
llama.cpp Speeds Selected Mac Tests by Fusing Computation
llama.cpp merged selective Apple-GPU optimizations on September 19, with a contributor-reported generation gain of about 16% in one M2 Ultra test.
Read More
llama.cpp added multi-token prediction for GLM-4.5-Air, lifting one contributor benchmark from 74.55 to 88.73 tokens per second, as Singapore's power prices jumped.
ZharfAI ResearchAugust 24, 20267 minutesSource: llama.cpp
llama.cpp release b10603 added multi-token prediction support for GLM-4.5-Air, giving local operators a new speculative-decoding path whose contributor benchmark raised mean throughput from 74.55 to 88.73 tokens per second on four RTX 3090 GPUs when one draft token was allowed. The same table shows why the result needs qualification: the gain shrank with two draft tokens and became a slowdown with three. Singapore's July inflation release supplies the physical counterweight. Electricity-and-gas inflation swung from −2.9% in June to +8.7% year over year, helping lift core inflation to 2.0% and all-items inflation to 2.2%. The developments are not causally linked. Together they expose the real AI cost equation: software can make each inference step more productive, but acceptance rates, hardware, cooling, electricity tariffs and capital still decide whether faster tokens become cheaper, reliable work.
The freshest useful AI signal is a small systems release with a large operating lesson. llama.cpp published b10603 at 18:45:59 UTC on August 23, adding multi-token prediction, or MTP, support for GLM-4.5-Air through pull request 26534. The implementation builds a draft graph for the model's prediction head, lets the converter retain or separate the MTP tensors and lets the runtime load a combined, trunk-only or MTP-only GGUF. In practical terms, the system can propose a token ahead and let the main model validate it, seeking more accepted output from each expensive target-model pass. This is an inference shortcut, not a new model release or a guarantee of lower bills.
The pull request's clearest benchmark used quantized target and draft models across four RTX 3090 GPUs. With Q4 target and Q4 draft weights, the author's baseline averaged 74.55 tokens per second; allowing one draft token raised the mean to 88.73, a reported 1.19× improvement. Two draft tokens produced 1.12×, while three fell to 0.97×—slower than the baseline. The Q8 comparison followed the same shape: 1.10× with one, 1.06× with two and 0.95× with three. A mixed Q4 target and Q8 draft reached 1.16× with one draft token but only 0.92× with three. The code currently asserts a single MTP block. Those details matter more than the best headline number: extra speculation pays only when enough proposed work is accepted.
This evidence is useful but narrow. It comes from the contributing developer, not an independent laboratory, and it measures tokens per second under specific quantization, batch, offload and multi-GPU settings. It does not report energy consumption, latency distribution, quality changes, memory pressure, interconnect overhead or cost per accepted token. A separate MTP-only GGUF artifact linked from the pull request was available on Hugging Face by 10:55:38 UTC on August 22, which makes the path reproducible for other testers, but availability is not validation. Production users should treat 1.19× as a configuration result to reproduce, not a universal property of GLM-4.5-Air.
The engineering implication is still meaningful. Speculative decoding changes the optimization target from raw generation speed to accepted work per unit of scarce capacity. A draft that runs cheaply and predicts well can reduce the number of serial target passes; a deeper or poorly matched draft can add memory traffic and computation that the verifier discards. The right dashboard therefore joins acceptance rate, p50 and p95 latency, tokens per joule, GPU-hours per completed task, output-quality checks and failure or retry rates. If an agent calls tools or waits on networks, faster token emission may not even be the binding constraint. Software leverage is real only after the full workflow becomes cheaper or more reliable.
Singapore's July consumer-price data, released at 13:00 local time on August 24, shows why that full cost boundary matters. Monetary Authority of Singapore core inflation rose to 2.0% year over year from 1.6% in June, while CPI-All Items inflation increased to 2.2% from 1.9%. Month over month, core prices rose 0.3%, even as the all-items index declined 0.2%. The main acceleration came from electricity and gas, services and food; accommodation inflation also increased. This is a macroeconomic release, not a measure of data-centre power prices, and it does not show that AI demand caused the change.
The electricity detail is nonetheless an unusually clear example of physical cost pass-through. Electricity-and-gas inflation moved from −2.9% in June to +8.7% in July because the regulated electricity tariff rose sharply. The official report explains that the tariff reflects the average natural-gas price in the preceding quarter—April through mid-June for the July–September tariff—so fuel costs arrive with a lag. Services inflation increased from 1.5% to 1.7%, food from 2.1% to 2.2% and accommodation from 0.6% to 0.8%; retail-goods inflation eased from 1.7% to 1.4%, and private-transport inflation from 8.4% to 8.0%. A softer monthly headline therefore coexisted with faster annual core inflation and a sharp utility reversal.
For AI operators, the comparison is operational rather than causal. An inference optimization can lower GPU time per accepted output, yet a tariff reset, cooling load, reserved-capacity contract or financing cost can absorb the saving. The unit-economics test is not whether a benchmark gets more tokens from the same card; it is whether the deployment delivers more correct, useful tasks per rial or dollar after electricity, cooling, networking, storage, observability and rejected speculative work. The same logic applies to cloud buyers: lower compute seconds help, but only if provider pricing and minimum commitments pass the efficiency through instead of retaining it as supplier margin.
Singapore's outlook also makes the two-way macro link explicit. MAS and the Ministry of Trade and Industry expect both core and all-items inflation to average 1.5%–2.5% in 2026. They expect core inflation to remain elevated into 2027 before moderating around mid-year if energy costs ease, while warning that oil prices, transport fares, weather and input-cost pass-through tilt risks upward. Their downside cases include tighter global financial conditions or a pullback in AI-related investment; stronger information-technology investment could instead create demand spillovers. AI is therefore both a capital-spending variable in the outlook and a workload whose economics remain exposed to the energy path.
The best interpretation avoids two opposite errors. It would be wrong to dismiss the llama.cpp result because it is not an audited production benchmark: open runtimes often turn narrow kernel or graph changes into meaningful capacity gains once independent operators reproduce them. It would also be wrong to annualize the top 19% throughput increase into a 19% cost reduction. The benchmark's own reversal at three draft tokens is the warning. Acceptance, model mix, memory placement and workload shape the result; electricity and capital determine how much of the technical gain reaches the income statement.
What to watch next is concrete. For GLM-4.5-Air, independent tests should publish acceptance rates, power draw, latency tails, quality parity and results beyond four RTX 3090s, including whether more MTP blocks or different draft quantization improve the curve. Operators should compare target-only and speculative runs at the completed-task level, not cherry-pick peak throughput. On the macro side, watch Singapore's October tariff reset, whether the July energy jump feeds services and food, and whether the assumed mid-2027 moderation remains credible. Today's defensible conclusion is modest but useful: inference software found a shorter path, while the energy system still owns the floor beneath it.
TagsAI inferencemulti-token predictionspeculative decodingGLM-4.5-Airllama.cppenergy costsSingapore inflationAI unit economics

llama.cpp merged selective Apple-GPU optimizations on September 19, with a contributor-reported generation gain of about 16% in one M2 Ultra test.
Read More
Liquid Compute's September 15 seed round backs a proposed market for compute price exposure, not an already approved futures exchange.
Read More
llama.cpp release b10707 stops an n-gram lookup from scanning all 256 possible sequence slots after it has found every sequence actually attached to a used KV-cache cell.
Read MoreIf this note maps to a real system in your organisation, start with the services page or a shipped case study.