We used dual-isotope mixing models (δ¹³C/δ¹⁵N and δ²H/δ¹⁵N) in a Bayesian framework to partition allochthonous and autochthonous energy sources for salmonids in 2 headwater streams in the Kenai Lowlands, Alaska (USA). Our 1^st objective was to estimate the production base for juvenile coho salmon (Oncorhynchus kisutch) and Dolly Varden (Salvelinus malma). We hypothesized that consumers would be reliant on both autochthonous (filamentous algae and periphyton) and allochthonous sources, but that autochthonous sources would dominate because of the open canopy and lower-quality litter inputs provided by the riparian wetland vegetation, primarily bluejoint grass (Calamagrostis canadensis). Our 2^nd objective was to evaluate the utility of stable H isotopes for tracing energy pathways in a northern-latitude ecosystem. We hypothesized that δ²H-based models would provide more precise estimates of source partitioning than δ¹³C-based models because of greater source separation. Allochthonous source contributions consistently exceeded autochthonous sources for all fish species and size classes at both study sites. However, diet shifted during ontogeny, and larger Dolly Varden relied more on autochthonous sources than did smaller individuals of both species. Last, we found good correspondence and similar levels of precision between the δ¹³C- and δ²H-based models despite greater source separation by δ²H. Our results highlight the importance of allochthonous sources in headwater streams, and we suggest that litter inputs from grasses may be an under-appreciated subsidy to salmon production. Stable H isotopes can be an effective foodweb tracer in northern-latitude streams, but source partitioning results were not sufficiently different from stable C isotope models for us to recommend unequivocally using them to replace or enhance δ¹³C in similar studies.