From aeb84784f264de6c8695d2afcd4d934d810328c9 Mon Sep 17 00:00:00 2001 From: matevip Date: Mon, 3 Aug 2026 03:15:19 -0400 Subject: [PATCH] feat(llm): add explicit context windows for GLM-5V-Turbo and the Kimi coding alias --- .../java/vip/mate/llm/probe/ModelContextWindowCatalog.java | 6 ++++++ .../vip/mate/llm/probe/ModelContextWindowCatalogTest.java | 4 ++++ 2 files changed, 10 insertions(+) diff --git a/mateclaw-server/src/main/java/vip/mate/llm/probe/ModelContextWindowCatalog.java b/mateclaw-server/src/main/java/vip/mate/llm/probe/ModelContextWindowCatalog.java index d7026264..284b2ae6 100644 --- a/mateclaw-server/src/main/java/vip/mate/llm/probe/ModelContextWindowCatalog.java +++ b/mateclaw-server/src/main/java/vip/mate/llm/probe/ModelContextWindowCatalog.java @@ -84,6 +84,9 @@ final class ModelContextWindowCatalog { // ===== Moonshot Kimi ===== m.put("kimi-k2", 262_144); + // Coding-plan alias (plus its -highspeed variant) for the K2.7 code + // model, which serves the same 256k window. + m.put("kimi-for-coding", 262_144); // ===== Zhipu GLM ===== m.put("glm-4.7", 204_800); @@ -91,6 +94,9 @@ final class ModelContextWindowCatalog { // 200k across the GLM-5 line (5 / 5.1 / turbo variants); 5.2 lifted it to 1M. m.put("glm-5", 204_800); m.put("glm-5.2", 1_000_000); + // Multimodal sibling, stated separately rather than inherited from the + // glm-5 prefix: its 200k window is documented in its own model page. + m.put("glm-5v", 204_800); // The 9B open weights ship at 128k (the separate -1m build is its own id). m.put("glm-4-9b", 131_072); diff --git a/mateclaw-server/src/test/java/vip/mate/llm/probe/ModelContextWindowCatalogTest.java b/mateclaw-server/src/test/java/vip/mate/llm/probe/ModelContextWindowCatalogTest.java index 70ae6f11..be9c93b4 100644 --- a/mateclaw-server/src/test/java/vip/mate/llm/probe/ModelContextWindowCatalogTest.java +++ b/mateclaw-server/src/test/java/vip/mate/llm/probe/ModelContextWindowCatalogTest.java @@ -41,6 +41,10 @@ class ModelContextWindowCatalogTest { assertEquals(204_800, ModelContextWindowCatalog.lookup("glm-5.1")); assertEquals(204_800, ModelContextWindowCatalog.lookup("glm-5-turbo")); assertEquals(1_000_000, ModelContextWindowCatalog.lookup("glm-5.2")); + assertEquals(204_800, ModelContextWindowCatalog.lookup("glm-5v-turbo")); + // Vendor alias for the K2.7 code model, including its -highspeed tier. + assertEquals(262_144, ModelContextWindowCatalog.lookup("kimi-for-coding")); + assertEquals(262_144, ModelContextWindowCatalog.lookup("kimi-for-coding-highspeed")); // Max line stays at 256k while plus/flash run 1M. assertEquals(262_144, ModelContextWindowCatalog.lookup("qwen3-max-2026-01-23")); assertEquals(1_000_000, ModelContextWindowCatalog.lookup("qwen3.6-plus-2026-04-02"));