Applied research lab curating data solutions to accelerate foundation model development.
Joined February 2025
- Tweets103
- Following11
- Followers4K
- Likes214
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.
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Late restocking, unresolved refunds, and decisions based on outdated information show that sustained operations remain challenging for AI agents.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…
2
312
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 research@afterquery.com.
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.
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…
This past week at GTC Berlin, @AfterQuery was one of the companies highlighted in the Nemotron partner ecosystem during @JensenHuang's keynote.
We're proud to be an official @NVIDIAAI post-training partner and look forward to supporting more enterprises across their entire post-training strategies!
Read the full blog below.
Full blog: afterquery.com/blog/post-tra…
3
383
Congrats to @spectorb @amspector100 @aidanmantine and the @flappyairplanes team on the $1B Series A!
Excited to see what their work on GPU systems and learning algorithms unlocks next ✈️
@AfterQuery just moved an open model past the cost-accuracy frontier in finance.
We post-trained @NVIDIAAI Nemotron 3.5 Super VL on expert credit tasks with @hebbia.
Perfect solved: 47% to 74%. That beats Sol 6, and lands within 3 points of Astra 6, at about 1/7 the cost.
Link to the full case study in replies.
Not all coding data is created equal.
Using only 500 tasks from @AfterQuery's SWE agent dataset, our internal post-training team was able to improve Qwen3.8-27B-Medium by 11.3 points in only 15 steps of GRPO.
Read more about our training experiment in the blog post linked in the comments.
Read more about our experiment using SkyRL here: afterquery.com/blog/improvin…
To learn more about our tasks and reward design, reach out to us at research@afterquery.com
Congratulations to the @motif_tech team on the release of Motif 3!
AfterQuery is proud to have served as the sole data partner on this model.
This is a historic moment for Korean sovereign AI and the open weights ecosystem writ large. We look forward to supporting Motif and the Korean AI ecosystem as they push the frontier of open intelligence.
Motif 3 leads its comparison set on τ³-Banking, and posts 94.7 on τ²-Bench Telecom, 74.9 on Terminal-Bench 2.1, and 76.2 on SWE-bench Verified. It’s also great to see @nvidia's open foundation put to use here - Motif 3's post-training was done via NeMo-RL.
In a world where it’s easy to assume that the AI race is already won, the team at Motif is proof that the field is open, and we’re still at the starting line.
Congrats to the @WeAreLegora team on the release of the Legora BAR, a benchmark for agentic legal work built on 5,161 real law firm cases across 28 practice areas!
@AfterQuery helped QA the benchmark with Legora. Proud to work with their team on measuring the frontier of agentic legal work.
Link to the full blog post in the replies.
Join our team.
The best operators at @AfterQuery chose us over Wall Street, and we’re hiring more of them.
They’re business or CS majors who left Goldman, JPMorgan, and Jane Street because they believe this is the most consequential work they could be doing right now.
If this is you, apply at the link in the comments or leave your email below.
Know someone cracked? Tag them - we pay $10,000 for successful referrals.
Congrats to the @PoolsideAI team on Laguna S 2.1!
With thinking enabled, Poolside reports scoring 70.2% on Terminal-Bench 2.1.
Today we're releasing Laguna S 2.1, our most capable model to date.
It's a 118B total parameter Mixture-of-Experts model with 8B activated per token, a context window of up to 1M tokens, and thinking and no-thinking modes.
Capable enough to hold its own against models many times its size. Small enough to run on a single @NVIDIAAI DGX Spark.
Laguna S 2.1 is fully open under OpenMDW-1.1, with weights available today on @huggingface
poolside.ai/blog/introducing…
Last Thursday, we hosted Poker Night at our SF office.
Researchers, founders, and builders joined the @AfterQuery team for 3 things:
1. Poker.
2. Drinks.
3. Lively conversation.
We're hosting more events soon.
Comment if you want to join the next one ↓
See what your codebase could be worth.
Introducing Atrium by @AfterQuery.
We’re paying companies for their private codebases. Connect a repo, get an automated quality analysis in minutes, and receive a quote in days.
Link in the replies.
See what your codebase could be worth: atrium.afterquery.com/
Kimi K3 ranks #1 on @AfterQuery's SpreadsheetBench 2, surpassing Claude Fable 5.
An open weight model now outperforms all closed-sourced models. Read more in the Kimi K3 blog and SpreadsheetBench 2 paper linked below.
Congrats to the @kimi_moonshot team on the incredible model!
Kimi K3: kimi.com/blog/kimi-k3
SpreadsheetBench 2, which AfterQuery helped build: arxiv.org/pdf/2606.29955
Congrats to the @thinkymachines team on the release of Inkling!
Inkling is a fully open-weight (Apache 2.0) 975B-A41B MoE model with native support for text, image, and audio inputs, featuring a context window of up to 1M tokens. It leads all U.S. open-weight models on @ArtificialAnlys’s Intelligence Index.
We’ve tested it internally and seen strong performance across reasoning, coding, and agentic tasks.
Inkling has day-one support on Tinker, and we're excited to start running experiments with it.
AfterQuery is proud to have supplied data that helped train the model.
Today, we are introducing Inkling.
Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available.
thinkingmachines.ai/news/int…
Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. 🧵
