Lower local qwen3-coder-next message budget
ober
779b5a3be0b04fdab9758a11d4e546f2a1c899fc
--- a/src/jcode/core/models.ss +++ b/src/jcode/core/models.ss @@ -460,11 +460,11 @@ ;; weights leave only ~4 GB under the 32 GB wired cap, so keep the window ;; conservative — compaction must fire before prefill OOMs the GPU. ((string-contains mid "TurboQuant") 16384) - ;; The local qwen3-coder-next MLX server stalls before completing large - ;; 26k-token prefill turns. Keep its budget below the generic qwen3 - ;; architectural window so auto-compaction starts while the server is - ;; still responsive. - ((string-contains mid "qwen3-coder-next-mlx") 16384) + ;; The local qwen3-coder-next MLX server stalls around 26k-token prefill + ;; turns. jcode's message estimator does not include the fixed tool + ;; schema overhead (roughly another 8-9k tokens), so this is an effective + ;; message budget rather than the model's architectural context window. + ((string-contains mid "qwen3-coder-next-mlx") 8192) ;; grok-build: Grok CLI proxy advertises a 512k context window in ;; ~/.grok/models_cache.json (info.context_window). Hardcode here so ;; the TUI ctx% bar has a value even before models_cache.json is read. --- a/test/run.ss +++ b/test/run.ss @@ -1821,7 +1821,7 @@ (check! "ctx window grok-build = 512000" (model-context-window "grok-build") 512000) (check! "ctx window grok-3 unaffected (no clause)" (model-context-window "grok-3") #f) (check! "ctx window qwen3-coder-next mlx conservative" - (model-context-window "/Users/user/models/qwen3-coder-next-mlx") 16384) + (model-context-window "/Users/user/models/qwen3-coder-next-mlx") 8192) (check! "ctx window generic qwen3 remains 32768" (model-context-window "qwen3:8b-q4_K_M") 32768)