Sinan Ozdemir

AI Education @FireworksAI_HQ 🔥🎆 NLP + Gen AI Expert / LLM whisperer 🤖 AI Author 📚 Founder @Aikylie (acquired) 📈 Fellow @YCombinator

[Sinan Ozdemir @Prof_OZ]

https://t.co/HwKgitTPFy

[Fireworks @FireworksAI_HQ]

DeepSeek V4.1 Flash 1017 runs at 1,521+ tok/s on Fireworks Ultrafast.

Fast enough to ship Tetris in 11 seconds, 7x faster than Claude Opus 5.5.

Join the Ultrafast waitlist: https://t.co/vE0jM7Lyou

[Sinan Ozdemir @Prof_OZ]

I saw a great post from @nutlope and wanted to take a shot at making my own Jev-like classifier as well on @FireworksAI_HQ.

I'm so happy people are starting to love LLM-based classifiers so anything I can do to keep the hype going, I'm game.

[Sinan Ozdemir @Prof_OZ]

https://t.co/HwKgitTPFy

[Sinan Ozdemir @Prof_OZ]

@natolambert @trillium_labs Simply Amazing!

[Sinan Ozdemir @Prof_OZ]

I'm especially looking at "mixed across multiple harnesses."

Harnesses change pretty often. Teaching a model to adapt across tasks AND harnesses at this scale will really help future proof our models.

Not a new technique, but cool to see it become more and more the default.

[Fuli Luo @_LuoFuli]

Nearly half a year of silence. We spent it studying one problem: how far RL can scale.

MiMo-V2.6 is in the middle of its RL run right now. Three things we scaled: compute (~2B tokens per step, 1568 prompts × 16 rollouts, fully async), environments and harnesses (multi-task agentic RL, mixed across multiple harnesses in one run), and grader compute (agentic in-group credit assignment, with test-case and rubric-based rewards). We'll open-source the details piece by piece over the coming weeks.

Streaming the run: https://t.co/ZSxahzJRju

[Sinan Ozdemir @Prof_OZ]

@Madisonkanna @FireworksAI_HQ Welcome aboard!

[Sinan Ozdemir @Prof_OZ]

Watch me break this down in detail and if you have any questions about where you are on this path and how to move forward, come on by and ask away!

https://t.co/6KhnI1LipU

[Sinan Ozdemir @Prof_OZ]

Companies adopting AI walk the same path:

- Rent closed frontier models
- Engineer harnesses around them
- Move to open models for performance & efficiency
- Train logic & skills into weights

@harvey, @cursor_ai, @cognition know this. AI-native companies just walk it faster.

[Sinan Ozdemir @Prof_OZ]

@mervenoyann @nvidia They're lucky to have you on board!

[Sinan Ozdemir @Prof_OZ]

@sohailmo Fun! Writing a book is no small feat :)

[Sinan Ozdemir @Prof_OZ]

I gotta try this

[World Labs @theworldlabs]

Introducing Atlas:

The world's first multimodal world model that generates image and video frames with pixel-perfect camera control and reconstructs them in 3D.

Model the world, move the camera, and simulate space & time.

[Sinan Ozdemir @Prof_OZ]

Dataset: https://t.co/LmpqIHktTU
Code: https://t.co/8I1gbwWpHW

[Sinan Ozdemir @Prof_OZ]

Something I wish I got asked a lot: "What reasoning level should I use?"

Let's try on a computer use dataset:

- NO model did best at the max reasoning
- K3 and GPT 5.5 really overthink at max
- Opus did best at low

Test reasoning efforts, please!

data + code below

[Sinan Ozdemir @Prof_OZ]

@LotusDecoder @FireworksAI_HQ 🙌 love to hear you’re still working on it!

[Sinan Ozdemir @Prof_OZ]

@sophie07_h @FireworksAI_HQ @harvey Great question! One way is the rubrics decompose tasks into binary checks tied to deliverables making it harder to hack and they tested many LLM judges. They can be confident it worked because they saw gains on held out benchmarks. A hacked policy usually doesn’t generalize well.

[Sinan Ozdemir @Prof_OZ]

@ryancrowe something hardware, maybe robotics I bet but even that, Huggingface comes with Pollen Robotics (which btw I just bought that duck they put out 👀 very excited

[Sinan Ozdemir @Prof_OZ]

Nvidia buying Huggingface is going to be a huge shift for local models and open models in general. I’m still wrapping my head around how the world’s largest repository of models and datasets might change because of this.

[Sinan Ozdemir @Prof_OZ]

Congratulations on being one of our winners! I took a look and was very impressed :) I'm curious where you got the inspiration for this!

[LotusDecoder @LotusDecoder]

#KimiK3onFireworks

I built MedRecap, a home medication reconciliation assistant, with Kimi K3 on @FireworksAI_HQ.

Upload photos of your pill bottles, prescriptions, and visit notes, and it organizes everything into one medication list, flags discrepancies across the records, and uses FDA drug labels to surface questions worth confirming with a doctor or pharmacist.

All demo data is fictional. 👇

I originally planned to build an outdoor hiking guide. Then I discovered a mature product already doing it well, so I switched ideas close to the deadline.

The core of MedRecap was built by K3 through Cursor IDE Agent, covering the frontend, backend, image understanding, and report workflow. I spent about $20 on API usage.

What impressed me most is how capable open-source models have become at delivering a complete web app, and how stable the @FireworksAI_HQ API was across vision tasks and multi-turn workflows.

[Sinan Ozdemir @Prof_OZ]

@natolambert @radixark A gentleman and a scholar as usual 🤩 thank you my friend

[Sinan Ozdemir @Prof_OZ]

@natolambert @radixark I'm not in town but can I get a signed copy? :)

[Sinan Ozdemir @Prof_OZ]

@Ishita__Sethi 🙏 I’m glad you think so!

[Sinan Ozdemir @Prof_OZ]

Yesss this is excellent stuff! More examples on the cookbook are coming all the time for us (and our agents) to use

https://t.co/t5PupFEWow

[Terry Chen @227Terry]

Almost every agent product has an intent-routing layer: an LLM that looks at a task and decides what should handle it.

For a proof of concept, a prompt and an OpenAI API key may be enough. But once you are routing millions of production requests, cost and latency start to matter.

You could send every decision to the best frontier model and expect it to work. In our test, GPT-5.5 reached 91.7% routing accuracy with 1.19-second P50 latency, costing $2.40 per 1K requests.

To see how open source performs, I asked my coding agent to “vibe tune” a router using Fireworks’ Training Agent skill. It selected managed LoRA SFT and trained Qwen3-8B on 600 labeled examples.

Fine-tuning improved Qwen’s routing accuracy from 80.0% to 93.3%, with 0.38-second P50 latency. On this small holdout, its quality was effectively comparable to GPT-5.5, while latency was roughly three times lower.

The economics depend on scale. One dedicated H200 costs $5,040 per month, breaking even with GPT around 2.1M monthly routing requests. At 10M requests, tuned Qwen is about 4.8 times cheaper. Below break-even, GPT remains cheaper.

Closed APIs are often the fastest way to get started. But when production volume makes cost, latency, and control important, custom-trained open-weight models become compelling.