IBM tests open-model agent factory for quantum chip design

IBM Research says its all-open agent factory improved the high-level-synthesis performance of a quantum error-correction hardware design by about 52 times over the baseline, at an estimated model-inference cost of $25.95.

Given the same budget of 100 optimization hypotheses, IBM’s all-proprietary configuration delivered the larger improvement—about 204 times—but carried an estimated model-inference cost of $299.45, roughly 11.5 times the all-open total. Each configuration was tested in a single search run, so the results describe those observed runs rather than establish a universal model ranking.

The experiment optimized a classical implementation of Relay-BP, a decoder used in quantum error correction. Four worker agents proposed changes across five rounds of 20 hypotheses, with a coordinator directing the search. Candidate designs were checked for correctness and synthesized with Vitis HLS for the same FPGA target. High-level synthesis estimates how code will map to hardware; the reported gains were not measurements from a fabricated production chip.

The IBM’s technical study compared three configurations. Claude Opus 5 served as both coordinator and workers in the proprietary setup, GLM-5.3 filled both roles in the all-open setup, and the hybrid paired an Opus coordinator with GLM workers. IBM said it served GLM-5.3 on premises through llm-d.

At the 100-hypothesis mark, the best HLS-estimated latency was 139 microseconds for the all-proprietary configuration, 223 microseconds for the hybrid and 544.7 microseconds for the all-open configuration, versus a 28,231-microsecond baseline. Those results corresponded to speedups of about 204 times, 126 times and 52 times, respectively. Estimated inference costs were $299.45, $105.25 and $25.95 in the same order.

Lower spending did not mean fewer tokens. IBM reported about 240.9 million tokens for the all-open search after excluding one runaway worker execution, compared with 215.9 million for the all-proprietary run and 197.5 million for the hybrid. The study attributed the difference to serving economics, estimating a blended on-premises cost of $0.1077 per million tokens for GLM-5.3. For Opus, it used standard API rates for input, output and cache-read tokens.

IBM said llm-d routed requests according to reusable cached context and separated prompt processing from token generation. The study reported a prefix-cache hit rate of about 98% for the workload.

The dollar totals cover model inference, not the full cost of hardware development. They do not include non-inference synthesis or verification costs. The opened study also did not provide enough raw billing detail to independently recompute every component, and the excluded runaway GLM worker was not quantified in the source material.

Extending the all-open search from 100 to 200 hypotheses improved its best estimated latency from 544.7 to 467.6 microseconds, raising the reported speedup from 51.8 times to 60.4 times. IBM reported diminishing incremental gains and said future matched-budget evaluations will also need to account for synthesis and verification costs alongside inference.