High-resolution regional weather forecasting supports applications such as early warning, hydrological planning, and regional risk assessment, but traditional numerical models such as the Weather Research and Forecasting (WRF) system remain computationally demanding for rapid ensemble or sensitivity experiments. Machine-learning emulators can substan- tially lower this cost, but many are tied to a single grid or depend on future boundary fields that are not always available. We present <strong>FiLMeR (v1.0)</strong>, a context-conditioned machine-learning emulator for WRF-scale near-surface regional weather fields. FiLMeR encodes dynamic atmospheric forcing and static physiographic information in two separate encoders and conditions the decoder with Feature-wise Linear Modulation (FiLM) using grid spacing and cyclic time-of-day/day-of-year metadata. The model predicts six near-surface variables (temperature, humidity, surface pressure, wind components, and precipitation) from historical Global Forecasting System (GFS) states, using a hurdle-style precipitation head. A single shared-weight model is trained and evaluated across four nested WRF domains over the Indian subcontinent, at two grid spacings (27 km and 9 km).</p> <p>We test whether FiLMeR&rsquo;s two central design choices, separating inputs into two encoders and conditioning with FiLM, are each responsible for its accuracy. This is done using a controlled comparison across seven combinations of these two choices, together with a diagnostic test on the trained model. Any form of conditioning substantially improves on an unconditioned model, but the dual-encoder design on its own performs worse than a simpler single encoder, and needs FiLM to become competitive again. Once FiLM is added, this configuration ties with a simpler single-encoder design that uses a more basic conditioning method, under a comparison rule fixed before the results were seen; neither is clearly better.</p> <p>The diagnostic test further shows that FiLM&rsquo;s measured contribution is mainly temporal (time-of-day and day-of-year), not domain-adaptive, even though the model is trained across four geographically distinct domains. On the independent 2025 test period, FiLMeR achieves a 2 m temperature RMSE of 1.50 K and a surface-pressure RMSE of 1.37 hPa, and produces the corresponding 80-step, 3-hourly output sequence (equivalent to a 10-day WRF forecast) in a fraction of a second on a single GPU, about 38,000 times faster than the WRF dynamical integration it emulates. FiLMeR is therefore a computationally efficient emulator for these WRF-scale surface fields: conditioning is necessary for its accuracy, and FiLM&rsquo;s benefit is primarily a temporal-conditioning mechanism.