v0.6.3

What's Changed

• Fix APC splitting on multimodal inputs by @lucasnewman in #1311
• Fix Qwen quantized KV cache prompt state by @Blaizzy in #1313
• Fix Gemma 4 unified silently dropping image/video inputs (video processor rejects standard HF config keys) by @sockeye44 in #1321
• Fix thinking enabled by default for chat templates by @lucasnewman in #1316
• Fix APC exact disk-hit entries not promoted to in-memory LRU by @mikeatlas in #1324
• Align Qwen3-VL visual masks to the current prefill / decode window by @lucasnewman in #1325
• Fix LFM2.5 VL model loading by @lucasnewman in #1328
• Fix Phi 3.5 VL eos tokens by @lucasnewman in #1326
• Fix Qwen3-VL deepstack visual embeds misaligned during chunked prefill by @sockeye44 in #1332
• Fix --system prompt ignored by one-shot generate by @sockeye44 in #1330
• Improve speculative prefill when using MTP drafters by @lucasnewman in #1334
• Respect generation_config sampling defaults by @Blaizzy in #1337
• Add server thinking defaults by @Blaizzy in #1342
• Add Gemma 4 DLM by @Blaizzy in #1347
• Fix DiffusionGemma long-context prefill by @Blaizzy in #1348

New Contributors

• @sockeye44 made their first contribution in #1321
• @mikeatlas made their first contribution in #1324

Full Changelog: v0.6.2...v0.6.3