Today we’re releasing two decision models, elephant-1 (9B) and mammoth-1 (27B), trained by the @baselabs team, available for free for one week on our Model APIs. mammoth-1 is #1 on the Jev Decision Index and elephant-1 is tied for #1 on the Artificial Analysis Decision Index. It’s a new model, but more importantly, this is an open-source model you can access, deploy, and post-train on Baseten (or wherever you do your best work). The free week ends 10/27/26 at 11:59 PM PT. Try it and let us know what you think. Try mammoth-1 here: https://t.co/x8oDALMuyZ Try elephant-1 here: https://t.co/fmCGb3eGth

+1 more

Today we are launching @ObsessionDB, the realtime data warehouse for agents. Your agents are hungry for data. Feeding them is getting expensive. Your data warehouse wasn't built (or priced) for that. We learned this the hard way. We scaled our last company to petabytes on ClickHouse and watched the infra bill go vertical. So we built @ObsessionDB . The ClickHouse engine, rebuilt with compute and storage separated, so your infra scales with your agents instead of your bill. 100x faster and 20x cheaper than Snowflake and Databricks. Already live, processing trillions of rows a day. Agents are just getting started...

Won 1st place at the @ycombinator hackathon! 🏆 Most agents don’t have access to a wallet or credit card, so how can they pay for things? To solve this @KonradsKrypto, @muaazshafiq6, and I built D402, a way for agents to do work to pay for things! Instead of requiring a wallet, D402 lets agents complete useful tasks that can be monetized to pay for access, allowing them to get past HTTP 402 paywalls without spending money directly. Every agent has tokens, but not every agent has a wallet. Huge thanks to @nemild for hosting this hackathon! P.S. We are in SF for the week looking to chat with founders. Reach out! 🙌

A step-by-step guide to building a TTS training dataset at the scale of tens of millions of samples. Over the past few months, we've been working on ways to collect and build high-quality TTS training datasets at scale. Compared to collecting text data for LLMs, working with speech data is significantly more complicated, with plenty of engineering challenges along the way. There's surprisingly little public discussion about this, especially when it comes to the practical challenges of collecting speech data at scale. So I wanted to share what we've learned over the past few months: the problems we've run into, the lessons learned, and how we designed and built our data engine from scratch. Hopefully, this will be useful to others working on TTS and large-scale speech datasets. We're also hoping to open-source the full codebase down the road so others can build on it. I also put together a video to make the article easier to follow.

🚨New paper: Reward Models (RMs) are used to align LLMs, but can they be steered toward user-specific value/style preferences? With EVALUESTEER, we find even the best RMs we tested exhibit their own value/style biases, and are unable to align with a user >25% of the time. 🧵 https://t.co/IcLKW6Fjor

Their 80% horizon is also about 15 minutes, so I'm inclined to believe this is a real update on the surprisingly short length of tasks AIs are able to do independently with reasonable reliability on real AI R&D.

I've made some big improvements to the PDF reader on Papers with Code! > The AI assistant now cites its sources, full RAG-style > This is powered by Chandra-OCR 2 by @datalabto which is openly available on @huggingface > This model returns bounding boxes besides markdown > Allows you to click on citations and highlight sections Try it out on any PDF, like https://t.co/u2ikzY3bdv

🌲 Tree packing has landed in AsyncGRPO in TRL! ⚡ 26× less trainer compute to reach the same reward on agentic RL. GRPO starts from the same prompt G times, and forked multi-turn rollouts share most of their history. Sequence packing forwards those shared tokens once per sample. Tree packing forwards each unique token once. The first version used FlexAttention, but the shipped one is a single FA3 varlen call. Each token retains its own target, advantage, and previous log-prob. Extensively tested against sequence packing on dense, sliding-window, softcap, MoE, and VLM models. When it can't pack anything, it costs ~1%. On a REAL agentic RL run on MiMo Envs, tree packing reached the same reward with 26× less trainer compute 👇

Introducing Drex for Search 🔍 Supercharge your agentic search with Decision-powered retrieval Standard file search holds AI agents back. Drex changes that: • 62% fewer steps to find answers • 56% tokens saved • +4% performance boost across Claude Code, Codex, OpenCode, Hermes, and PI Validated across 6 professional work benchmarks like OfficeQA (529 real-world queries) Upgrade your agentic workflows with the File Search API Try for free: https://t.co/vdEavE1dsd Faster agents. Cheaper runs. Sharper accuracy.

The wonderful “Don’t just Fix it in Post” paper identifies predicting memorization as a key open problem. A progress report: we can now accurately predict verbatim memorization using just three well-understood variables. Experiments by Jing Huang, using the Hubble model suite: https://t.co/BdnqtoeX45

AlphaGo and coding agents need reasoning, so does robotics, which is arguably orders of magnitude harder. But what is the right representation to perform reasoning? Excited to introduce DreamCan, led by @GChongkai ! DreamCan performs sparse visual planning, reasoning over sparse visual frames that are both **reachable** by the robot and **makes progress** towards the task goal, and it can learn this capability from action-free suboptimal videos! Check out Chongkai’s post for more details 🧵⬇️

AlphaGo and coding agents need reasoning, so does robotics, which is arguably orders of magnitude harder. But what is the right representation to perform reasoning? Excited to introduce DreamCan, led by @GChongkai ! DreamCan performs sparse visual planning, reasoning over sparse visual frames that are both **reachable** by the robot and **makes progress** towards the task goal, and it can learn this capability from action-free suboptimal videos! Check out Chongkai’s post for more details 🧵⬇️

I gave Qwen3.8 27B one prompt and one hour with /goal in Hermes Agent on my RTX 5090, and it built Mont-Saint-Michel (yeah, I'm French) as a voxel diorama. this is the model I run every day. > the tide rises and falls around the mount on a one-minute loop > pilgrims on the causeway, sheep on the salt meadows, gulls overhead > T speeds up the tide, W brings a rain shower, N turns it to night with the windows lit a video shows what a model builds when nobody steers it for an hour. I think it stills need some tweaks and improvements but it is actually very fun to build. reference image in the first comment.

Genie 4 brings realistic dynamics and sound to realtime interactive world models! 🧞🚀🎶 This was the last project I worked on at Google 🥲, and the quality and dynamics are still so much better than the 3D creations from the latest frontier models. Bridging this gap to enable accurate, efficient, and high-quality simulations of the real world remains an open research challenge that is key to unlocking useful physical intelligence. Congrats to the whole team, and have fun exploring with Genie 4!

We are releasing Echo-3 today. Its the best model in the world at 3D world reconstruction. Believe it or not, this world below was generated from just an iPhone capture, no specialize capture hardware needed. You scan a space with an iPhone and Echo-3 reconstructs it in super high quality while keeping metric consistent. Think of the possibilities. Try it today at with our app @SpAItial_AI

You can now run pplx-decider-flash-v1.2 locally on your Mac, without an API key. Decider for Mac is an app that runs the 4B decision model on device. Ask a question about text or an image and get a probability back, all on your machine. https://t.co/1BJ4Qv3Urr https://t.co/raBBZ26kbC

Introducing Flora, an open model suite for scaling robotics data engines. It watches robot and human video and annotate what happened (timestamped instructions) and how well it went (rewards and failure reasons). 🔗 https://t.co/yZUS5pHvv9 We use it to train NVIDIA GR00T. 🧵

AI shouldn’t be locked to one hardware vendor or chip. To democratize AI compute, we’ve made MAX source available with Modular 26.7. 🧵

You can animate flowing water in splats now with more splats. And yes it also works with lava 😂 https://t.co/oN8FMYqxon

.@HamelHusain ranked every AI eval method from worst to best. This point stood out to me: You need a human subject matter expert to review your agent's output. Your engineers can't do it (unless you're building an engineering agent). We're all building self-verification into our AI engineering loops. You get AI to prove to itself that what it built works or needs iteration. But there are some things that require judgement or taste from a human. Anything subjective such as marketing materials or GTM strategies.

Gave a talk at PTC on flex_gemm which is a new gemm_epilogue API in pytorch built on Quack. I think its neat, not ground breaking but neat nonetheless :) https://t.co/rkE8i90qI1 planning to bootstrap this and continue to add fun examples

i've found defects in ~65% of the TAU-banking tasks. these are tasks where the policy isn't enforced correctly. for example, if there are rules the model can violate and not get penalized for (transferring to human reqs, tasks where the tool call can fail and the model is still rewarded, etc). some grading may be directionally correct, but still contains noticeable defects that needs to be hardened prior to use. high-level defects for people working on benchmarks right now: > models are capable of taking many out of scope actions without being penalized (only tool-call checks, no db) > models can directly lie to the customer with no penalty (no communication checks, which for a customer service benchmark feels important, though i didn't label tasks that only fail this as p0) > many of the tool call checks are naive and if the tool call fails there are no repercussions. i.e. a test of email change may just check if the tool was called, with no field checks or follow through check, so if the tool call fails or if i change the wrong email it may still pass. to expound a bit more: (1) i'm going to try and go through every public benchmark. the reason for these audits is that i've become frustrated with people using defective benchmarks to choose what type of data they want to purchase and data co's making flawed data to help a model hill-climb a benchmark. it is very clear that bad RL data is a contributor to reward hacking, and the benchmarks that inform purchasing decisions lie upstream of that, thus handing a huge power to companies that publish benchmarks. (2) sierra and the TAU series have had tremendous contributions to tool-calling evals. tau-bench is actually what got me into RL data in the first place more than two years ago. this is in no means to discredit sierra and the effort they've put into creating the benchmark. big inspiration! (3) please reach out and correct me if i'm wrong on any of these asap! i want to give sierra credit where it is due.

Meet Fibo 2.0 from @bria_ai_ 🔥 The #1 *open weights* model in image editing is supported in Diffusers (day-0, of course)! 1. JSON-native control 2. Pretty good text rendering 3. Comes in two variants: Turbo and Dev Check out the model here: https://t.co/gPdrEG5JpN https://t.co/R2pyWGG6gQ

Introducing Metaphor: The most cost-efficient cloud coding agent. We outperform Claude Code and Codex on Terminal-Bench 4.0 while achieving a 95% cache token rate. We use ~25% fewer tokens than Claude Code and Codex. DM us to try it free for a month. https://t.co/QeOT0Ju1jQ

Agent Manager is finally out! Orchestrate dozens of agents across harnesses (Claude Code, Codex, OpenCode etc.) effortlessly. Agents communicate and self-organize to help you achieve complex tasks and always keep running on HF spaces. Try it: https://t.co/JSPum13a0g https://t.co/grfIsuCQPK

🔦 Beam, Reflection's first model with open weights, is live today on CoreWeave Serverless Inference. Call it through an API and leave the GPUs to us. We are also proud to share that @reflection_ai used @wandb in CoreWeave Forge during Beam's training. Developers in the private preview have been calling it fast and accurate, and they like that it doesn't ramble. See what @TheCodingSoup on our team built with private preview access below. Get started with Beam here: https://t.co/tyRVazUtUm

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. 🧵:

StreamEval: The first benchmark for interactive streaming video generation It evaluates 37 models across 5 tasks with 300 scripted test cases, revealing that reliable request execution remains a central challenge. https://t.co/UIovU4Tgg0

The evals book is here. Hamel is the evals 🐐 and I’m lucky he wanted to write this book with me!! We are grateful for our reviewers, editors, illustrators and readers. Order it on Amazon; or come to our book signing to get a copy (link in replies).

Hermes Agent can now pick a different model for each sub-agent from routes you define. fast for lookups. deep for debugging. local for private work. one plugin. Advanced Delegation. official plugin from @NousResearch two modes: ▸ model routes define named routes. the agent picks per task. delegation: backend: model_map backends: model_map: options: - name: fast description: "cheap and quick: lookups, summaries." provider: openrouter model: google/gemini-3-flash-preview - name: deep description: "slow and careful: hard debugging." provider: anthropic model: claude-opus-5 - name: local description: "private: runs on this machine." base_url: http://localhost:8080/v1 model: qwen3-coder default: fast the agent sees only the names and descriptions. never the raw provider, model ID, or API key. it can't invent a route. only pick from your list. ▸ profile routes sub-agents run as your Hermes profiles. each profile carries its own model, credentials, skills, and SOUL.md. delegation: backend: profile_routes backends: profile_routes: profiles: - name: coder description: writes and refactors code. - name: researcher description: web research and summaries. default: coder delegate a research task. it runs as your researcher profile with researcher's model, researcher's skills, researcher's memory. fully isolated. switch modes in any chat: /delegation models --now /delegation profiles --now /delegation default or configure from CLI: hermes delegation config hermes delegation config --show install: hermes plugins install advanced-delegation requires Hermes 0.22.5 or newer.

We are really excited to build independent evaluations of RSI progress and are actively hiring. We constantly see conflicting evidence about RSI progress. Most hinges on information coming from labs. To illustrate this, let's just look at evidence discussed *this week* that conflicts in whether it suggests RSI progress is fast or slow. https://t.co/F4vi5Z3fch Results from OpenAI and Google evals indicate the time horizons for R&D tasks being automated are around ~15 minutes, which are much slower than METR time horizons for software tasks. This suggests we should have *slower* RSI timelines. https://t.co/RfFobR22bw Results from Mirendil indicate that pretraining at a certain scale can be heavily automated such that it can be managed by a single pretraining researcher. This suggests we should have *faster* RSI timelines. https://t.co/lIYID0eFII Results from Nvidia indicate that frontier models provide significant uplift for CUDA optimization. This suggests we should have *faster* RSI timelines. Each piece of evidence is more complex than what I write here. Interpreting them all together is confusing even if you trusted them all fully. And factoring in the underlying level of credibility is basically impossible: each result is basically just some plots and numbers selectively disclosed by labs in a tweet/talk/report. So deep sense-making about RSI progress is very hard. If you are interested in building independent RSI evals, apply here: https://t.co/w6CfTsvJfI

We are really excited to build independent evaluations of RSI progress and are actively hiring. We constantly see conflicting evidence about RSI progress. Most hinges on information coming from labs. To illustrate this, let's just look at evidence discussed *this week* that conflicts in whether it suggests RSI progress is fast or slow. https://t.co/F4vi5Z3fch Results from OpenAI and Google evals indicate the time horizons for R&D tasks being automated are around ~15 minutes, which are much slower than METR time horizons for software tasks. This suggests we should have *slower* RSI timelines. https://t.co/RfFobR22bw Results from Mirendil indicate that pretraining at a certain scale can be heavily automated such that it can be managed by a single pretraining researcher. This suggests we should have *faster* RSI timelines. https://t.co/lIYID0eFII Results from Nvidia indicate that frontier models provide significant uplift for CUDA optimization. This suggests we should have *faster* RSI timelines. Each piece of evidence is more complex than what I write here. Interpreting them all together is confusing even if you trusted them all fully. And factoring in the underlying level of credibility is basically impossible: each result is basically just some plots and numbers selectively disclosed by labs in a tweet/talk/report. So deep sense-making about RSI progress is very hard. If you are interested in building independent RSI evals, apply here: https://t.co/w6CfTsvJfI