We’re Handshake AI 👋🏼
We work with frontier labs on the expert data, environments and evaluations their models train on. We publish the benchmarks. And we bring the same approach to enterprises, so they can build AI on their own experts’ judgment.
We post-trained Qwen3.8-Flash-Next on a subset of our Terminal-Bench RL environments. The model improved on Terminal-Bench 4 from 11.6% to 15.7% across 4 domains: security, machine learning, software, and operations.
Great work by @ulyatkch, @ErrkC, and others on the research team. Thank you to @appliedcompute for the collaboration!
RL environments like these target widely applicable capabilities and are essential for AI teams looking to give their intelligence an edge.
Our research team post-trained Qwen3.8-Flash-Next with GRPO in the mini-swe-agent harness. Just a fraction of @joinHandshakeAI’s coding RL environments improved the model from 11.6% -> 15.7% on Terminal-Bench 4 (beating GPT-6 Luna Max).
More details:
https://t.co/9tRbdT4USB
@WesAField at Reuters today: "Start benchmarking and evaluating your work... Author your own future. You can do that tomorrow, and you will be miles ahead."
Companies that do this, especially in finance and healthcare, will move AI from pilot to production.
Just wrapped a fireside chat at Reuters Momentum AI Finance 2026 in New York. Thanks Richelle and the Reuters Events team for a great conversation with a bit of levity.
Benchmarking and evaluating enterprise work was the headline. Enterprises should put in the foundational layers of codifying what the enterprise does, why it’s valuable, and how the enterprise thinks about the world. This moves an AI pilot to production.
Post-training Qwen3.8 27B on Handshake AI’s life science data raised its accuracy on the VIALS benchmark from 20.5% to 27.1%. Better than GPT-6 Luna High and 6.2pp higher than Kimi K3 Medium at 50 times lower cost. Thank you to @Osmosis_AI for the collab.
Handshake AI’s data and RL environments help research teams bridge the gap to the frontier.
Post-training Qwen3.8 27B on @joinHandshakeAI life science data raised accuracy on life science tasks by 7% (~20% -> 27%) on ATLAS-VIALS bringing it to the Pareto frontier of OSS models at 1/50th the cost of Kimi K3 and with similar accuracy as GPT-6 Luna.
The gains transfer to other benchmarks, e.g.
LAB-Bench FigQA: 75.3%→77.2%
MicroVQA: 63.8%→68.3%
PerceptionBench: 53.8%→54.8%
HLE multimodal: 17.5%→19.3%
We saw similar gains on Nemo where our data doubled it's accuracy on VIALS. Full blog post here: https://t.co/DTin0xH0B8
See more on ATLAS VIALS on @ArtificialAnlys's page here: https://t.co/gPYZRl9hon
@thinkymachines put two of our benchmarks on the Inkling 2 Small card.
ATLAS-Finance, deal tasks built by bankers inside a simulated firm: 6.0%.
VIALS, reading gel blots and plasmid maps from biotech labs: 34.6%, or 57.8% with tools.
@jomulr on his take 👇🏽
Despite controversy clouding American open-source AI, the models are finally starting to shine.
Thinking Machines' Inkling 2 Small model card reports strong results on two of our benchmarks:
- ATLAS-Finance: Can agents complete long-horizon assignments in banking environments?
- ATLAS - Visual Life Sciences (VIALS): Can multimodal AI interpret scientific artifacts from biotech/pharma workflows?
Awesome to see this progress in high-value use-cases, and that Inkling is open-sourced under Apache 2.0!
https://t.co/nlAl9V3H6W
@Sbhaiwala03, Handshake AI’s Chief Strategy and Innovation Officer, predicts on @latentspacepod that the data market for AI will increase 10x over the next three years.
High quality data is the main fuel driving the growth.
Artificial Analysis ran our VIALS benchmark as an agent eval, VIALS-AA.
Give a model tools, and scores nearly double, from about 25% to 47%.
Even then, a scientist still finishes these tasks faster and more accurately.
Take away its tools, and Claude Fable 5.2 scores below Claude Haiku 5.5 with tools on VIALS-AA; across models tool use nearly doubles scores, from ~25% to ~47% on average
VIALS-AA is our implementation of @joinHandshakeAI's VIALS benchmark for visual reasoning in life sciences research. By default, models work as agents: they think, run code and view images in a sandbox. We reran VIALS-AA in a one-shot format: one request with the question and images, and no tools.
Lower cost models in an agent harness can beat single-shot large models on both cost and performance in visual reasoning and interpretation of scientific artifacts.
Claude Fable 5.2 (max) leads VIALS-AA at 67% with tools, at a cost of $4.20 per task; without tools it scores 44% at $0.73 per task. Claude Haiku 5.5 (max) scores 49% with tools, at $0.07 per task: higher than Fable without tools, at about a tenth of the cost.
Our VIALS benchmark reveals clear weaknesses of current vision encoders. Human scientists can accurately do these tasks in minutes without using tools.
While it's good that agents with the right tools can perform significantly better in these VQA tasks, I believe this indicates we have a long way to go in computer vision research. For instance, I think Astra represented a fundamental breakthrough in computer vision when it was released (only VLM to reach 50% on VIALS), and we need more fundamental breakthroughs.
Generalization in image space seems to be harder than in text space, thanks to the fact that society has compressed the world into language so effectively.
We’re honored to be part of the inaugural 2026 AI 100, @NewcomerMedia and @BessemerVP's ranking of the top 100 private AI companies in the world.
The credit goes to the team and the experts in our network who do the work.
@latentspacepod asked @NickMagazine1 how good models need to be before a lab would trust them without a human checking. His answer: somewhere under 90% is unserviceable, because getting this wrong has an economic cost and sometimes a public health one.
On VIALS the best frontier models are under 60%.
Models are confident narrators, which can be a good thing.
When does it become a problem?
When they’re confidently wrong.
Tune in to @jomulr and @NickMagazine1 👇🏽
In this example, @NickMagazine1 goes through Opus 5.5 analyzing an image from a biopsy. In this case, the model chose the wrong answer, and that wrong answer would cause harm to a real patient.
It’s a great demonstration of why using AI models in medicine can be so high-stakes.
Handshake AI now has its own home on X!
Follow for our research and benchmarks, our work with frontier labs, and the enterprise AI we’re building on top of it. 👇
@joinHandshakeAI