Pinned Tweet
After 7 months on the job market, I am happy to announce:
- I joined @allen_ai
- Professor at @CarnegieMellon from Fall 2025
- New bitsandbytes maintainer @Titus_vK
My main focus will be to strengthen open-source for real-world problems and bring the best AI to laptops 🧵
- I joined @allen_ai
- Professor at @CarnegieMellon from Fall 2025
- New bitsandbytes maintainer @Titus_vK
My main focus will be to strengthen open-source for real-world problems and bring the best AI to laptops 🧵
156
87
2,380
280,083
Tim Dettmers retweeted
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
🔹 Full-project coding: games in Unreal, Unity & Godot.
🔹 Excels at parametric 3D modeling in Blender & CAD.
🔹 More reliable at finance with hallucination down by 65%.
🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Open weights by Nov 4, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.ai/blog/kimi-k3-1…
🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
🔹 Full-project coding: games in Unreal, Unity & Godot.
🔹 Excels at parametric 3D modeling in Blender & CAD.
🔹 More reliable at finance with hallucination down by 65%.
🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Open weights by Nov 4, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.ai/blog/kimi-k3-1…
1,257
3,234
30,183
4,058,025
Tim Dettmers retweeted
Very excited to share that we've been selected to receive NSF funding through the NSF-STRIDE AI Efficiency Challenge!
People are increasingly using LLMs to analyze large datasets, but doing so with current AI models and systems is often extremely expensive, slow, and inefficient. We need better infrastructure to make this kind of analysis practical at scale.
I am hiring PhD students and postdocs to work in this area starting in 2027—prospective PhD students, please apply at the CMU website, and prospective postdocs, please send me an email.
People are increasingly using LLMs to analyze large datasets, but doing so with current AI models and systems is often extremely expensive, slow, and inefficient. We need better infrastructure to make this kind of analysis practical at scale.
I am hiring PhD students and postdocs to work in this area starting in 2027—prospective PhD students, please apply at the CMU website, and prospective postdocs, please send me an email.
18
15
206
13,789
Tim Dettmers retweeted
For those who are like me trying not to drawn in this AI perturbance, @Tim_Dettmers shared amazing inspirational grounded in reality notes on how to survive this and thrive next:
timdettmers.com/2026/09/21/dla…
timdettmers.com/2026/09/21/dla…
Thanks a lot to @jeffra45 for sharing this with me.
timdettmers.com/2026/09/21/dla…
timdettmers.com/2026/09/21/dla…
Thanks a lot to @jeffra45 for sharing this with me.
0
8
61
4,183
This is a very big deal. DeepSeeks inference system is on par with OpenAIs stack and better what Anthropic is running (at least for their user size).
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/De…
1/4
17
59
781
78,705
Tim Dettmers retweeted
Today, Dream Labs is coming out of stealth with Dream-1, a robot foundation model that controls bimanual robots it has never seen, on tasks it was never shown, in places it has never been, from $3/hour.
Dream API is live for everyone, today. 🧵
Dream API is live for everyone, today. 🧵
150
181
1,656
612,789
Tim Dettmers retweeted
Tons of bangers and useful advice for undergrads and grad students trying to make sense of how to navigate academia and entering industry in the age of agents in @Tim_Dettmers recent blog.
Highly recommend people read it.
timdettmers.com/2026/09/21/dla…
Highly recommend people read it.
timdettmers.com/2026/09/21/dla…
12
78
948
98,416
This is really cool work on our synthetic hospital benchmark. For patient diagnosis, this shows that small models + SFT + the right tools > GPT-6 Luna. Interesting is that the tools help small models, but not big models. So gaps in capability can be replaced by tools
spent a few days trying to train a tiny model to diagnose patients on the Synthetic Hospital environment out of CMU.
RL didn't help much, but some SFT and a simple search tool got qwen3-4b to perform as well as a ~1T param model
this is because the agent is more than just the model.
🧵:
20
16
186
17,532
Tim Dettmers retweeted
🚀 DeepSeek-V4.1-Flash-1017 is here!
Stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API.
🛠️ Optimized for @DeepSeekHarness, Tencent CodeBuddy, grade to DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the new model.
1/4
Stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API.
🛠️ Optimized for @DeepSeekHarness, Tencent CodeBuddy, grade to DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the new model.
1/4
1,626
4,129
37,168
8,093,130
Tim Dettmers retweeted
Excited to introduce Scout 1.0, the model powering Bits Investigation, Datadog’s AI SRE agent for complex troubleshooting and late-night alerts.
We post-trained GLM-5.3 Flash to match frontier models at much lower cost on our internal AlertEval benchmark. 🧵
We post-trained GLM-5.3 Flash to match frontier models at much lower cost on our internal AlertEval benchmark. 🧵
8
27
116
8,480
Tim Dettmers retweeted
🚢Marin 535B-A23B training has crossed the halfway mark! Some observations and learnings so far:
23
99
1,073
71,932
We built an all-synthetic simulation of a hospital that can be shared publicly without any issues. Our data quality is so high, physicians cannot tell the difference between real/synthetic patients. The data is fully verified -- perfect for RL.
Introducing Synthetic Hospital: an open, fully synthetic longitudinal EHR benchmark with verifiable ground truth!
1,268 patients, 5,602 encounters, zero PHI. Physicians could not reliably distinguish its charts from real ones.
📄 arxiv.org/abs/2609.30027
💻 github.com/sparkcpark/syn…
✍️ sparkcpark.github.io/posts/finding-…
26
85
1,157
143,787
Tim Dettmers retweeted
Excited to share Quail, our new open source AI-SQL engine (a collab with Modal)! By planning queries and LLM inference together, it reaches 1B+ input tokens/min on one H100 for one query 😱🚀
AI-powered data operators create a new, interesting inference workload👇
AI-powered data operators create a new, interesting inference workload👇
85
164
1,644
380,249
Tim Dettmers retweeted
Marin believes strongly in the power of the open community. There are so many datasets that have been produced, living on Hugging Face. We have built scalable infrastructure that ingests this data, performs deduplication, decontamination, mixing, producing 25T tokens.
There’s an enormous amount of open training data on Hugging Face. What does it take to make it work together 🤗?
For Marin’s 535B run, we built on 25T tokens from 152 datasets with licenses permitting training.
Here’s the work between downloading those and training a model 🧵
13
28
358
37,968
Our 2nd release in our open-source week.
CliffCompaction lets you run sessions for millions of tokens and save 50% on token costs. It integrates easily with Claude Code and Codex. We have been using it for months, and it has better vibes than Codex/Claude Code compaction.
CliffCompaction lets you run sessions for millions of tokens and save 50% on token costs. It integrates easily with Claude Code and Codex. We have been using it for months, and it has better vibes than Codex/Claude Code compaction.
CliffCompaction, our autocompaction tool for coding agents, is out!
Sessions run for millions of tokens, agents stay on track, and costs drop by up to 50%.
Paper: arxiv.org/abs/2609.26779
Code: github.com/nguyenvuthient…
PyPI: pypi.org/project/cliffc…
Blog: nguyenvuthientrang.github.io/cliffcompactio…
22
21
178
15,196
Cliff compactin is very easy to install and use with Codex, Claude Code and is integrated in bitsandbytes code our harness for autnomous research.
Install: pip install cliffcompaction
Use: cliff enable
It works by managing a separate compaction state between you and any OpenAI/Anthropic API
Install: pip install cliffcompaction
Use: cliff enable
It works by managing a separate compaction state between you and any OpenAI/Anthropic API
2
1
7
956
This research has been driven by my students Trang (@thientrangngv ) and Eulrang Cho. They did absolute fantastic work. Also check out Trang's blog and video -- they are a piece of art!
blog: nguyenvuthientrang.github.io/cliffcompactio…
blog: nguyenvuthientrang.github.io/cliffcompactio…
3
0
9
879