正在加载视频...
视频加载失败
Meet Apodex 1.1: Scaling Agentic Intelligence for Complex Work Open Source Harness: Open Weights: We’re excited to introduce Apodex 1.1, our new model family built to scale agentic intelligence for professional work. 🧠 Frontier-level intelligence for complex work Apodex 1.1 brings frontier-level agentic performance across complex professional work, scientific... show more
38 条评论
Where Apodex 1.1 lands. ▸ Professional work — APEX-Agent 38.5 · GDPVal 78.8 win rate ▸ Finance — FrontierFinance 54.3 ▸ Scientific research — FrontierScience-Research 63.3 · BioMysteryBench Human-difficult 35.3 ▸ Reasoning — HLE 56.1 📊 Across the four benchmarks evaluated in both ReAct and Agent Team configurations, Agent Team adds 4.1–9.3 points. More on our blog:
Two scaling surfaces, one shared execution foundation. 1. Environment Scaling — expanding the file, search, code, and other executable environments the model can learn and act in. 2. Agentic Coordination Scaling — expanding a task’s ability to be broken down, coordinated, integrated, and reorganized across multiple Agents, branches, and points in time. 3. Persistent task state — keeping the plan, artifacts, dependencies, failures, and next steps intact. For researchers, that means: 🧱 End-to-end execution — from raw input to an inspectable deliverable. 🗂️ Long-horizon task management — plans, state, dependencies, and progress remain intact throughout the run. ✋ Mid-task intervention — without discarding work that remains valid. 👁️ A visible, controllable process — with clear steps, artifacts, failures, and next actions. 🔀 Asynchronous Agent Team coordination — useful work branches out and comes back together. 🔍 Independent Statement Review — key claims go through an independent check before they reach the user.
Apodex 1.1 mini — 35B, open weights, and built for local deployment. With Agent Team: 50.2 FrontierFinance · 51.7 FrontierScience-Research · 27.7 APEX-Agents—virtually matching the 1T-parameter Kimi K2.6 at 27.9. With ReAct: 40.0 · 45.0 · 24.2. FrontierAgent is open-sourced alongside it, with a native command-line TUI, ReAct and Agent Team modes, and one-command launch on macOS and Linux—no preinstallation or Docker requirement.
The persistent state and mid-task intervention are what really I really love here. Scaling intelligence is not just about a bigger model, it’s about building agents that can actually manage complex work from start to finish.
open-sourcing the workbench is the smart move here,models get compared on benchmarks for a week then forgotten, but tooling people can actually build on sticks around,that’s the bet apodex is making.
I just finished tinkering with DeepSeek Harness, and now I’ve come across another open-source Harness with 35B open weights and local deployment—my tinkering spirit has been reawakened.
Well, tinkering spirits are our favorite kind. Join our discord!😊
Since you can step in and redirect the agents mid-task, do you think the new bottleneck will shift from model capabilities to how well we can actually manage and manage our own AI teams?
从 TRACES 的发现型范式，到现在能本地跑的 1.1 + Agent Team，落地速度挺快。准备测试一下长链条任务。
It performs rather poorly in my testing (Svelte 5, doesn‘t use/recognize svelte:window, uses $effect where onMount would be suitable). Could be a lack of knowledge and not intelligence, idk. But I‘d be careful with these bechnmarks. Could be benchmaxxed.
Impressive release from Apodex 1.1 🚀 Asynchronous multi‑agent teamwork unlocks complex professional tasks. It’s notable that its 35B open‑weight mini rivals trillion‑parameter Kimi K2.6 on one agent benchmark. Still curious how these multi‑agent systems hold up in real‑world use cases. 🤔
The Agent Team component coordinates complex, long-horizon tasks through a multi-agent workflow. It decomposes tasks, preserves state, operates asynchronously, and incorporates direction during execution. These functions may provide greater flexibility for tasks that require parallel work or ongoing coordination than workflows based on a single agent
Complex, evolving tasks are already challenging enough, and having agents collaborate asynchronously while responding to new developments indeed requires a more flexible and reliable workflow. Apodex 1.1 looks especially compelling for this kind of project-based work.
Asynchronous Agent Teams deserve more attention than simply stacking up context: they can break down tasks in parallel, and also allow humans to intervene at any time. Looking forward to more stability data from real long tasks.
Open source 💪
A strong launch. 🔥 The asynchronous Agent Team + open-source FrontierAgent combo makes Apodex 1.1 especially interesting. Excited to see what people build with it.
Apodex 1.1 出来了 搞了个开源的 Agent 框架 + 开放权重模型，主打异步多 Agent 一起干活，能搞复杂专业任务（科研、金融、深度搜索这些）。 还有个 1.1 mini（35B），本地就能跑，部分分数都快追上万亿参数的模型了。
The combo of open weights and the async Agent Team harness is huge for local workflows! Excited to test FrontierAgent on some long-horizon research tasks. Great work on this release! 🚀
This approach is pretty interesting—it lets you step in and guide things midway, unlike other AIs that just go haywire once you let them loose. Planning to carve out some time to test how it performs. Is the locally run Mini version any good?
🎉Awesome! Huge props for open-sourcing FrontierAgent + the 35B weights with Apodex 1.1. This is a massive gift to the open-source community — async Agent Team + fully local deployable research workbench. Real frontier agentic capabilities for everyone. 👏
35B的mini版就能在FrontierFinance拿到50.2、几乎追平1T参数的Kimi K2.6，这个性价比太香了。本地部署加上开源FrontierAgent，终于可以不用把敏感金融数据丢到云端了。
So impressive! From the TRACES discovery paradigm to the direct implementation of 1.1 + asynchronous Agent Team, the speed is astonishing. Especially looking forward to the locally runnable 1.1 mini + FrontierAgent, ready to stress-test it with a few real long-chain research/analysis tasks. This design of asynchronous multi-Agent coordination + mid-process intervention feels like it's truly heading toward the "co-discoverer" direction. Starred it already, gonna try it right away
I love being able to jump in anytime during a task to change direction when complex research goes off track, no need to scrap the whole thing and start over.
The interesting claim isn’t another benchmark bump. It’s whether mid-task intervention actually keeps valid work instead of quietly restarting under a new board. That’s the part worth testing.
This open-source pace is way too fast, huh? The 35B mini can run locally, the agent team is this badass, and I feel like my computer's about to start pulling overtime lol. But hey, on the flip side, being able to deploy and research tools on your own is genuinely awesome—gonna give it a shot later to see if it can help me crank out reports in the dead of night.
The mini version with 35B parameters can score 50.2 on FrontierFinance, nearly matching the 1T-parameter Kimi K2.6—this cost-performance ratio is just too enticing. With local deployment plus the open-source FrontierAgent, I can finally stop tossing sensitive financial data onto the cloud.
现在很多 Agent 只是玩具，这个直接瞄准科研、金融真实复杂工作。已经标记链接，等大佬跑 demo 看效果。
Congrats on the launch! Very interesting to see models optimized for multi-agent coordinations!
That a 35B open weight mini can already rival far larger models on agent benchmarks when given the right orchestration layer is the quiet signal: the bottleneck is shifting from “how big is the model” to “how well can we decompose, coordinate, and persist state across agents.
Open source + local deployment + native TUI, very friendly for those who want to run complex tasks themselves! Looking forward to checking out the stability on complex research/file-based tasks after actually trying it out.
Pressure is being continuously applied and I love it
Leverages your Attention Is All You Need
如果多个 Agent 真能做到并行研究再互相整合，复杂任务的效率应该会提升不少。
Open-sourcing the workbench plus a local mini model is a strong move. A lot of these launches stay closed and demo-only.
the asynchronous agent team approach makes this seriously interesting for complex research.
直接放权重和工作台代码，比那些靠 PPT 圈内测资格的团队痛快多了。 就冲开源这一点先 star 一个
这种既要云端高性能上限，又要本地化极度安全，还把调度框架直接开源的套路，本质上是在用工具链捆绑深度用户。 但是对当金融分析师习惯了这种多 Agent 协作的工作流程，基础大模型的替代成本就会变得极高。 智能化竞争的下半场早就不是单体模型跑分的比拼， 而是看谁能把协同架构和实际工作流缝得更无缝。 未来的路还很长
Agent Team adding a few points over ReAct on the same model is the number I care about. Workflow, not a new brain.
