llama.cpp b9979–b9990 — 12 builds in 24 hours: per-request reasoning budgets, NUL-byte truncation fix, reasoning leak fix, Minimax2 Eagle3
| Build | UTC | Change |
|-------|-----|--------|
| b9990 | Jul 13, 19:51 | Added Minimax2 EAGLE3 speculative decoding support |
| b9986 | Jul 13, 09:06 | Fixed reasoning content leakage with force-opened bare template prefills; Nemotron Nano v2 regression fix |
| b9985 | Jul 13, 08:32 | SYCL: fused top-k MoE kernel with GGML_SYCL_ENABLE_FUSION env variable |
| b9984 | Jul 13, 07:38 | Q2_K quantization extended to SYCL DMMV reorder pathway |
| b9982 | Jul 13, 01:28 | Server now honors per-request budget_tokens in chat completions — previously discarded silently |
| b9980 | Jul 13, 00:07 | Models with mmproj disabled no longer declared as multimodal on /v1/models endpoint |
| b9979 | Jul 12, 23:36 | Fixed silent prompt truncation caused by embedded NUL bytes in multi-turn mode |
budget_tokens in your per-request chat completion payload expecting to control reasoning length per call, those parameters were previously being silently discarded — the server was ignoring both the token count and message fields. Upgrade to b9982 to get per-request reasoning budget control. This is specifically useful for applications that want tight per-request cost control over reasoning depth without restarting the server. (2) b9979 — NUL byte truncation: If you send multi-turn conversation history that could contain NUL bytes (e.g., binary data embedded in tool call results, certain tokenization edge cases), messages after the NUL byte were being silently truncated. This is a correctness bug that can cause context loss without any error signal. Upgrade to b9979. (3) b9986 — reasoning leak with force-opened templates: If you use bare template prefill (a technique that forces the model to continue from a specific prefix) with models that have a <think> reasoning block, the reasoning content was leaking into the output rather than staying in the thinking section. This fix also resolves a Nemotron Nano v2 regression that was introduced in a recent build. (4) b9990 — Minimax2 EAGLE3: Minimax2 (the open-weight version of MiniMax-M3, which was added to vLLM v0.25.0 last week) now has EAGLE3 speculative decoding support in llama.cpp. EAGLE3 typically improves token throughput by 2–3× for models that support it. GGUF weights for Minimax2 that include the draft model are required — check the model card on HuggingFace. (5) b9985 — SYCL MoE fusion (Intel GPU users only): The fused top-k MoE kernel for SYCL is now available as an opt-in; enable with GGML_SYCL_ENABLE_FUSION=1. This reduces kernel launch overhead for MoE models on Intel Arc and Flex GPUs.