diff --git a/network-poc/tests/model-benchmark.mjs b/network-poc/tests/model-benchmark.mjs
index f2e87e5..feedba7 100644
--- a/network-poc/tests/model-benchmark.mjs
+++ b/network-poc/tests/model-benchmark.mjs
@@ -79,7 +79,9 @@ async function ollamaChat(model, prompt, systemPrompt, maxTokens = 2048) {
     if (!resp.ok) throw new Error(`Ollama HTTP ${resp.status}: ${await resp.text()}`);
     const data = await resp.json();
     const elapsed = Date.now() - start;
-    const text = stripThinking((data.message?.content || '').trim());
+    // Ollama: jotkin mallit (qwen3.5) palauttavat ajattelun erillisessä thinking-kentässä
+    const rawContent = (data.message?.content || '').trim();
+    const text = stripThinking(rawContent);
     const evalCount = data.eval_count || 0;
     const evalDurationNs = data.eval_duration || 1;
     const tokPerSec = evalCount / (evalDurationNs / 1e9);
@@ -283,7 +285,7 @@ async function runPipeline(model, scenario) {
     try {
         // 1. Vaatimukset
         console.log(`    [1/5] Vaatimukset...`);
-        const req = await ollamaChat(model, scenario.prompt, CLIENT_SYSTEM, 1024);
+        const req = await ollamaChat(model, scenario.prompt, CLIENT_SYSTEM, 2048);
         timings.push(req);
         if (!req.text || req.text.length < 50) { result.error = 'Vaatimukset liian lyhyet'; return result; }
         result.reqOk = true;
@@ -291,7 +293,7 @@ async function runPipeline(model, scenario) {
 
         // 2. JSON-speksi
         console.log(`    [2/5] JSON-speksi...`);
-        const specResp = await ollamaChat(model, `${req.text}\n\nOutput a JSON spec for this project.`, SPEC_SYSTEM, 2048);
+        const specResp = await ollamaChat(model, `${req.text}\n\nOutput a JSON spec for this project.`, SPEC_SYSTEM, 4096);
         timings.push(specResp);
         const spec = extractJson(specResp.text);
         if (!spec || !spec.entities || spec.entities.length === 0) { result.error = 'JSON-speksi epäonnistui'; writeFileSync(`${dir}/_spec_raw.txt`, specResp.text); return result; }