob Ferguson Lavinia(Xiaolei) Liu, SHRM-SCP Mark McAndrew Tingting Bi David D. Lee George Hu
-
Lin Qiao reposted thisLin Qiao reposted thisWe didn’t ship DeepSeek V4 on Day 0 like we always do. Why? We love speed at Fireworks AI, but quality >> speed. While running our extensive evals, we found that even the official reference model code was producing corrupted outputs. When we tested over the weekend, all endpoints except the official DeepSeek API had these issues. After 2 days of extensive debugging with DeepSeek AI, SGLang, and vLLM communities, the issues are fixed, and we’re proud to serve DeepSeek V4 Pro in all its glory. Check the full story in our blog: https://lnkd.in/gja2cTjUDeepSeek V4 Pro: Validating Frontier Models for ProductionDeepSeek V4 Pro: Validating Frontier Models for Production
-
Lin Qiao shared thisI'm thrilled to welcome George Hu as President of Fireworks AI! When the seven cofounders started the company, we wanted to honor our roots in PyTorch; torch carries the fire. We want the fire of AI to serve people everywhere. We named our company Fireworks to move the industry forward to be AI-first. Three and a half years in, we have grown substantially. Today, we process 30 trillion tokens a day, the biggest inference provider outside of the frontier labs’ APIs. We got to this impact by relentlessly defining what frontier inference is. It is the control we give to all application developers to customize their own model, deliver magically fast experience, and scale with maximum efficiency. We have built multiple product lines with Fireworks Training, Fireworks Inference and the best open models on one platform. We are accelerating to the next phase of scaling – enabling customers everywhere to experience the power of their own private models on Fireworks. This growth demands a world-class business leader, a unicorn, a one-of-a-kind. George is this leader. He is one of a very small number of people who have scaled not one but two iconic software companies. He worked alongside Marc Benioff at Salesforce, where he helped grow revenue from $100M to $5B+, and then did it again with Jeff Lawson at Twilio, driving nearly 10x growth in five years. When George and I first met a year ago, I said to him – we are too small for you. But today, we are ready! I admire the conviction he brings to Fireworks, the same conviction I've had with the other cofounders since Day One. George, thank you for your trust in me and the Fireworks team. Let's build something generational! We're still early. The companies winning with AI are the ones who own their intelligence and their differentiation. Our job is to make that possible for every company in the world. And if your conviction is as strong as George’s and mine, we're hiring. Let’s bring the 🔥
Lin Qiao commented on a post
1mo
Welcome Leo Zheng! Very excited to work together!
-
Lin Qiao liked thisLin Qiao liked thisWe heard you. You wanted GLM 5.2 to go faster. Introducing GLM 5.2 Fast, the same model and the same quality as GLM 5.2 Standard, now running at 140 tok/s on shared serverless. No reserved GPUs, no stack changes. GLM 5.2 is built for agent loops: reading repo context, planning, calling tools, editing files, running tests over long horizons. Speed is what makes those loops feel usable and keeps long context practical. Same weights, same OpenAI- and Anthropic-compatible APIs, full 1M-token context. It's just a faster path to your tokens. Need to go even faster? We've hit 446 tok/s on Artificial Analysis. Reach out and chat with us about a custom deployment, we'll tune one to your workload. Learn more → https://lnkd.in/gBVjvtQf
-
Lin Qiao liked thisLin Qiao liked this🔬 𝗔𝗜 𝗗𝗲𝘀𝗶𝗴𝗻: 𝗥𝗼𝗯𝗼𝘁 𝗧𝗵𝗮𝘁 𝗚𝗼𝗲𝘀 𝗔𝗻𝘆𝘄𝗵𝗲𝗿𝗲! 𝗢𝘃𝗲𝗿𝘃𝗶𝗲𝘄 Artificial Intelligence is NOT right for everything. There. I said it! Now that we're done with that, it's probably a good idea to focus on use cases where it can be very helpful. For example, Pliant Energy Systems Inc robots use several modes of locomotion found in the animal kingdom with just one pair of “fins”. These fins are best described as four-dimensional objects with a hyperbolic geometry that allows the robot to swim like a ray, crawl like a millipede, jet like a squid, and slide like a snake. A craft equipped with this system has unprecedented freedom to travel through a range of environments in a single mission. As an underwater vehicle, the robot’s ability to instantly reverse direction and do quick turns make it ideal for tasks such as coral reef inspection or lion fish hunting where a craft must rapidly maneuver to look around and between objects. 𝗛𝗶𝗴𝗵𝗹𝗶𝗴𝗵𝘁𝘀 1. Efficient undulating propulsion system 2. Excellent static thrust/Watt 3. Stealth characteristics 4. Amphibious capabilities over sand, snow, ice and other solid substrates 5. Resistant to entanglement in plants or other aquatic debris 6. Lower environmental impact than spinning propeller-thrusters 𝗟𝗶𝗻𝗸 https://lnkd.in/ec9-s2VF T. Scott Clendaniel
-
Lin Qiao liked thisLin Qiao Looking forward more MAI model serving tokens from Fireworks AI !
-
Lin Qiao reacted on thisLin Qiao reacted on thisReflecting on last week at #MSBuild which had so many highlights and sharing a few that were particularly meaningful for me: 1. GA of Fireworks AI on Foundry. We know enterprise customers make a portfolio choice on models based on performance, cost, and use case specificity. At Microsoft we strongly believe in model choice and when choosing OSS - Fireworks AI on Foundry provides 20+ frontier open models, BYOW to support custom models, high throughput, low latency, and enterprise grade controls customers expect from Azure. TY to Lin Qiao, George Hu, Brenna Shubin and the entire FW team for the deep collaboration. Excited to scale and drive even more joint customer success from here! 2. The genuine buzz and excitement from key AI partners and customers from across the ecosystem on building with Azure including Nvidia, Meta, Mistral, MongoDB, IBM, to name a few. Gained key demand signals to inform meaningful partnership opportunities to meet customers' needs. Joseph Fu Dion Ubert ☁️ Stuart Francis Michael Leyva 3. Spending some time with my team and the extended team of stakeholders I get the privilege to work alongside at Microsoft. Nothing beats in person connections even in a world full of Teams videos and chat. Jason Graefe Sheel Majumdar Ayelet Steinitz Greg Goldkamp Mackenzie Kosut Nanditha Ram Bitrus Baba Chibuzo Obi Derek Loar Naomi Moneypenny Long Chen Yina Arenas Trupti Parkar Felipe Becerra Phil Kim Michelle Gonzalez Matthew Kerner
Experience & Education
-
Fireworks AI
View Lin’s full experience
See their title, tenure and more.
or
Publications
-
On brewing fresh espresso: LinkedIn's distributed data serving platform
SIGMOD
Other authors -
-
Blink: Not Your Father's Database!
Proceedings of the 5th International Workshop on Business Intelligence for the Real-Time Enterprises (BIRTE)
The Blink project’s ambitious goals are to answer all Business Intelligence (BI) queries in mere seconds, regardless of the database size, with an extremely low total cost of ownership. It takes a very innovative and counter-intuitive approach to processing BI queries, one that exploits several disruptive hardware and software technology trends. Specifically, it is a new, workload-optimized DBMS aimed primarily at BI query processing, and exploits scale-out of commodity multi-core processors…
The Blink project’s ambitious goals are to answer all Business Intelligence (BI) queries in mere seconds, regardless of the database size, with an extremely low total cost of ownership. It takes a very innovative and counter-intuitive approach to processing BI queries, one that exploits several disruptive hardware and software technology trends. Specifically, it is a new, workload-optimized DBMS aimed primarily at BI query processing, and exploits scale-out of commodity multi-core processors and cheap DRAM to retain a (copy of a) data mart completely in main memory. Additionally, it exploits proprietary compression technology and cache-conscious algorithms that reduce memory bandwidth consumption and allow most SQL query processing to be performed on the compressed data. Ignoring the general wisdom of the last three decades that the only way to scalably search large databases is with indexes, Blink always performs simple, “brute force” scans of the entire data mart in parallel on all nodes, without using any indexes or materialized views, and without any query optimizer to choose among them. The Blink technology has thus far been incorporated into two products: (1) an accelerator appliance product for DB2 for z/OS (on the “mainframe”), called the IBM Smart Analytics Optimizer for DB2 for z/OS, V1.1, which was generally available in November 2010; and (2) the Informix Warehouse Accelerator (IWA), a software-only version that was generally available in March 2011. We are now working on the next generation of Blink, called BLink Ultra, or BLU, which will significantly expand the “sweet spot” of Blink technology to much larger, disk-based warehouses and allow BLU to “own” the data, rather than copies of it.
Other authors -
-
Integration of Server, Storage and Database Stack: Moving Processing Towards Data
ICDE
Other authors -
-
A framework for enforcing application policies in database systems
SIGMOD
Other authors -
Languages
-
English
-
Recommendations received
1 person has recommended Lin
Join now to viewOther similar profiles
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top contentOthers named Lin Qiao in United States
17 others named Lin Qiao in United States are on LinkedIn
See others named Lin Qiao