llama.cpp 0.4.1 adds three new model architectures

llama.cpp 0.4.1 is a maintenance patch on 0.4.0. It adds three model architectures, reorganizes how JSON schemas are handled, and pulls the ggml compute library up from v0.23.0 to v0.24.0.

Three architectures join the list

The release adds Maple 20B-A1B, a ternary MoE architecture running on CPU first (PR #27000), the Tencent Hy 4 preview architecture under the tag hy_v4 (#28127), and Spark2.5 (#27868). A --fuse-qkv flag was added to the HF-to-GGUF conversion path (#22780).

Existing families get repairs rather than features. Kimi-K3 can now roll back recurrent state (#28466), MTP context KV cache allocation was fixed for DeepSeek2 and GLM-MoE (#28630), and GDN normalization for Qwen, Kimi and GLM models switched from max to rsqrt (#28068). Granite and Granite3 MoE report their parameter counts correctly again (#28643, #28632).

Plumbing: schemas, parsers, logs

JSON schema handling moved onto a shared representation called common_schema (#28736), and the specialized chat parsers were split out into common/parsers (#27764). For anyone running constrained decoding or tool calls in production, that plumbing is arguably the more consequential half of this release.

Structured logging arrives as JSONL through --log-jsonl and the LOG_JSON environment variable (#28437, #28586). The long-deprecated --mmap, --mlock and --direct-io arguments are gone (#28334), so launch scripts that still pass them need editing before the upgrade.

Server and web UI

Router child processes are now watched by a single monitor thread (#28555), backed by new server_subproc and waiter types in server-common.h. A hang in LRU handling when several requests hit the same model is fixed (#28539), and model downloads are allowed even once --models-max is reached (#28530).

Context checkpoints are no longer evicted on prompts shorter than checkpoint_min_step (#28302), and --reasoning-preserve is enabled by default (#28437). On the UI side, chat messages mount lazily (#28460), assets get a cache (#28802) and are embedded directly via CMake (#28445), plus a fix for MCP image attachments inside tool blocks (#28089).

Exactly one C API signature changes: llama_sampler_chain_n() returns int32_t rather than int (#28631).

What the release page does not say

The notes give September 14 as the date but no year. The page's assets block failed to load when we fetched it, so binary file names, platforms and sizes could not be checked. There is no contributor count either; the page only notes that five commits have landed on the branch since this tag.

This story rests on the release page alone - a single source. It carries no benchmarks, no speed comparisons and no claim about how mature the three new architectures are, so neither does this article.