What happened
Kapitsa’s GitHub release notes for version 0.2.35 say that every nace-ai KQuant pack now writes with its drafter. The notes give one specific command and hardware result: kapitsa serve --model nace-ai/Qwen3.5-9B-MLX-Q4-KQuant writes at about 145 tokens per second on an M5 Pro, up from about 52 tokens per second. This is a project-reported result for that model pack and machine, rather than a general performance guarantee for every Kapitsa installation.
The same release describes faster handling of long prompts on M5 Macs. For the 9B model, Kapitsa says a 4,000-token prompt is read in 2.4 seconds instead of 3.9 seconds. That figure concerns prompt reading, not necessarily total time to a finished response: an end-to-end interaction also depends on generation length and the rest of the serving workflow. The release presents the number as a comparison for KQuant packs on M5 hardware.
The configuration change is limited to one Apple Silicon generation. On M5 Macs, the default 9B option is now the project’s Q4-KQuant pack. Kapitsa describes answers from that pack as closer to the original model, while saying writing is about 5% slower. The release therefore frames the default not as a pure speed choice but as a quality-and-throughput trade-off. M3 and M4 Macs keep the older pack, so version 0.2.35 does not make the same default change across all Apple Silicon systems.
GitHub marks version 0.2.35 as released on November 12 and identifies the release commit as 30b18f9. The notes point to project issues #1210 and #1191 for the drafter-writing change, #1199 for long-prompt reading, and #1196 for the M5 default. They do not announce a separate pricing tier, hosted-service access change, or new hardware requirement. The release is principally a model-pack behavior and performance update for Kapitsa users on Macs.
What the archive adds
New Runtime covered Kapitsa v0.2.15 on October 24, when the project added stability fixes and faster MoE prefill. An October 28 note on Kapitsa 0.2.22 covered EXL3 support, 8-bit expert packs, and a Flash-Next result on an M5 Max. This is a new release in that continuing Kapitsa thread: it introduces an M5-specific default and new KQuant measurements rather than restating those earlier changes.
Trend signal
This weakly supports agent economics moves to completed work. Faster local generation and prompt ingestion can lower latency in an agent loop, but Kapitsa reports tokens per second and prompt-read time rather than the cost of verified completed work. The more immediate signal is practical: hardware-specific pack defaults can shape the quality and responsiveness that a local-model workflow actually delivers, even when the nominal model size remains unchanged.
Why it matters
For a team assessing local inference, the release makes a useful distinction between throughput and output behavior. The M5 default is described as closer to the original model but about 5% slower to write, while the stated M5 Pro example shows a much larger gain for one KQuant serving setup. Those claims may be relevant to teams choosing whether a local 9B model is quick enough for interactive tools, evaluation loops, or internal assistants.
The prompt-reading result is also potentially material for workloads that send substantial context repeatedly. A 4,000-token prompt is a concrete test case, not proof that all context sizes improve by the same proportion. Builders should treat the release as a reason to benchmark the revised default against their own prompts, response lengths, and acceptance criteria, rather than treating a token-rate result as a complete production comparison.
Limits and uncertainty
Kapitsa’s release notes contain the performance claims, and they have not been independently checked here. The 145-tokens-per-second example names an M5 Pro but does not specify the full hardware configuration, memory capacity, power conditions, sampling settings, repeated-run distribution, or resource use. The long-prompt comparison identifies M5 Macs and a 9B model but does not describe a range of context sizes or total request latency.
The claim that answers are closer to the original model is not accompanied in the release notes by an evaluation method, test set, or quality score. Nor do the notes quantify how drafter-assisted writing varies by model pack, prompt type, or generation length. The stated 5% writing-speed difference should therefore be tested against the workload that matters to each user.
Skill to build now
Create a small local evaluation harness before changing a default model pack. Run the previous and Q4-KQuant options on a fixed set of representative prompts, recording prompt-read latency, generation rate, end-to-end response time, and a human quality decision. Use the same Mac, request settings, and output-length cap for both runs. Adopt the new default only if its quality improvement remains worthwhile after the measured latency trade-off for the actual task.
Reader takeaway
Kapitsa 0.2.35 is an M5-focused local-inference update, not a universal Apple Silicon switch. It makes Q4-KQuant the default 9B pack on M5 Macs and reports faster KQuant writing and prompt reading, while explicitly acknowledging a small speed cost for the new default. Validate the quality-versus-latency trade-off on your own workload.
Open question
How does Q4-KQuant compare with the previous pack on answer quality and end-to-end latency across representative workloads on M5 Macs?