Apodex 发布 Apodex 1.1 Deep Discover 模型

Apodex has released Apodex 1.1, a model family built to complete work instead of describing it. The previous generation read web pages and returned a written answer. This one opens the files themselves- a spreadsheet, a PDF, a dataset- picks a method, writes and runs the analysis code, and hands back the tables, charts, and scripts it produced, with figures traceable to source. It went live on August 24.

The problem it targets is the cost of a long task that collapses partway through. An agent that meets a missing value or a failed assumption at step forty either stops or returns something vague, and the hours spent getting there get spent again. Apodex calls its answer Adaptive Recovery, where the system explains the adjustment it is making and continues. Two training directions sit behind it. Environment Scaling widens the file, search, and code environments the model learns inside, and Agentic Coordination Scaling teaches it to break a task apart. Both execute on a runtime called AgentOS, which holds tool calls, file state, and progress for a whole job.

The headline capability is Deep Discover, a mode where the model assembles what Apodex calls an Agent Team without being told how, deciding whether a task splits, how many subagents run, and when they consolidate. The team is asynchronous, so each branch reports to a shared task state as it finishes rather than at the end, and the main task absorbs findings in real time while a live task board carries plans, steps, and exceptions. Adding a file or changing a requirement mid-run keeps the intermediate results that still hold and replans only the affected parts. A step called Statement Review then checks the conclusions against their sources, data, and computations, and flags what conflicts.

On benchmarks Apodex published and ran through its own harness, the Agent Team configuration scores 38.5 on APEX-Agents (1.0 scored 16.5) and 63.3 on FrontierScience-Research (28.3). In one inquiry run on the web product, a single Deep Discover task ran 395 steps across 28 sources and returned a 31 KB report, with the reviewer checking 51 claims before delivery. The flagship is a 397B model with a 260K context window, and a 35B open-weight sibling, Apodex 1.1 mini, carries the same coordination layer. Both are free to call on the Apodex API platform for now, at list prices of $0.30 and $0.10 per million input tokens. Deep Discover sits behind a preview program called Apodex Frontier, while the Deep Research and Deep Solve tiers are open to everyone.

Apodex shipped 1.0 in June alongside AgentHarness, an evaluation harness limited to single-agent loops. FrontierAgent, released with 1.1 under Apache 2.0, replaces it with a runtime and terminal interface that covers both modes, starting with one command on macOS and Linux. Apodex also maintains TRACES, a benchmark for discovery-oriented AI, and says pretraining for Apodex 2.0 is underway, built for what it calls a Heavy-Duty Solver.

Star FrontierAgent on GitHub, check the open weights on Hugging Face, or test the API platform.