diff --git a/README.md b/README.md
index fd6afa8d..aeb1d86b 100644
--- a/README.md
+++ b/README.md
@@ -142,6 +142,30 @@ to run in OpenCL mode add the flag `-gpu_backend opencl`.
 
 There are more flags you can use to configure sampling; [read about them here](doc/flags.md#sampling).
 
+## Step 4 (optional): Serve samples from an HTTP server
+If you want to generate new text on demand without loading the model for every sample, you can use the script `server.lua`. 
+You will need to install [Turbo](https://github.com/kernelsauce/turbo) framework first:
+```bash
+luarocks install turbo
+```
+
+Check [Turbo installation manual](https://github.com/kernelsauce/turbo#installation) in case of problems.
+Then run the server:
+
+```bash
+th server.lua -checkpoint cv/checkpoint_10000.t7 -port 8888
+```
+
+Now you can generate new sample by sending an HTTP GET request:
+```bash
+curl -G -d "length=100&temperature=0.9" http://localhost:8888/sample
+```
+
+The following command line arguments of `sample.lua` remain unchanged for `server.lua`: 
+`-checkpoint`, `-gpu`, `-gpu_backend`, `-verbose`. `-port` argument configures the HTTP port.
+The other arguments (`length`, `start_text`, `temperature` and `sample`) should be passed as GET parameters.
+
+
 # Benchmarks
 To benchmark `torch-rnn` against `char-rnn`, we use each to train LSTM language models for the tiny-shakespeare dataset
 with 1, 2 or 3 layers and with an RNN size of 64, 128, 256, or 512. For each we use a minibatch size of 50, a sequence 
diff --git a/server.lua b/server.lua
new file mode 100644
index 00000000..d0f7978c
--- /dev/null
+++ b/server.lua
@@ -0,0 +1,59 @@
+require 'torch'
+require 'nn'
+
+require 'LanguageModel'
+
+local turbo = require("turbo")
+
+local cmd = torch.CmdLine()
+cmd:option('-checkpoint', 'cv/checkpoint_4000.t7')
+cmd:option('-gpu', 0)
+cmd:option('-gpu_backend', 'cuda')
+cmd:option('-verbose', 0)
+cmd:option('-port', 8888) -- http port to listen
+local opt = cmd:parse(arg)
+
+
+local checkpoint = torch.load(opt.checkpoint)
+local model = checkpoint.model
+
+local msg
+if opt.gpu >= 0 and opt.gpu_backend == 'cuda' then
+  require 'cutorch'
+  require 'cunn'
+  cutorch.setDevice(opt.gpu + 1)
+  model:cuda()
+  msg = string.format('Running with CUDA on GPU %d', opt.gpu)
+elseif opt.gpu >= 0 and opt.gpu_backend == 'opencl' then
+  require 'cltorch'
+  require 'clnn'
+  model:cl()
+  msg = string.format('Running with OpenCL on GPU %d', opt.gpu)
+else
+  msg = 'Running in CPU mode'
+end
+if opt.verbose == 1 then print(msg) end
+
+model:evaluate()
+
+
+local SampleHandler = class("SampleHandler", turbo.web.RequestHandler)
+
+function SampleHandler:get()
+    -- Get the 'name' argument, or use 'Santa Claus' if it does not exist
+    opt['length'] = self:get_argument("length", 2000)
+    opt['start_text'] = self:get_argument("start_text", "")
+    opt['sample'] = self:get_argument("sample", 1)
+    opt['temperature'] = self:get_argument("temperature", 1)
+    
+    local sample = model:sample(opt)
+    
+    self:write(sample)
+end
+
+local app = turbo.web.Application:new({
+    {"/sample", SampleHandler}
+})
+
+app:listen(opt.port)
+turbo.ioloop.instance():start()