Mercor published a detailed account on September 1 of how it post-trained Qwen3.5-397B-A17B, an open-weight mixture-of-experts model, using reinforcement learning on 1,928 expert-created knowledge-work tasks. The company says Pass@1 on its own APEX-Agents benchmark rose 70 percent relative, from 16.11 percent to 27.29 percent. That is a large jump on a hard test, and it is the third post in a series where Mercor has scaled this recipe from under 1,000 tasks to roughly 2,000.
The infrastructure matters here. Mercor built the run on SkyRL, an asynchronous reinforcement-learning framework from Berkeley’s Sky Computing Lab developed with Anyscale, paired with Harbor, an environment and rollout manager from the Terminal-Bench team. Each task runs inside a simulated company sandbox stocked with PDFs, spreadsheets, and chat logs, and the agent works through MCP tool calls or code execution. Getting that sandbox to fail cleanly, rather than silently, consumed most of the reported engineering effort.
Mercor’s own numbers make the case for where the gains actually came from. Fixing harness bugs alone, before any reinforcement learning began, raised the base Qwen3.6-35B-A3B model’s mean reward from 22.74 percent to 28.69 percent with zero training. In a separate ablation study, the single best algorithm choice (a loss-aggregation method called prompt_mean) added 3.9 points. Full post-training moved both models 10 to 12 points overall. By Mercor’s own accounting, environment quality and harness correctness did more work than the reinforcement-learning algorithm chosen to run on top of them.
That comparison deserves a caveat the blog post does not supply. Mercor sells expert-labeled training data and task environments as a commercial product, so a finding that expert-built environments matter more than algorithm choice is also a finding that favors Mercor’s own business. APEX-Agents is Mercor’s proprietary benchmark family, built from its own task set, so the 70 percent figure is a result on a test the company designed and sells access to, not an independent, third-party evaluation. None of that makes the underlying numbers wrong, but readers should treat the 70 percent figure as a vendor claim rather than a peer-reviewed result until an outside lab reproduces it on a benchmark Mercor does not control.
The transfer results are the part least tied to Mercor’s own scoring. The post-trained models were also evaluated on Terminal-Bench 2.1 under a different harness, OpenCode, with no MCP servers exposed at all, and the gains largely carried over. Scores on HLE and GPQA, two reasoning benchmarks unrelated to agentic tool use, showed no regression. That combination, a capability gain that survives a harness swap without breaking general reasoning, is the more durable evidence in the post.
The company states its own takeaway without hedging: what they fed the model mattered more than which algorithm they picked. If that holds at frontier scale generally rather than only inside Mercor’s pipeline, the scarce input in frontier post-training is not GPU-hours or a clever loss function. It is expert-authored task environments built by people who actually do the underlying knowledge work, which reprices the entire layer of vendors selling labeled data and simulated environments to labs training agentic models. Mercor released the training script, model weights, and eval traces publicly, which gives outside teams a first chance to test that claim on data Mercor did not curate.
Teams evaluating open-weight agent models for knowledge-work deployment should treat the 70 percent APEX-Agents figure as a vendor benchmark and prioritize the Terminal-Bench transfer result when deciding how much weight the claim deserves.
Reporting based on Mercor’s engineering blog post, “Training frontier knowledge work agents: A 397B RL training guide with SkyRL,” published September 1, 2026.
