Alibaba's Accio ships Occamy-1.0, an open 35B agent model
Three papers this morning, one argument between them: the interesting frontier in agents is no longer peak capability but the cost, plumbing, and audit trail of getting work actually done.
Occamy-1.0, released by Alibaba International's Accio team under Apache-2.0, is a 35B-total, 3B-active Mixture-of-Experts model specialized for "co-work" — long, stateful tasks that chain search, code, tools, and files. Built by further training the Qwen3.6-35B-A3B checkpoint (256 experts, 8 routed plus 1 shared active per token), it scores 82.20 average on Claw-Eval against GPT-5.6 Sol's 81.80, and lifts Automation pass@1 from its base model's 7.5 to 27.6; the paper is honest that on GDPval the frontier still leads, 1,128 against 1,741, and positions the model at the low-cost knee of the cost-performance Pareto frontier rather than above it. The more consequential release is the scaffolding: weights, part of a 14,998-sample, 403 million-token execution-grounded fine-tuning corpus, and the multi-harness reinforcement-learning infrastructure behind it are all public. Our read: a product team is saying the binding constraint for agents is now per-episode cost, not peak score — and at 3B active parameters, an overnight workflow stops being a budget line item. This is the same Accio group that launched CommerceAgentBench last week, and it sits in the current wave of small open models squeezing the frontier on price — A 33B model from Singapore just matched DeepSeek V4 Pro — on free tokens.
An independent study priced the harness question, and the answer is "it depends on the task." On a private, contamination-controlled suite of 256 repository and post-cutoff contest problems, the same model ran under its vendor's native agent SDK and under a neutral one — claude-opus-4.8 under claude-agent-sdk versus deepagents, gpt-5.5 under the OpenAI Codex SDK versus deepagents — and the 792 graded paired runs resolve to no consistent average edge either way: −1.25 and +1.25 points, with confidence intervals straddling zero. The Opus average, though, hides two opposite halves: on 61 repository tasks the native harness trails by 9.0 points, on 19 contest tasks it leads by 23.7 (p = 0.003 on a label-permutation test — and the author concedes the split was chosen after seeing the data and needs designed replication). The paper also corrects an August manuscript whose cost claims rested on a defect in its own telemetry; in the re-pricing the neutral harness costs 1.2 to 1.6 times as much per solved task, but 58 runs on the Anthropic account left no usage record at all, so even the billing order is unresolved.
A systematization of jailbreak research argues the field has been grading the wrong artifact. The paper re-maps attacks and defenses across the whole agentic pipeline — user interaction, planning and reasoning, memory, tool use, inter-agent communication — and lands three findings from a controlled study: strong native alignment does not imply adversarial robustness; defense effectiveness swings model by model, attack by attack, component by component, often at real cost in over-refusal, utility, and latency; and a low final-response attack-success rate can mask severe intermediate compromise — planner, memory, and tool interactions remain unsafe even when the final answer gets filtered out. The practical read for anyone shipping agents: if your guardrail inspects only the outbound message, your audit trail is decorative.
What to watch: whether Occamy's released RL harness gets forked into other vertical agents, and whether the harness-versus-model contrast draws the designed replication its author calls for.
If a 3B-active open model handles your overnight workflows and a frontier model only wins on the hard one-offs, which one is your stack actually shipping this quarter? Tell us in the comments.
