Models are becoming more capable per byte. Phones are gaining memory and faster silicon. Where do those curves meet—and what else has to be true?
Early-to-mid 2030s is a useful engineering scenario to explore. It is not an evidence-backed arrival date for full frontier intelligence. Hardware fit can be modeled; preservation of frontier capability after compression remains the decisive unknown.
“Frontier” is a bundle of abilities, not a parameter count or an IQ score. The target here is the capability available on 12 September 2026, exemplified by GPT‑6 Astra and Claude Fable 5. Their release claims establish the reference cohort; they do not reveal a minimum equivalent local model size.1, 2
Match a frozen reference across reasoning, coding, knowledge, instruction following, multilingual use, vision and tool selection. For a future validation study: require each domain to fall within a predeclared margin, such as 5 percentage points in task success, at matched tools and budgets. This brief has not run that study.
A future premium phone with locally stored weights, local computation and local documents/tools. Default engineering gates: an 8K-token prompt, ≤5 seconds of prompt processing, ≥20 generated tokens/second and ≤4 watts for the AI workload. These are chosen usability targets, not a universal definition.
A phone could match today’s model while cloud systems keep improving through larger training runs, longer reasoning and more agents. Future cloud parity is a moving target. Also, an offline model cannot reproduce a connected assistant’s live web, remote databases or server-side tools.
The falling curve is the assumed memory required to preserve the frozen capability. The rising curve is RAM available to the AI. Their intersection is the first hurdle; bandwidth, prompt processing and energy determine whether the result is usable.
The dotted reference doubles RAM every two years from the same starting point. It is deliberately a counterfactual: Moore’s law does not promise that phone RAM grows at that rate. Everything to the right of the starting point is a scenario projection.
1× means all modeled resource constraints pass. This index measures engineering headroom, not intelligence. Preset lines keep the same starting workload, phone and user requirements as your selection; they change five annual improvement assumptions. They are not confidence intervals.
The matrix below recomputes all resource gates for different starting weight footprints and footprint halving times. Other assumptions remain as currently selected. Later-than-horizon cells mean no crossing through September 2038.
Historical comparisons show that useful capabilities can become much cheaper and smaller. They do not provide a stable exchange rate between “frontier intelligence” and gigabytes. Below, measurements and vendor claims retain their original scope.
Stanford’s AI Index reports a shift from PaLM 540B to Phi‑3-mini 3.8B for models exceeding 60% on MMLU. That is evidence of substantial capability diffusion on a specific test, not proof of equal general intelligence.3
Google reports Gemma 4 E2B using under 1.5 GB on some devices with low-bit weights and memory-mapped embeddings. Device, backend and context matter; mapped storage is not the same thing as a full model download or a universal peak-RAM budget.4
Qualcomm reports 37% faster NPU performance and 16% better performance per watt for Snapdragon 8 Elite Gen 5 versus its predecessor. These are different metrics, and neither translates directly into the same gain in sustained LLM decoding.5
Within Google’s published Gemma 4 evaluation table, E4B trails 31B by very different amounts across tasks. This comparison uses one model family’s reporting, reducing—but not removing—evaluation differences. It is not a comparison with the September cloud frontier.6
| Reported task score | E4B | 31B |
|---|---|---|
| GPQA Diamond | 58.6% | 84.3% |
| LiveCodeBench v6 | 52.0% | 80.0% |
| MRCR v2, 128K | 25.4% | 66.4% |
Published model scores are not verified scores of the quantized phone builds. Precision, thinking budget and deployment settings must be tested again.
Training efficiency: Ho et al. estimate a roughly eight-month compute halving time on historical language-modeling evaluations (95% interval: about 5–14 months). That does not measure an inference-memory halving time.7
Scale dependence: later experimental work finds small-model algorithmic progress slower than a simple transfer of those estimates suggests; a separate note identifies possible selection bias in observational estimates.8, 9
API price: the AI Index reports a >280× decline at a GPT‑3.5-level MMLU threshold between November 2022 and October 2024. Those endpoints span about 23 months. Prices also reflect hardware, batching, utilization, margins and competition. They are not a measurement of on-device energy or model size.3
Model weights are only part of memory. Decoding repeatedly reads weights and cached context. Processing a long prompt can instead be limited by arithmetic. More accelerator TOPS can improve one stage much more than the other.10
At 20 generated tokens/second, 2,000 hidden reasoning tokens take 100 seconds before counting the answer. For a total 2,500-token response, that is 125 seconds. A smaller model that thinks longer may fit RAM while missing the interactive experience. Test-time compute can improve selected tasks, but its benefit depends on the problem and model.14
About 3.3% of an assumed 20 Wh battery, for the AI workload alone. The display, radios, OS and other apps consume additional energy. Four watts continuously would use that battery in five hours even before those other loads. These are arithmetic examples, not measured phone battery tests.
The forecast’s “4× / 16× thinking” control holds answer throughput constant by requiring more internal tokens/second. Allowing minutes per answer would produce a different convergence date.
The familiar two-year doubling concerns transistor counts. It does not guarantee matching growth in affordable RAM, memory bandwidth, useful tokens per joule or battery capacity. The end of classic Dennard scaling also separated transistor density from automatic power benefits.15
Smaller transistors, nanosheets, larger caches and matrix accelerators can improve efficiency. TSMC’s N2 announcement described 10–15% more speed at the same power or 25–30% less power at the same speed, rather than a free doubling of everything.16 The model uses separate growth assumptions for useful compute, bandwidth and energy efficiency.
Samsung lists 12 GB and 16 GB configurations for Galaxy S26 Ultra. The model’s 8 GB usable budget is an analyst allowance after OS/apps, not an OS guarantee.17 Samsung has announced LPDDR5X packages up to 32 GB and 10.7 Gb/s per pin; component capability does not mean mass-market phones ship with it.18
Apple attributes up to 40% better sustained performance in iPhone 17 Pro to A19 Pro paired with improved cooling.19 Samsung’s announced UFS 5.0 reaches 10.8 GB/s sequential transfer, but flash bandwidth and access patterns differ from RAM. Offloading weights trades memory capacity for I/O and energy.20, 21
No single equation can credibly include every contribution to LLM performance. This map covers the major model, hardware, system and deployment mechanisms. “Net shrink rate” bundles overlapping improvements; counting each again would exaggerate progress.
| Factor | How it helps | What can get missed | Forecast treatment |
|---|---|---|---|
| Training compute & data | Larger training budgets, better data mixtures, curricula and longer training can put more capability into a given model size.22 | Training-optimal and inference-optimal size differ. Synthetic data quality, data coverage and teacher access can constrain transfer. | Net shrink rate; no training hardware on phone |
| Distillation & post-training | Teacher outputs, reasoning traces and reinforcement learning can improve smaller models. DeepSeek released distilled variants from 1.5B to 70B.23 | Selected math/coding gains do not prove broad knowledge, robustness or long-task equivalence. Teacher-generated tokens still cost resources. | Net shrink rate; quality gate unresolved |
| Quantization & low-bit training | 4-bit packing gives a raw 4× weight reduction versus 16-bit. Quantization-aware or native low-bit training can go further.12, 13 | Metadata, outliers and higher-precision operations reduce savings. Quality can fall. Precision cannot halve indefinitely. | Already in starting footprint; net future gains + floor |
| Sparsity, pruning & MoE | Use fewer weights per token. Gemma 4’s MoE has 25.2B total parameters and 3.8B active.6 | Inactive experts still need storage and usually residency. Routing, cache misses and uneven expert access limit ideal speedups. | Active fraction changes traffic/compute, not total weights |
| Attention & state architecture | GQA, shared KV, sliding windows, recurrent/state-space or hybrid designs can reduce context memory and traffic.11, 6 | Long-context retrieval and exact recall must survive. Efficient attention kernels do not remove the need to read model weights. | Separate KV/state improvement; approximate attention term |
| Speculative & multi-token decoding | Draft several tokens, then verify. Gemma 4 includes matching MTP drafters for local inference.24 | The drafter takes RAM and compute; acceptance rates vary. Algorithm/backend/device support matters. | Within effective throughput assumptions; no extra multiplier |
| Reasoning, search & ensembles | More inference-time work can solve harder problems, including with smaller models.14 | Hidden tokens, candidate rejection and parallel agents raise latency, cache residency and joules per finished task. | Reasoning multiplier; concurrent agents excluded |
| Retrieval, tools & memory | Local search, a calculator, code execution and personal context improve the complete assistant. | A stronger system is not necessarily a stronger base model. Fresh web data and cloud tools break full offline equivalence; local indexes need storage. | Must match evaluation tools; additional app workload excluded |
| Multimodal perception | Efficient audio/vision encoders or unified architectures can reduce processing stages.25 | Camera/video input, audio output and continuous sensing add tokens, compute and power; text-only fit is not full multimodal fit. | Required in quality contract; text resource proxy only |
| Compute silicon & kernels | NPUs, GPU matrix units, fused operators, compiler tuning and cache locality improve usable arithmetic.5, 10 | Peak TOPS mix precisions and sometimes sparsity conventions. Unsupported operators can fall back to slower processors. | Sustained effective TOPS; prompt + decode compute gates |
| RAM capacity & bandwidth | More affordable RAM fits weights; wider/faster memory moves them. Packaging and larger caches can help both. | RAM capacity, bandwidth and flash storage are three separate budgets. OS app limits and other tasks reduce availability. | Separate usable RAM and effective bandwidth curves |
| Battery, thermal design & duty cycle | Lower joules per operation/byte and better cooling sustain work longer. | Hot surroundings, a charging phone, battery aging and other apps can lower the power budget; a peak burst is not a long session. | Fixed AI power budget and declining energy coefficients |
| Economics & deployment | Demand may justify more RAM, better packaging and hardware/software co-design. | Bill of materials, manufacturing yield, DRAM supply, release cycles and consumer replacement rates delay broad adoption. | Premium-device scenario; no mass-adoption forecast |
| Availability & reliability | Open weights, local runtimes and good OS integration make technically possible inference accessible. | Licenses, withheld weights, update policies and app distribution can block deployment. Benchmarks can saturate or miss rare, consequential failures. | External conditions; must verify before declaring arrival |
The most defensible near-term expectation is uneven progress: compact specialists first, broader assistants later. A high score on one benchmark, an impressive demo and an all-day local assistant are different milestones.
Small local multimodal and reasoning models exist. They are useful starting points, with substantial variation by task. “On-device AI exists” is already true; “it matches the current frontier broadly” remains unproven.
Track the smallest phone-deployable model that matches the frozen reference on held-out tasks. Measure peak RAM, prompt latency, sustained speed and joules per successful task on real devices. Further capability diffusion is plausible; the pace is uncertain.
Default improvement scenarios clear the modeled resource gates in this window. Broad equivalence needs independent evidence. Long context, extensive reasoning, dense architectures or a compression floor can push feasibility outside it.
The roadmap for mobile hardware is visible in parts. A reliable curve for the minimum bytes needed to preserve all of today’s frontier abilities is not. The document’s most consequential slider is therefore a research question, not an engineering specification.
Earlier: independently evaluated 4–10 GB models matching the frozen frontier at a matched reasoning budget; strong gains in sustained tokens per joule; more usable RAM in actual phones.
Later: stubborn long-task or knowledge gaps after compression, quality loss in phone quantization, stalled bandwidth growth, RAM remaining expensive, or reasoning lengths growing faster than throughput.
METR’s time-horizon work is useful for tracking long-task competence, but it measures task difficulty at a success threshold, not how long a model thinks or how fast phone inference runs. METR also documents sensitivity and limits in these measurements.26
The baseline curves do not fit the last two years of capability growth, and they contain no explicit recursive-improvement mechanism. They assume constant improvement rates. This additional experiment asks how much earlier resource feasibility could arrive if AI research acceleration produces faster reductions in the size of an equivalent model.
OpenAI’s September 6 report describes increasing agent use and experimental activity, but notes that human decisions and compute can constrain the whole research process.27 Anthropic reports faster engineering and experimental work while explicitly saying fully autonomous successor development has not yet been achieved.28
These are first-party observations, not independently measured rates of general intelligence growth. More code, more experiments and stronger benchmark scores measure different things. Neither report supplies a reliable rate for shrinking a frontier-equivalent model into phone memory.
More capable AI → more effective AI research → better algorithms and models → more capable research assistants. If each cycle strengthens the next, improvement rates can rise over time. If experiments, data, compute or validation become the bottleneck, that acceleration can level off. METR models these possibilities and stresses that feedback strength remains poorly constrained.29
For pockets, the relevant transfer is from research gains to fixed-quality inference efficiency. Gains spent on larger cloud models or more test-time compute do not automatically make a phone model smaller. Hardware design may accelerate too, but fabrication and product cycles add separate delays.
This compares the main controls above with an additional acceleration of weight-footprint reduction. Hardware and KV/state trends remain the same. It is a conditional sensitivity calculation, not a calibrated forecast of RSI or a simulation of autonomous AI research.
Let u be elapsed years after the assumed onset plus deployment delay. The additional research-rate multiplier is m(u) = min(cap, (1 + gain)^u). The weight-shrinkage rate multiplier is 1 + transfer × (m(u) − 1). We integrate that rate to compute cumulative weight reductions. Setting gain = 0, transfer = 0 or cap = 1 reproduces the baseline exactly.
The model does not estimate when full recursive self-improvement will arrive, or derive acceleration from model capability endogenously. Its parameters are scenario choices. It isolates a possible extra effect instead of multiplying historical benchmark, training-efficiency and API-price trends together. It also assumes weight reductions reduce active arithmetic, as in the base model; gains from precision alone may transfer less strongly.
Earlier resource feasibility remains conditional on capability retention, mobile-accessible weights and practical software support. An accelerating cloud frontier can simultaneously move further ahead even as today’s fixed target becomes cheaper.
This is a scenario analysis informed by primary papers, model cards, hardware specifications and original benchmark analysis. It is not a fitted statistical model. No new device benchmarks were run, no probability distribution is estimated, and no private model size is inferred.
Let t be years since 12 September 2026. Units are decimal GB, seconds, joules and tera-operations (10¹² operations). User controls and presets set the inputs. Weight bytes, active compute and traffic shrink together as an explicit simplifying assumption; this may overstate gains when compression mainly reduces precision rather than computation.
The arithmetic term approximates two operations per active parameter at a nominal 4-bit weight representation. Attention is an illustrative 32-layer, 4096-hidden-dimension full-attention workload; its average prompt cost is approximated by half its final-context decode cost. The same state-efficiency trend reduces the attention term as a simplifying architecture assumption. The energy coefficients are analyst choices: 0.08 J/GB transferred (80 pJ/byte) and 1 J/tera-op (1 pJ/op), plus 1 W idle. They are not measurements of a named phone.
Prompt compute assumes the effective TOPS value is sustainable within the selected power envelope. This simplified model does not separately simulate prompt energy, kernel launch overhead, model cold-load time, application limits, concurrent agents or audio/video processing. The ≥20 tokens/s gate is about total generation throughput; time to first visible answer may include hidden reasoning. The reported convergence remains an approximate engineering milestone.
| Input | Default | Provenance / interpretation |
|---|---|---|
| Equivalent packed weights | 100 GB | Analyst hypothesis, approximately the raw weight size of 200B parameters at 4-bit. Not a claim about any frontier model. Sensitivity includes 25–200 GB. |
| Footprint halving / lower floor | 24 months / 4 GB | Analyst scenario for fixed quality, not the historical eight-month training result. A nonzero floor prevents endless shrinkage. |
| Usable RAM | 8 GB, +15%/year | Illustrative allocation on a premium 12–16 GB phone; future growth is assumed.17 |
| Effective bandwidth | 40 GB/s, +18%/year | Assumed sustained model traffic. A 64-bit bus at 10.7 Gb/s per pin would have 85.6 GB/s theoretical transfer; actual device bus width and utilization must be verified.18 |
| Effective compute | 8 TOPS, +25%/year | Assumed useful low-precision operation rate, already discounted from peak marketing specifications. |
| Active weight fraction | 25% | Hypothetical sparse/MoE architecture. Set to 100% to test a dense model. No further annual sparsity multiplier is added. |
| Runtime / state | 2 GB / 1.074 GB at 8K | Runtime allowance includes working buffers and unmodeled metadata. State uses an illustrative GQA cache; state footprint halves every six years by assumption. |
| AI power / idle | 4 W / 1 W | Analyst design budgets. Dynamic energy falls by a factor of 1.15/year, after shrinking the workload itself. No battery-capacity growth is assumed. |
| Quality and price trends | No additional multiplier | Training efficiency, distillation, quantization, API-price declines and benchmark improvements are contextual evidence, not independent multipliers. |
There is no accepted scalar for intelligence. A rigorous follow-up would freeze tasks and frontier-model versions; record success, quality, tool access and reasoning budgets; test multiple local candidates at their deployed quantization; and measure peak memory, sustained thermal behavior, full-response latency and energy per successful task. Use several model families and phone generations to estimate a fixed-quality efficiency frontier with uncertainty. Avoid contaminated or saturated benchmarks and report each domain separately.
The present model extrapolates smooth exponential hardware growth through 2038, with a weight floor but no hardware saturation curve. Real releases are discrete and constrained by cost, heat and supply. Beyond the next three to five years, the values are increasingly speculative. Crossing on a premium phone is also different from reaching most people’s existing phones.
All sources checked on 12 September 2026. Vendor performance claims are attributed as such. Papers support mechanisms and historical findings; none endorses the forecast dates or assumed starting footprint.