Pinned Tweet
Nvidia is reportedly in talks to acquire Reflection AI. On Vercel’s AI Gateway, open-weight share of tokens went from 7% in December to 56% in August, while taking just 14% of spend. By one day in mid-September, it hit 78%.
Anthropic shows the split. From June to July, its spend share rose from 61% to 64% while its token share slipped from 32% to about 30%. Pricing power is holding for now.
The usage data tells us an interesting pattern: open wins the token volume, closed keep the dollars.
The driver is agents looping in the background, planning, retrieving, and revising. Developers route that repetitive middle to open models at lower cost and save closed APIs for the calls where a mistake is costly.
So how would the token volume and dollars flow in the next few years?
Little Ornith be cooking! This time, the open weight is pretrained from scratch, achieved frontier performance, and can run on a single NVIDIA DGX. Congrats team!
Aloha!🌺 Meet Ornith-1.5-122B: No Tradeoff Between Strong and Fast
Stronger. Faster. Leaner. Ornith-1.5-122B delivers frontier-level capabilities with exceptional inference efficiency. With superior instruction following, stronger task decomposition, and deeper understanding of real-world workflows, it achieves state-of-the-art performance among open-source models of comparable size, while running on a single NVIDIA DGX card!!
Open weights under the MIT license.
Coding
✅SWE-bench Verified: 0.845
✅TB2.1: 0.813
✅CyberGym: 0.800
✅NL2Repo: 0.512
Agentic
✅HLE: 0.446
✅BrowseComp: 0.785
✅MCP-Atlas: 0.746
✅Toolathlon: 0.708
🤗Hugging Face: huggingface.co/ornith-ai/Orn…
We can’t wait to see what you build with Ornith-1.5-122B.
21x faster than Microsoft DeepSpeed and 7x than NVIDIA Megatron! Very exciting
The AI industry has a massive compute efficiency problem. At Physion Labs, we think it's time to challenge it.
Today, we're sharing the technology behind NexTron, our world-intelligence training engine.
Our benchmarks on matched hardware:
⚡ Up to 20.9× SFT throughput vs. DeepSpeed (H100)
⚡ Up to 6.6× SFT throughput vs. Megatron (H200)
⚡ Up to 7.2× SFT throughput vs. Megatron (B200)
⚡ Up to 3.1× end-to-end RL throughput vs. Miles–Megatron
🔥 180B model fine-tuning on just 8 H100 GPUs.
Same hardware. Dramatically higher throughput. Comparable model quality.
We believe the next frontier isn't just building bigger GPU clusters.
It's unlocking more intelligence from every GPU, every watt, and every dollar spent on training.
#PhysionLabs #NexTron #WorldIntelligence #AIInfrastructure
Prediction 1: The spend gap narrows. Closed labs lose per-token pricing power as open models match them on routine work, prices keep falling, and API bill shocks push enterprises to self-host and fine-tune open weights. But the spend pool itself restructures. Dollars relocate down to the metal and up the stack: to inference providers and neoclouds billing against GPUs, and to the application layer. Open weights become the most valuable strategic asset.
Prediction 2: This could slow if the frontier re-widens. Recursive self-improvement plus a compute edge could give closed labs escape velocity. They also own the product, see how people actually work, and train on that signal. Compute is the main bottleneck for open weight companies, and if it stops being binding, closing the gap is only a matter of time.
Frontier intelligence shouldn’t stay concentrated in a few closed labs. The game on Western open weights has just begun. I wrote down some thoughts on open weight AI and the frontier escape velocity. If you have your own predictions, I’d love to compare notes.
charlottexia.substack.com/p/…
Excited to be moderating a panel at our Continual Learning After Hours on Monday, featuring @jasonwu0731 (Jason Wu, Sr Director of @salesforce AI Research), @JiweiLi1 (Jiwei Li, Cofounder of DeepReinforce/@ornith_) and @kingsquare2 (Di Jin, VP of @nebiusai, Cofounder of Eigen AI).
Whether you're coming from @COLM_conf, making the rounds at #SFTechWeek, or just deep in the space, we'd love to have you join us! partiful.com/e/M451cLAPnXY0U…
Charlotte Xia retweeted
💡Ornith-1.5-DFlash
🤗 huggingface.co/collections/o…
Join us at Continual Learning After Hours next Monday Oct 6 to kick off #COLM2026 and/or SF Tech Week!
A few spots are open, RSVP in comment 👇
Join us next time!
Thanks to everyone who came out for a slower Sunday at Stanford and special thanks to Michael @mmmbchang 🐐 instructor!
We brought researchers and builders together for a morning of Pilates and brunch to unplug and connect. More mornings like this soon! 🧘
Charlotte Xia retweeted
🚀 Physion Arc 1.0 — Anime Track is now live:
physionlabs.ai/leaderboards/…
@openart_ai takes #1 overall, while @character_ai leads the objective evaluation.
We benchmarked frontier end-to-end video agents on ~1-minute anime creation across text-only, single-reference, and multi-reference tasks, using scripts and visual assets created with real anime directors — testing not just generation quality, but the ability to turn scripts and visual references into coherent anime sequences.
Anime is an especially demanding test of video agents: character identity, expressive acting, visual style, shot composition, pacing, continuity, and storytelling all have to hold together across an entire sequence.
Huge congratulations to @openart_ai and @character_ai for pushing the frontier of AI-native anime creation forward. 🎬🔥
Charlotte Xia retweeted
🐶🏁 Three models. One race.
We independently tested H3 Max by @fal, minimax H3 by @MiniMax_AI, and FastH3 preview by @haoailab @haozhangml @wlsaidhi across robotics, animation, movies, and ads.
H3 Max takes the overall lead 🏆. FastH3_preview, meanwhile, is a community OSS effort that’s already keeping up surprisingly well, even beating H3 on overall Prompt Adherence 🥳
The bigger gaps show up in Visual Integrity and Human Preference.
Full evaluation: physionlabs.ai/blog/minimax-…
🤖 Made with AI
Charlotte Xia retweeted
“Faster” and “better” are meaningless without context.
Physion Labs would be happy to sponsor an independent evaluation of both models—not to pick a winner, but to map the real quality–speed frontier and show which tradeoff is worth making for each use case.
Would both (@haozhangml, @isidentical) be interested in collaborating on this?
oh, thanks for evaluting this for us. We wish we had more compute to distill with more data (like you would have), but it is happy to see that, with our very limited dozens of GPUs, we are already generating good videos.
Also, I hope you could be less mean to us next time, as I believe fastH3 generates sufficient decent videos that many users can already use at much lower cost, definitely not a "random pixel generator"
Charlotte Xia retweeted
Congrats to the @ornith_ team on launching the Ornith-1.5 family!
Very cool work on models that improve themselves, and thanks for including SGLang in the model cards🧡
Cookbook is on the way. Excited to see what people build and run with them!
Commands👇
Aloha! 🌺Introducing Ornith-1.5, a family of open-source LLMs spanning 9B Dense, 35B MoE, and 397B MoE, trained with self-improving strategies.
It achieves state-of-the-art performance among open-source models of comparable size and delivers performance comparable to Claude Opus 4.8 across reasoning, agentic, and coding tasks:
✅Terminal-Bench 2.1 (86.1)
✅SWE-Bench (86 on verified, 65.1 on pro, 79.6 on Multilingual)
✅DeepSWE (56)
✅HLE (44.6)
✅ClawEval (81.4)
✅Tool Decathlon (71.2)
Ornith-1.5 takes a major step toward training foundation models through end-to-end self-improvement, extending the self-scaffolding strategies introduced in Ornith-1.0 into a more complete self-improvement loop: the model proposes new tasks, generates task-specific scaffolds, and produces solution rollouts for reinforcement learning, continuously creating new learning experiences from which it can improve.
All models, along with their quantized versions (FP8, GGUF, MLX, and NVFP4), have been released under the MIT License, enabling unrestricted commercial and research use.
📘Tech Blog: ornith.ai/ornith_1_5.html
🤗Huggingface: huggingface.co/collections/o…
Go Ornith 1.5! A major step toward end-to-end self-improvement.
Aloha! 🌺Introducing Ornith-1.5, a family of open-source LLMs spanning 9B Dense, 35B MoE, and 397B MoE, trained with self-improving strategies.
It achieves state-of-the-art performance among open-source models of comparable size and delivers performance comparable to Claude Opus 4.8 across reasoning, agentic, and coding tasks:
✅Terminal-Bench 2.1 (86.1)
✅SWE-Bench (86 on verified, 65.1 on pro, 79.6 on Multilingual)
✅DeepSWE (56)
✅HLE (44.6)
✅ClawEval (81.4)
✅Tool Decathlon (71.2)
Ornith-1.5 takes a major step toward training foundation models through end-to-end self-improvement, extending the self-scaffolding strategies introduced in Ornith-1.0 into a more complete self-improvement loop: the model proposes new tasks, generates task-specific scaffolds, and produces solution rollouts for reinforcement learning, continuously creating new learning experiences from which it can improve.
All models, along with their quantized versions (FP8, GGUF, MLX, and NVFP4), have been released under the MIT License, enabling unrestricted commercial and research use.
📘Tech Blog: ornith.ai/ornith_1_5.html
🤗Huggingface: huggingface.co/collections/o…
Charlotte Xia retweeted
We’re pleased to announce Charlotte Xia’s promotion to Principal at Fusion Fund!
@xia_char brings deep technical expertise, investment judgment, and a founder-first approach across #AI, #PhysicalAI, and #DeepTech. Congrats, Charlotte, on this well-deserved milestone!
Charlotte Xia retweeted
We’re hosting a breakfast for the ai x bio community in nyc on fri 24th
Join us to meet the @PheironAI team and discuss probability of success, predictive validity and human evidence
Fun + food guaranteed!
Ending ICML with a great workshop talk from Noam Brown (@polynoamial):
- Plateau in test-time compute is very far away
- Multi-agent system significantly speed up execution. Infra for multi-agent system is non-trivial to support parallel scaling capability
- Rethinking eval: time, cost, token latency should all be plotted on x-axis; Safety eval is currently broken
- Transformers with more juice can take us to AGI (!!) Though alternatives are still worth exploring because why is AGI the endpoint
Nexa AI has been acquired by @Qualcomm! What an incredible journey to witness, from proving frontier AI could run on any device, to Octopus, NexaSDK, and Hyperlink, and now Qualcomm AI Hub GenieX. Watching this team grow and hit remarkable traction has been a privilege. Congrats to @alexchen4ai, @zacklearner, @alanzhuly and the whole Nexa team.
