[Sync] llm stuff

common/.config/aichat/config.yaml

@@ -1,3 +1,4 @@
+# NOTE: the lw/ and wm/ host overlays carry full-file copies of this config — mirror changes there.
 # see https://github.com/sigoden/aichat/blob/main/config.example.yaml
 keybindings: vi
 editor: nvim
@@ -22,34 +23,35 @@ clients:
   name: local
   api_base: http://192.168.0.204:11343/v1
   models: &lan_models
+  # Speeds are benched decode t/s (clean-night sweeps) — see fl/.config/llamacpp/README.md roster.
   - name: Qwen3-Coder-30B-Instruct-UD-Q3_K_XL
-    max_input_tokens: 30000            # ctx 32768
+    max_input_tokens: 30000            # ctx 32768 · ~30 t/s — main agent coder
   - name: Qwen3-Coder-Next-UD-IQ3_XXS
-    max_input_tokens: 128000           # ctx 131072
+    max_input_tokens: 128000           # ctx 131072 · ~16 t/s — long sessions (128k)
   - name: Qwen3.6-35B-A3B-MTP-UD-IQ3_XXS
-    max_input_tokens: 22000            # ctx 24576
+    max_input_tokens: 22000            # ctx 24576 · ~36 t/s — daily driver
     supports_vision: true
   - name: Qwen3.6-35B-A3B-Thinking
-    max_input_tokens: 22000            # ctx 24576
+    max_input_tokens: 22000            # ctx 24576 · ~39 t/s — hard problems
     supports_vision: true
   - name: Qwen3.5-9B-UD-Q6_K_XL
-    max_input_tokens: 30000            # ctx 32768
+    max_input_tokens: 30000            # ctx 32768 · ~32 t/s — quick tasks
     supports_vision: true
   - name: Qwen3.8-27B-UD-IQ3_XXS
-    max_input_tokens: 22000            # ctx 24576
+    max_input_tokens: 22000            # ctx 24576 · ~15 t/s — hybrid reasoner, quiet desktop only
     supports_vision: true
   - name: gemma-4-26B-A4B-it-UD-IQ4_XS
-    max_input_tokens: 22000            # ctx 24576
+    max_input_tokens: 22000            # ctx 24576 · ~34 t/s — quality generalist, best vision
     supports_vision: true
   - name: gemma-4-E4B-it-UD-Q8_K_XL
-    max_input_tokens: 62000            # ctx 65536
+    max_input_tokens: 62000            # ctx 65536 · ~57 t/s — fast generalist, long docs
     supports_vision: true
   - name: GLM-4.7-Flash-UD-Q4_K_XL
-    max_input_tokens: 22000            # ctx 24576
+    max_input_tokens: 22000            # ctx 24576 · ~21 t/s — quality coder
   - name: gpt-oss-20b
-    max_input_tokens: 62000            # ctx 65536
+    max_input_tokens: 62000            # ctx 65536 · ~38 t/s — fast reasoning + tools
   - name: gpt-oss-20b-low
-    max_input_tokens: 62000            # ctx 65536
+    max_input_tokens: 62000            # ctx 65536 · ~37 t/s — snappy answers
 # Remote access via duskadiy.com (reachable from anywhere; requires an API key —
 # $DUSKADIY_API_KEY, from the gitignored conf.d/secrets.fish). The LAN endpoint above is keyless.
 # Same server, same model ids — reuse the list above via a YAML anchor.