What happened
The nace-ai/kapitsa project published version 0.2.24 on October 28, 2026. Its release notes describe two notable additions: a way to use Splash as an engine behind Kapitsa’s server, and use of a pack’s own MTP head for speculative drafting with Qwen3.5 MLX packs. The announcement is part of a larger release that also contains documentation, benchmarks, and fixes; the release notes do not frame every listed change as a new user-facing capability.
For Qwen3.5, the release says that MLX packs containing their own MTP head—identified by the language_model.mtp.* files and associated with Jundot’s oQ-mtp packs—can use that head to draft tokens instead of relying on plain decoding. The reported comparison is for a 27B oQ4e pack: 16.1 to 37.1 tokens per second on an M5 Pro, with the same greedy text. That is a throughput result reported by the project, not a general benchmark across Macs or packs. The release excerpt does not provide the test setup, prompt, run count, or variance.
The Splash work adds an engine option to Kapitsa’s serving command: kapitsa serve --engine splash. The release notes describe a native engine protocol v11 client codec, a child-process runtime for splash serve-native, discovery of Splash’s binary and model folder when flags are omitted, and documentation and checks. The authors say the protocol codec was checked against Splash’s own bytes. They report token-identical output to plain Splash on three requests, and identical token IDs on three requests in the documented A/B checks. The notes also list handling for client history and cache salt, opt-in conversation salt and persistent cache, a memory plan, and Splash sampling defaults.
The release identifies Splash 1.4.0 in its protocol and seed-rule notes. Several implementation details are explicitly limited: the runtime was tested against a fake engine in one step; certain cache, salt, and memory-plan behavior was not verified on a GPU; and the benchmark notes say results were not verified on other Macs. The release excerpt gives no price or commercial-access terms. It describes a software release and command-line option, rather than a hosted service or a new general-availability offering.
What the archive adds
This is a new release with additional facts on a known Kapitsa story. Kapitsa v0.2.15 covered stability fixes and faster MoE prefill; Kapitsa 0.2.22 covered EXL3 support and a Flash-Next benchmark. Version 0.2.24 extends that thread with a Splash engine path and Qwen3.5 MTP drafting, rather than restating either earlier release.
Trend signal
The release offers a small, concrete example of the harness architecture outliving model choice pattern: Kapitsa’s serving interface can route work through a distinct Splash engine, while the model pack’s own drafting head changes decoding behavior. This is evidence of an integration boundary in one local inference project, not proof that this architecture is becoming a general industry standard.
Why now
The release is dated October 28, 2026, and combines these changes in version 0.2.24. The supplied source does not explain why the project chose this release date or connect the timing to an external event.
Why it matters
For teams running models locally, the release highlights two separate levers: choosing an engine under a serving interface and using model-pack-provided drafting components to accelerate decoding. The reported Qwen3.5 result is noteworthy because the project says the faster run produced the same greedy text, making it more relevant than a speed figure alone. It is still a project-reported result on a particular pack and M5 Pro, so builders should treat it as a reason to test their own workload rather than as a promised speedup.
The Splash backend may also reduce the need to choose between Kapitsa’s serving interface and Splash’s engine, but the supplied material establishes only a narrow set of checks: three requests for token identity, protocol-byte checks, and several features not GPU-verified. A team considering it should validate its own model loading, cache behavior, memory use, and failure handling before depending on it. The practical consequence is incremental: local inference stacks can expose more than one engine and can exploit draft heads packaged with a model, but operational confidence still depends on workload-specific testing.
Limits and uncertainty
The release notes do not establish that the reported throughput applies beyond the 27B oQ4e pack on an M5 Pro, nor do they provide enough benchmark methodology to assess variance or reproduce the result exactly. “Same greedy text” does not establish equivalent behavior for sampling or other decoding settings. The Splash identity checks cover three requests, not broad compatibility. The source explicitly says some cache, salt, and memory-plan behavior was not verified on a GPU and one benchmark was not verified on other Macs. It also does not establish production reliability, comparative performance against other engines, or pricing and support terms.
Reader takeaway
Kapitsa 0.2.24 gives local-inference builders a Splash-backed serving option and reports a 2.3x Qwen3.5 drafting speed increase for one Apple Silicon setup, with the same greedy text. Both are worth testing if they match your stack; neither the speed result nor the backend checks should be generalized beyond the limited evidence in the release notes.