About
Technologist at heart. Always learning, always building — compilers, distributed systems,…
Articles by Yun
Activity
8K followers
-
Yun Jin shared thisWe missed day0 launch of a major model, intentionally, again. ⏸️ When GLM-5.3-Flash weights dropped, we are ready on every aspect, from performance to benchmark evaluation scores. Except we found that common open-source engines were burning roughly 2x more tokens to get there on AIME and GPQA. Same answers, much longer thinking. Score parity said ship. The token distribution said something was still unexplained. So we held the public launch, moved the model into a disclosed private preview, and dug in alongside Z.AI, vLLM and SGLang. Two days later the official API's behavior changed and the reasoning lengths converged. One trace still sticks with me: the model correctly computed an answer, then talked itself into doubt over a competition problem it had misremembered, and spent thousands of tokens re-litigating a step it had already gotten right. 🤔 I wrote up what the investigation taught me, and why quality is a distribution rather than a number. 📊 #AIInference #LLM #OpenWeights #ModelEvaluation #GLM #ArtificialIntelligence #Fireworks
-
Yun Jin shared thisHarvey — the legal-AI category leader — just shipped Tenet, its first post-trained model. Built on Kimi K3, trained with Fireworks AI for long-horizon legal work: diligence runs, evidence sweeps, the tasks that take hours, not seconds. Here's what that buys: By tuning on its own data, Tenet outperforms the best closed models. And that data never leaves Harvey's hands, so every improvement compounds into a moat no lab can absorb. That second part matters beyond Harvey. When application companies can build data moats, the industry stays healthy. When they can't, a couple frontier labs capture all the value — and the applications starve. This is the trend: scaling a business requires owning intelligence. Fine tuning over open models is the way to go. Cursor did it, Harvey did it, all category leaders will. We are proud to support the movement.Yun Jin shared thisCongrats Harvey team for launching Tenet, a specialized frontier model in legal! We are excited to drive the training work of Tenet with Harvey and collaborate with Mercor on data. Together, we delivered a frontier-quality model across 24 areas of corporate law. Tenet exceeded Opus5 and Fable at many dimensions, covering 1300+ legal tasks.
-
Yun Jin shared thisI watched Llama leading the open weight model movement when I was in Meta. Then the center of gravity drifted, and for a while the frontier of open weight lived elsewhere. Meta is back. Muse Glimmer 30B is a strong American open-weight model — and one of the best its size for the work that matters now: agentic tasks, reliable tool use, long-horizon reasoning. Open model matters. A closed frontier API is rent, not ownership. It improves on everyone's data yet belongs to no one; the vendor sets the roadmap and the price. That's why Fireworks AI sees specialized intelligence - models built by your data, for your use cases, and owned by yourself - is the critical piece of the AI ecosystem. Try Fireworks' Dedicated Training API today for Glimmer, LoRA or full-parameter. 👇Yun Jin shared thisMuse Glimmer 30B is now available on Fireworks' Dedicated Training API for both LoRA and Full-Parameter fine-tuning: a U.S.-developed, open-weight model and one of the strongest of its size for agentic work, reliable tool use, and long-horizon reasoning. Serverless Training API access is coming shortly. Try it out today: https://lnkd.in/gCNfffCy
-
Yun Jin shared thisI spent years in Internet companies building systems on one sacred rule: scale out beats scale up. Cheap machines, loose coupling, failure as routine. Racks like GB300 NVL-72 feel like an anti-pattern through this lens. 72 GPUs, liquid-cooled, wired to behave like one machine. It's on the scale up path. Except AI workloads demand intense computation with tightly coupled communication — and not just training; inference is pushing the same boundary now. A 2.8T-parameter MoE model all-to-alls across its expert group every single token. Push that group past the 8-GPU NVLink island and your collectives crawl across InfiniBand at a fraction of the bandwidth. A very expensive way to be slow. The rule that actually holds: Scale out the request. Scale up the token step. Replicas, routing, multi-region — still scale-out. But the token step now wants a rack-scale fast domain, gang scheduling, per-token lockstep. The scheduler isn't a bin-packer anymore; it's part of the inference system. My take: why the rack is the new node 👇 #AIInfrastructure #Inference #NVIDIA #NVL72 #LLM #ArtificialIntelligence #GB300 #Kimi #HPC #SuperComputerThe Rack Is the New Node: NVL-72 and Frontier InferenceThe Rack Is the New Node: NVL-72 and Frontier Inference
-
Yun Jin shared thisSame weights. Same model name. Not the same model. ⚖️ Artificial Analysis ran the same evals against every endpoint serving GLM-5.2: scores ranged from 102% of the reference down to 73%. Whatever drives a gap that size, it isn't the weights — those are identical everywhere. So where does it come from? My understanding points to three places: the numerics (quantization method, kernel choice, floating-point ordering), the plumbing (chat templates and parsers that silently break tool calls), and the budget (output caps that ration reasoning). None show up on a spec sheet. 🔬 Speed has numbers anyone can read. Quality doesn't — and the unmeasured term is the one that gets given up quietly. Day 3 with correct outputs beats Day 0 with silent corruption. Please read the full piece. 👇 #AIInfrastructure #LLM #Inference #MachineLearning #Fireworks
-
Yun Jin posted thisWhat a week for open models—and for the Fireworks team. 🎆 Fireworks AI launched four amazing open-weight models on day0, each bringing something different to production agents: - Muse Glimmer brings frontier-adjacent agent quality to a compact 30B multimodal model, with 131K context and a Meta-reported 75.5 on MCP Atlas. - Qwen3.8-Max is the second open model after Kimi K3 to exceed two trillion parameters. It finished Qwen’s year-long E-Commerce Bench 38% ahead of second place. - DeepSeek-V4-Pro-0813 reached 87.9 on Terminal-Bench 2.1 through improved post-training—without changing the underlying architecture—and supports 1M context. - Nemotron 3.5 Lightning scored a reported 86% on PinchBench while completing 10,000 tasks 30% faster than models at the similar size. Three trends stand out for me: - Long context is becoming an architectural challenge, not just a larger token limit. Sliding-window attention, recurrent linear attention, compressed sparse attention and Mamba–Transformer hybrids all make history cheaper to process while preserving global recall. - The benchmark unit is shifting from an answer to a trajectory. These evaluations measure whether agents can plan, use tools, recover and finish real work across hundreds or thousands of interactions. - Post-training and harnesses are becoming part of the model. DeepSeek’s quality jump and released agent harness show how better environments and feedback loops can create major gains without another architecture or pretraining run. 🧠 At Fireworks, being early matters—but correctness matters more. ⚙️We are consistently among the earliest teams to launch new models, but only after validating that their quality persist the serving stack. Huge thanks to our team for moving incredibly fast while holding that bar—and to Meta, Qwen, DeepSeek and NVIDIA for pushing open intelligence forward. 🙏 Check them out 👇 * Muse Glimmer: https://lnkd.in/gRMeKKcu * Qwen3.8-Max: https://lnkd.in/gCgXc-sE * DeepSeek-V4-Pro-0813:https://lnkd.in/g45xNP9z * NVIDIA Nemotron 3.5 Lightning:https://lnkd.in/g5KGnnNJ #LLM #AgenticAI #AIInfrastructure #DeepSeek #Qwen #MuseGlimmer #NVIDIA #FireworksAI
-
Yun Jin shared this🎆 Kimi K3 shipped on Fireworks day zero — inference and training, US-hosted, zero data retention (https://lnkd.in/gS4CDgJA). The first frontier open model in the 3-trillion-parameter class: 1M context, native vision, reasoning that rivals the top closed models. 🤔 I've watched people reason about what K3 costs to serve, and almost every take picks one dimension and runs with it: → It should be slow — 2.78T parameters. → It should be cheap — 75% less KV cache. → It's a ~50B model per token — only 16 of 896 experts fire. Each is anchored to something real. None survives contact with the others. And the third is off by about 2x — Moonshot's own model card reports 104B activated, the highest of any open-weight frontier model. 📐 Serving cost was never one number. K3 moved parameters, depth, context, sparsity and attention at the same time, and which of them binds you shifts with context length, batch size, cache precision, and how the engine is configured. 🔍 So I worked out the full picture: what each architectural change actually bought, where the popular intuitions break down, and how different constraints shape the serving speed. #KimiK3 #LLMInference #Fireworks #AI #LLM
-
Yun Jin reposted thisYun Jin reposted thisKimi K3 is live on Fireworks, day zero: the first open model in the 3 trillion parameter class to reach frontier quality on independent benchmarks. Built by the Kimi (Moonshot AI) team, K3 sports a 1M-token context window and native vision. On independent evals it's #1 in the world at frontend code (Arena, ahead of every closed model) and #3 on the Artificial Analysis Intelligence Index. Kimi K3 rivals Anthropic and OpenAI’s best models at a fraction of the cost. On Fireworks, you serve K3 on our own US infrastructure with zero data retention by default. You can fine-tune it on your private data from day zero, and own the weights you train. You want a moat? This is how you get a moat. Own your intelligence. Read more at the link in the comments.
-
Yun Jin shared thisAn open model just tied the closed frontier. Kimi K3 vs Fable 5 on coding: 92.4% to 92.6%. The interesting part isn't the tie. It's that they fail in completely different places. K3 is sharpest on symbolic math and dev tooling; Fable wins on web & data visualization work. Same tier. Different price. Complimentary strengths. Combining the two will yield the best quality plus economics. Mixture-of-Experts asked which MLP should handle a token. Mixture-of-Models asks which model should handle a task. The intelligence is the router, not the model. New piece 👇 #LLM #AI #Kimi #GPT #AIInference #FireworksAI
-
Yun Jin liked thisYun Jin liked thisFactory's droids write and commit code autonomously, at a volume no human reviewer can keep up with. That makes the safety layer critical: the check that catches secrets (API keys, passwords) before they reach production has to be trustworthy without getting in the way. Factory used the Fireworks Training API to fine-tune an open Qwen base into two specialized models: one to catch secrets, and one to clear false alarms. On their own held-out, repo-level evals, the trained model caught about 70% of real secrets at a 5% false-alarm rate versus ~59% for GPT-5.5, while being faster and more cost-effective. And with Fireworks handling the infrastructure, they got there without standing up training and inference infrastructure on their own. Build your own frontier: https://lnkd.in/gBrJhMRB
-
Yun Jin liked thisYun Jin liked thisExcited to introduce Sico, a Microsoft open-source project exploring the co-evolution of humans and Digital Workers — and a future of carbon–silicon symbiosis. As AI takes on more meaningful work, the opportunity goes far beyond automation. It is about humans and Digital Workers working, learning, and evolving together — forming a new kind of C+Si team where each side continuously amplifies the other. Sico is an exploration of what that future could look like. Check it out, try it, and contribute: https://lnkd.in/gkTG2uST
-
Yun Jin liked thisYun Jin liked this⚡️ Qwen3.8-Flash Language Performance & Visual Language Performance
-
Yun Jin liked thisYun Jin liked thisA belated post: I’ve left Google a little while ago. In the last few weeks, I spent lots and lots of time with my family, and focused on my health. While our family has been going through some difficult times, this time has also been an incredibly rich and reflective time and I’m so grateful that I was able to do it. We made memories together, had epic adventures, and cherished time with each other as much as we could. I’m back now and looking to the future with optimism. I will post more on my next steps later. Reflecting back on my time at Google, the word that comes to mind most is gratitude. I’m really proud of a lot of the work we accomplished together. But most of all, I’m grateful that I got to share in part of this journey with a lot of great colleagues, some of whom have become dear friends. Some of you know that I was a “boomerang” at Google, having left and returned. This is because I still felt really connected to the company, people, and work even after I’d left for the first time. The same is true now that I've left for the second time. Thank you to all of you who have worked with me, supported me, challenged me, made me a better manager and coach, pushed me to deliver things I didn’t think we could deliver, and simply shared the journey with me. I’m cheering you on.
-
Yun Jin liked thisYun Jin liked thisHonored to be in Time100AI. The award has my name on it, the work has hundreds of names on it. My cofounders, the engineers who worked tirelessly and whose capacity for invention is extraordinary. Before their work nothing like our chip had ever existed. Awards go to individuals. Breakthroughs come from teams. Thank you to TIME for the recognition. And thank you to the team at Cerebras - the fearless engineers who turned a previously impossible idea into reality.
-
Yun Jin liked thisTo ship frontier intelligence, speed isn’t enough!
-
Yun Jin liked thisYun Jin liked thisWe've partnered with Fireworks AI to bring Voyage AI models to their dedicated inference platform. That means embed, retrieve, rerank, and generate all in one place, right alongside the open models you're already running. https://lnkd.in/eWXCB-N3Your AI Performance Stack is Fireworks Models with Voyage AI embeddingsYour AI Performance Stack is Fireworks Models with Voyage AI embeddings
Experience & Education
Publications
-
Dynamo: Facebook’s Data Center-Wide Power Management System
ISCA 2016 (ACM/IEEE International Symposim on Computer Architecture)
See publicationA data center-wide power management system that monitors the entire power hierarchy and makes coordinated control decisions to safely and efficiently use provisioned data center power.
-
Watching videos from everywhere: a study of the PPTV mobile VoD system.
Internet Measurement Conference 2012
Patents
-
Graphical Programming Object Population User Interface Autogeneration
Filed US 20120227028
-
Scheduling by Growing and Shrinking Resource Allocation
Filed EU EP20080826472
A scheduler for computing resources may periodically analyze running jobs to determine if additional resources may be allocated to the job to help the job finish quicker and may also check if a minimum amount of resources is available to start a waiting job. A job may consist of many tasks that may be defined with parallel or serial relationships between the tasks. At various points during execution, the resource allocation of active jobs may be adjusted to add or remove resources in response…
A scheduler for computing resources may periodically analyze running jobs to determine if additional resources may be allocated to the job to help the job finish quicker and may also check if a minimum amount of resources is available to start a waiting job. A job may consist of many tasks that may be defined with parallel or serial relationships between the tasks. At various points during execution, the resource allocation of active jobs may be adjusted to add or remove resources in response to a priority system. A job may be started with a minimum amount of resources and the resources may be increased and decreased over the life of the job.
Other inventorsSee patent
Languages
-
English
-
-
Chinese
-
Other similar profiles
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content