Offer Recursive Flint as a model on this Mac

Sources/Lamp/LocalModels.swift

@@ -10,6 +10,8 @@ struct LocalModelSpec: Identifiable, Hashable {
     let sizeGB: Double
     let minRAMGB: Int
     let blurb: String
+    /// Tokens the model is given; most get Lamp's usual size, one made for long work more.
+    var contextSize: Int = LocalModels.contextSize
 }
 
 /// A model on disk that Lamp can run.
@@ -70,6 +72,12 @@ final class LocalModels: ObservableObject {
         LocalModelSpec(id: "qwen3.6-35b-a3b", name: "Qwen3.6 35B-A3B", repo: "unsloth/Qwen3.6-35B-A3B-GGUF",
                        files: ["Qwen3.6-35B-A3B-UD-Q4_K_M.gguf", "mmproj-F16.gguf"], sizeGB: 23.0, minRAMGB: 48,
                        blurb: "Best local option. Fast for its size, and can see screenshots."),
+        LocalModelSpec(id: "flint", name: "Recursive Flint", repo: "RecursiveSI/flint-int4-gguf",
+                       files: ["flint-int4-q4_1.gguf"], sizeGB: 26.0, minRAMGB: 48,
+                       blurb: "Very fast and made for coding, but can't see screenshots.",
+                       // Its own card runs it at 262K; at 128K its cache is about 3.4 GB, since all but 13 of its
+                       // layers look back only 513 tokens. Lamp's prompt alone is over 20K, so 32K ran out in a few turns.
+                       contextSize: 131_072),
         LocalModelSpec(id: "qwen3.8-27b", name: "Qwen3.8 27B", repo: "unsloth/Qwen3.8-27B-GGUF",
                        files: ["Qwen3.8-27B-UD-Q4_K_M.gguf", "mmproj-F16.gguf"], sizeGB: 17.4, minRAMGB: 32,
                        blurb: "Strong and careful, but slower. Can see screenshots."),
@@ -320,6 +328,16 @@ final class LocalModels: ObservableObject {
         writePiModels()
     }
 
+    /// The context a model on disk is run with: its catalog entry's, else Lamp's usual size.
+    func contextSize(of model: InstalledModel) -> Int {
+        Self.catalog.first { spec in model.repo == spec.repo && model.files.contains { $0.lastPathComponent == spec.files[0] } }?
+            .contextSize ?? Self.contextSize
+    }
+
+    /// The router's presets: the context for each model. A size given to llama-server itself would win over
+    /// these, so it's only set here.
+    nonisolated static let presetsFile = Edition.supportFolder.appendingPathComponent("models.ini")
+
     /// llama-server's models folder: one subfolder per GGUF model, symlinking the real files.
     private func rebuildRouter() {
         let fm = FileManager.default
@@ -333,6 +351,11 @@ final class LocalModels: ObservableObject {
                                            withDestinationURL: file.resolvingSymlinksInPath())
             }
         }
+        var presets = "version = 1\n\n[*]\nc = \(Self.contextSize)\n"
+        for model in installed where model.format == .gguf && contextSize(of: model) != Self.contextSize {
+            presets += "\n[\(model.id)]\nc = \(contextSize(of: model))\n"
+        }
+        try? presets.write(to: Self.presetsFile, atomically: true, encoding: .utf8)
     }
 
     /// Deletes a GGUF model's files from the Hugging Face cache (blobs only if nothing else uses them).
@@ -564,8 +587,9 @@ final class LocalModels: ObservableObject {
         rebuildRouter()
         let process = Process()
         process.executableURL = binary
-        process.arguments = ["--models-dir", Self.routerDir.path, "--jinja", "--host", "127.0.0.1", "--api-key", Self.apiKey,
-                             "--port", String(Self.port), "-ngl", "999", "-c", String(Self.contextSize)]
+        process.arguments = ["--models-dir", Self.routerDir.path, "--models-preset", Self.presetsFile.path,
+                             "--jinja", "--host", "127.0.0.1", "--api-key", Self.apiKey,
+                             "--port", String(Self.port), "-ngl", "999"]
         // The router runs each loaded model in a worker of its own, on a loopback port of its own, and leaves the key
         // out of the worker's arguments; the worker inherits the environment, and reads the key from there.
         var env = ProcessInfo.processInfo.environment
@@ -646,7 +670,7 @@ final class LocalModels: ObservableObject {
         if !gguf.isEmpty {
             providers[Builder.localProvider] = [
                 "baseUrl": "\(Self.baseURL)/v1", "api": "openai-completions", "apiKey": Self.apiKey,
-                "models": gguf.map { ["id": $0.id, "contextWindow": Self.contextSize,
+                "models": gguf.map { ["id": $0.id, "contextWindow": contextSize(of: $0),
                                       "input": $0.vision ? ["text", "image"] : ["text"]] as [String: Any] },
             ]
         }

build.sh

@@ -36,10 +36,10 @@ PI_TARBALL="vendor/pi-$PI_VERSION-darwin-$PI_ARCH.tar.gz"
 # llama.cpp's server runs downloaded models on this Mac (MIT, github.com/ggml-org/llama.cpp). Lamp doesn't carry it:
 # it's signed here and zipped as an asset of the release (below), which Lamp fetches the first time a local model is
 # downloaded or picked (LlamaInstall.swift).
-LLAMA_BUILD="b11346"
+LLAMA_BUILD="b11814"
 case "$(uname -m)" in
-  arm64)  LLAMA_ARCH="arm64"; LLAMA_SHA256="25cefd22cf708a14eacea623327a2320fa07493a29b4c37bf3376623de581553" ;;
-  x86_64) LLAMA_ARCH="x64";   LLAMA_SHA256="1f91f85611d1573c53409c99d96d43451cd89e928ce1882b6739fd1cf7befaf2" ;;
+  arm64)  LLAMA_ARCH="arm64"; LLAMA_SHA256="698d16c10104ad5212fdd9cb084e7076b3ab4255c68433ee8ed70f3d041432cb" ;;
+  x86_64) LLAMA_ARCH="x64";   LLAMA_SHA256="cb38196573bcf57c441523f15da3129ef162f202fdacf7c14d7b41c2156cc858" ;;
 esac
 LLAMA_TARBALL="vendor/llama-$LLAMA_BUILD-bin-macos-$LLAMA_ARCH.tar.gz"
 if [ ! -f "$LLAMA_TARBALL" ] || ! echo "$LLAMA_SHA256  $LLAMA_TARBALL" | shasum -a 256 -c - >/dev/null 2>&1; then