diff --git a/local/ai-provider/lua/ai-provider/curl.lua b/local/ai-provider/lua/ai-provider/curl.lua
index 7d6f151..1574aa8 100644
--- a/local/ai-provider/lua/ai-provider/curl.lua
+++ b/local/ai-provider/lua/ai-provider/curl.lua
@@ -138,6 +138,39 @@ function M.stream_json_lines(request)
   local stderr = {}
   local stdout_lines = {}
   local http_status = nil
+  local stdout_buffer = ''
+
+  local function process_stdout_line(line)
+    line = line:gsub('^%s+', ''):gsub('%s+$', '')
+    if line == '' then
+      return
+    end
+
+    local status = line:match('^' .. status_marker .. '(%d+)$')
+    if status then
+      http_status = tonumber(status)
+      return
+    end
+
+    local payload = line
+    local event_data = line:match '^data:%s*(.*)$'
+    if event_data then
+      if event_data == '[DONE]' then
+        return
+      end
+      payload = event_data
+    end
+
+    if #stdout_lines < 20 then
+      table.insert(stdout_lines, payload)
+    end
+
+    local ok, data = pcall(vim.json.decode, payload)
+    if ok and type(data) == 'table' and request.on_json_line then
+      schedule(request.on_json_line, data, line)
+    end
+  end
+
   local job = Job:new {
     command = 'curl',
     args = args,
@@ -146,33 +179,21 @@ function M.stream_json_lines(request)
         return
       end
 
-      for line in tostring(output):gmatch('[^\r\n]+') do
-        line = line:gsub('^%s+', ''):gsub('%s+$', '')
-        if line ~= '' then
-          local status = line:match('^' .. status_marker .. '(%d+)$')
-          if status then
-            http_status = tonumber(status)
-          else
-            local payload = line
-            local event_data = line:match '^data:%s*(.*)$'
-            if event_data then
-              if event_data == '[DONE]' then
-                goto continue
-              end
-              payload = event_data
-            end
-
-            if #stdout_lines < 20 then
-              table.insert(stdout_lines, payload)
-            end
-
-            local ok, data = pcall(vim.json.decode, payload)
-            if ok and type(data) == 'table' and request.on_json_line then
-              schedule(request.on_json_line, data, line)
-            end
-          end
+      local text = tostring(output)
+      if stdout_buffer == '' and not text:find('[\r\n]') then
+        process_stdout_line(text)
+        return
+      end
+
+      stdout_buffer = stdout_buffer .. text
+      while true do
+        local newline_at = stdout_buffer:find('[\r\n]')
+        if not newline_at then
+          break
         end
-        ::continue::
+
+        process_stdout_line(stdout_buffer:sub(1, newline_at - 1))
+        stdout_buffer = stdout_buffer:sub(newline_at + 1)
       end
     end,
     on_stderr = function(_, line)
@@ -181,6 +202,11 @@ function M.stream_json_lines(request)
       end
     end,
     on_exit = function(_, code)
+      if stdout_buffer ~= '' and not (request.is_cancelled and request.is_cancelled()) then
+        process_stdout_line(stdout_buffer)
+        stdout_buffer = ''
+      end
+
       local error_parts = {}
       if #stderr > 0 then
         table.insert(error_parts, table.concat(stderr, '\n'))
diff --git a/local/ai-provider/lua/ai-provider/providers/ollama.lua b/local/ai-provider/lua/ai-provider/providers/ollama.lua
index 88d623d..5487bcb 100644
--- a/local/ai-provider/lua/ai-provider/providers/ollama.lua
+++ b/local/ai-provider/lua/ai-provider/providers/ollama.lua
@@ -6,7 +6,6 @@ local DEFAULT_ENDPOINT = 'http://127.0.0.1:11434'
 local HEALTH_CACHE_TTL = 60
 local MODEL_CACHE_TTL = 60
 local STATUS_THROTTLE_MS = 100
-local MAX_PENDING_CHUNK_CHARS = 4000
 
 local state = {
   health = nil,
@@ -190,7 +189,6 @@ function M.chat(request)
   local pending_statuses = {}
   local pending_status_flush_scheduled = false
   local pending_chunks = {}
-  local pending_chunk_chars = 0
   local pending_chunk_flush_scheduled = false
 
   local function flush_pending_statuses()
@@ -211,13 +209,11 @@ function M.chat(request)
     pending_chunk_flush_scheduled = false
     if not request.on_chunk or is_cancelled(request) then
       pending_chunks = {}
-      pending_chunk_chars = 0
       return
     end
 
     local chunks_to_flush = pending_chunks
     pending_chunks = {}
-    pending_chunk_chars = 0
     for _, item in ipairs(chunks_to_flush) do
       request.on_chunk(item.chunk, item.data, item.kind)
     end
@@ -244,13 +240,12 @@ function M.chat(request)
       return
     end
 
-    local item = { chunk = chunk, data = data, kind = kind }
-    table.insert(pending_chunks, item)
-    pending_chunk_chars = pending_chunk_chars + #chunk
-
-    while pending_chunk_chars > MAX_PENDING_CHUNK_CHARS and #pending_chunks > 1 do
-      local removed = table.remove(pending_chunks, 1)
-      pending_chunk_chars = pending_chunk_chars - #(removed.chunk or '')
+    local previous = pending_chunks[#pending_chunks]
+    if previous and previous.kind == kind then
+      previous.chunk = previous.chunk .. chunk
+      previous.data = data
+    else
+      table.insert(pending_chunks, { chunk = chunk, data = data, kind = kind })
     end
     schedule_chunk_flush()
   end
diff --git a/lua/plugins/ai-provider.lua b/lua/plugins/ai-provider.lua
index b97484e..d8b94a6 100644
--- a/lua/plugins/ai-provider.lua
+++ b/lua/plugins/ai-provider.lua
@@ -12,7 +12,7 @@ return {
         default_model = 'auto',
       },
       ollama = {
-        default_model = 'gemma4:e4b',
+        default_model = 'gemma4:e2b',
         context_size = 1024 * 8,
         timeout = 180000,
         keep_alive = '1h',