Pinned Tweet
Congratulations to @nvidia on the release of Nemotron 3 Ultra!
Ultra now leads GDPval-AA by a good margin among US open source models. @AfterQuery is proud to have provided the GDPval training data used to hill climb the benchmark.
We added support for SDPO in SkyRL alongside UC Berkeley SkyLab researchers. Our new blog explores self-distillation across math, document, and chart-to-code tasks.
On MATH-500, Qwen3-8B improved from 85.6% to 91.2% (+5.6 points), while using 8× fewer rollouts than GRPO. afterquery.com/blog/on-polic…
In our chart-to-code runs, EMA teachers declined while frozen teachers stayed stable.
Separately, top-100-plus-tail KL reached comparable scores in about half as many updates as sampled-token KL.
1
6
136
Training data is increasingly produced by systems where humans and models work together. That means data labs must be able to show exactly what produced their data and why it should be trusted.
Today we're publishing our closed-model policy, how we enforce it, and a call for responsible training data production. afterquery.com/blog/a-call-f…
Today we're releasing CrossSearch, a new benchmark testing whether AI agents can do something people do at work every day: look through their files, search the internet, and reason through ambiguity.
Unlike benchmarks that test web search or file understanding separately, CrossSearch tests whether agents can do both together to solve real-world problems.
Across 100 tasks and 14 models, Gemini 4 Argon leads at just 64% pass³.
Full blog, leaderboard, and dataset in the replies.
Data Set: huggingface.co/datasets/Afte…
Blog: afterquery.com/blog/crosssea…
5
539
Can AI agents run a business as conditions change?
Introducing StoreBench, a high-fidelity RL environment where agents manage a simulated online apparel store, with safeguards against reward hacking.
StoreBench includes 12 evaluation tasks spanning 30–365 simulated days. Across the 11 tasks lasting 30–45 days, all seven models score lower on average than human participants and smart-triage, a rule-based program that prioritizes urgent work.
Late restocking, unresolved refunds, and decisions based on outdated information show that sustained operations remain challenging for AI agents.
Check out the comments to read more.
GitHub: github.com/AfterQuery/storeb…
Paper: arxiv.org/abs/2610.10942
Blog: afterquery.com/blog/storeben…
1
4
657
AfterQuery retweeted
Congrats to @cognition on SWE-2 Flash!
Exciting to see an increasing number of companies train models around the work they know best.
Strong open foundations like @thinkymachines' Inkling-Small enable teams to shape intelligence around their own needs.
Excited to have one of our post-training runs showcased by @thinkymachines.
Using Tinker, we post-trained gpt-oss-20b with SFT, then RL on our terminal agent data. Its Terminal-Bench 2.0 pass rate went from 3.1% to 17.0%.
@AfterQuery datasets are validated with Tinker against held-out benchmarks before they ship.
If you're building frontier models and want data with measured impact, reach out at [email protected].
AfterQuery retweeted
Can collecting robot data be as easy as 𝗽𝗿𝗲𝘀𝘀𝗶𝗻𝗴 𝗿𝗲𝗰𝗼𝗿𝗱?
Introducing UMI-Live, a 𝗰𝗼𝗺𝗽𝗹𝗲𝘁𝗲𝗹𝘆 𝗼𝗽𝗲𝗻-𝘀𝗼𝘂𝗿𝗰𝗲 bimanual UMI we made that takes you from handheld demonstrations all the way to a trained robot policy.
It's built to be easy to use. Each gripper has its own camera, LiDAR, compute, and battery, so there are no base stations to mount and nothing to calibrate in the scene. You press one button and both hands start recording.
While you keep collecting, our pipeline processes finished episodes in parallel and turns them into LeRobot datasets, automatically uploaded on Huggingface. VLAs fine-tuned mostly on this data succeed in 𝟭𝟬/𝟭𝟬 (π0.5) and 𝟴/𝟭𝟬 (MolmoAct2) trials on a real bimanual task.
We're open-sourcing everything, from the hardware designs and build guides to the code and ~40 hours of data 🧵
We improved models on professional work tasks using NVIDIA AI’s xToken, NeMo Gym, and NeMo RL for cross tokenizer distillation.
Read the full blog below.
We used @NVIDIAAI's xToken recipe along with NeMo Gym and NeMo RL open post-training libraries to improve models on professional work tasks through cross tokenizer distillation.
Held-out GDP.pdf strict pass@1 (avg 5) in NeMo Gym:
- Nemotron 3.5 Super VL: 6.67% to 11.83%
- Qwen3.5-9B: 4.40% to 8.20%
Read the full blog below.
Full blog: afterquery.com/blog/xtoken-d…
1
542
AfterQuery retweeted
Huge congrats to the @GoogleDeepMind team on Gemini 4 Carbon!
Impressive to see performance climb across many core capabilities.
I did say we'd be releasing our models periodically...introducing Gemini 4 Carbon, our next frontier model.
Following Argon's release just two weeks ago, Carbon shows the pace at which we’re pushing the frontier, it ranks #1 on the recently updated v5.0 of @ArtificialAnlys Intelligence Index at 59. It also leads across key cybersecurity evaluations scoring 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching.
We’re seeing significant improvements in long-horizon tasks even compared to Argon, and we’ve strengthened frontier safeguards.
As we did with Argon, we’re going to roll it out gradually + responsibly. Some trusted cyber defenders will get access starting this week. We'll be expanding it to more partners through our Fairwind Program, ahead of wider availability soon, so hang tight!
Congratulations to @Recursive_SI on the release of Flint!
Flint runs locally on your laptop and competes with models that require 10x the compute.
@AfterQuery is proud to support the @Recursive_SI team.
Today, we release our first model, Flint-40B-A2B, a fast local coding model built with our automated research system. huggingface.co/RecursiveSI/f…
1. Strong performance. 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on SWE-bench Verified.
2. It is trained from scratch (tokenizer->Pretrain->Midtrain->SFT->RL).
3. Fast in your local laptop. Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark.
4. We build the complete model training pipeline in 4 month!
Frontier models nearly run through the public internet for pre-training alone. The knowledge they need next is locked inside businesses.
Today we're launching @AfterQuery Data Partnerships.
If you're a business owner who is curious, or hesitant, about licensing your data, we would love to hear from you.
Learn more in the replies.
Read the full blog: afterquery.com/blog/announci…
AfterQuery Data Partnerships: afterquery.com/data-partners…












