From a72accde1a4a47db6efee11d4c16cbfa9228a729 Mon Sep 17 00:00:00 2001
From: ltadeu6 <ltadeu6@protonmail.com>
Date: Sun, 12 Apr 2026 11:49:34 -0300
Subject: [PATCH] Use unstable ollama-cuda and pin gemma4:e4b

---
 flake.lock                    | 19 ++++++++++++++++++-
 flake.nix                     | 15 +++++++++++++--
 home/ltadeu6.nix              |  2 +-
 hosts/Nixos/configuration.nix | 16 +++++++++++++++-
 4 files changed, 47 insertions(+), 5 deletions(-)

diff --git a/flake.lock b/flake.lock
index 53639fa..2b304bb 100644
--- a/flake.lock
+++ b/flake.lock
@@ -179,6 +179,22 @@
         "type": "github"
       }
     },
+    "nixpkgs-unstable": {
+      "locked": {
+        "lastModified": 1775710090,
+        "narHash": "sha256-ar3rofg+awPB8QXDaFJhJ2jJhu+KqN/PRCXeyuXR76E=",
+        "owner": "NixOS",
+        "repo": "nixpkgs",
+        "rev": "4c1018dae018162ec878d42fec712642d214fdfa",
+        "type": "github"
+      },
+      "original": {
+        "owner": "NixOS",
+        "ref": "nixos-unstable",
+        "repo": "nixpkgs",
+        "type": "github"
+      }
+    },
     "nixpkgs_2": {
       "locked": {
         "lastModified": 1767364772,
@@ -232,7 +248,8 @@
         "agenix": "agenix",
         "home-manager": "home-manager_2",
         "nix-openclaw": "nix-openclaw",
-        "nixpkgs": "nixpkgs_4"
+        "nixpkgs": "nixpkgs_4",
+        "nixpkgs-unstable": "nixpkgs-unstable"
       }
     },
     "systems": {
diff --git a/flake.nix b/flake.nix
index f4864b9..7f28643 100644
--- a/flake.nix
+++ b/flake.nix
@@ -3,6 +3,7 @@
 
   inputs = {
     nixpkgs.url = "github:NixOS/nixpkgs/nixos-25.11";
+    nixpkgs-unstable.url = "github:NixOS/nixpkgs/nixos-unstable";
     home-manager = {
       url = "github:nix-community/home-manager/release-25.11";
       inputs.nixpkgs.follows = "nixpkgs";
@@ -11,7 +12,7 @@
     nix-openclaw.url = "github:openclaw/nix-openclaw";
   };
 
-  outputs = { self, nixpkgs, home-manager, agenix, nix-openclaw, ... }:
+  outputs = { self, nixpkgs, nixpkgs-unstable, home-manager, agenix, nix-openclaw, ... }:
     let
       system = "x86_64-linux";
     in {
@@ -21,7 +22,17 @@
         modules = [
           ./hosts/Nixos/configuration.nix
           agenix.nixosModules.default
-          { nixpkgs.overlays = [ nix-openclaw.overlays.default ]; }
+          {
+            nixpkgs.overlays = [
+              nix-openclaw.overlays.default
+              (final: prev: {
+                unstable = import nixpkgs-unstable {
+                  inherit system;
+                  config = prev.config;
+                };
+              })
+            ];
+          }
           home-manager.nixosModules.home-manager
           {
             home-manager.useGlobalPkgs = true;
diff --git a/home/ltadeu6.nix b/home/ltadeu6.nix
index e5439e8..5d85277 100644
--- a/home/ltadeu6.nix
+++ b/home/ltadeu6.nix
@@ -349,7 +349,7 @@
           id = "OPENCLAW_GATEWAY_TOKEN";
         };
 
-        agents.defaults.model.primary = "ollama/qwen2.5:3b";
+        agents.defaults.model.primary = "ollama/gemma4:e4b";
         models.providers.ollama = {
           baseUrl = "http://127.0.0.1:11434";
           api = "ollama";
diff --git a/hosts/Nixos/configuration.nix b/hosts/Nixos/configuration.nix
index c2ac204..6c19d2c 100644
--- a/hosts/Nixos/configuration.nix
+++ b/hosts/Nixos/configuration.nix
@@ -43,6 +43,14 @@
       };
   };
 
+  nix.settings = {
+    # CUDA builds are heavy; use a binary cache to avoid local compilation.
+    substituters = lib.mkAfter [ "https://cache.nixos-cuda.org" ];
+    trusted-public-keys = lib.mkAfter [
+      "cache.nixos-cuda.org:74DUi4Ye579gUqzH4ziL9IyiJBlDpMRn9MBN8oNan9M="
+    ];
+  };
+
   powerManagement.cpuFreqGovernor = "performance";
   # Bootloader.
   boot = {
@@ -248,10 +256,16 @@
     };
     ollama = {
       enable = true;
+      # Nixpkgs' plain `ollama` is typically CPU-only; use the CUDA variant on NVIDIA.
+      package = pkgs.unstable.ollama-cuda;
       acceleration = "cuda";
       host = "[::]";
+      environmentVariables = {
+        # Lower default context to keep KV cache smaller (faster + more likely to stay on GPU).
+        OLLAMA_CONTEXT_LENGTH = "8192";
+      };
       loadModels = [
-        "qwen2.5:3b"
+        "gemma4:e4b"
       ];
       syncModels = true;
       # listenAddress = "10.0.0.2:11434";