Two Years of Local AI on a Laptop: When Open Models Outpaced Moore's Law

Update

August 1, 2026: Added DeepSeek V4 Flash 0731 at 50, added a section on local coding via Terminal-Bench v2.1, and re-based every score in this post from Artificial Analysis Intelligence Index v4.0 to the current v4.1 — which moved most of them down and made the trend steeper, not shallower.

TL;DR

Between May 2024 and July 2026, the most expensive MacBook Pro you could buy stayed at 128 GB of unified memory. The hardware ceiling barely moved. But the smartest open-weight model you could actually run on it went from a score of 3 (Llama 3 70B) to 50 (DeepSeek V4 Flash 0731 on Unsloth's dynamic 2-bit GGUF) on the Artificial Analysis Intelligence Index v4.1.

That is 16.7× in 26 months, or a doubling of intelligence every 6.4 months.

Moore's Law (transistor count) doubles every 24 months. Local open-weight AI on a laptop has been improving almost four times as fast as Moore's Law, on completely unchanged hardware.

A caution on that ratio, up front: a 3 is close to the floor of v4.1, whose nine evaluations are hard enough that 2024-era models barely register, and ratios computed off a near-floor denominator overstate things. A sturdier anchor gives nearly the same answer — Llama 3.3 70B scored 9 in March 2025, so 9 → 50 in 16 months is a doubling every 6.5 months. Both roads lead to roughly six and a half.

Smartest open-weight model on a 128 GB MacBook Pro
Artificial Analysis Intelligence Index v4.1   (higher score better)

May 2024  Llama 3 70B          ███                                                  3
Oct 2024  Qwen 2.5 72B         ██████████                                          10
Mar 2025  Llama 3.3 70B        █████████                                            9
Oct 2025  gpt-oss-120B         ████████████████████████                            24
May 2026  DS V4 Flash          ████████████████████████████████████████            40
Jul 2026  DS V4 Flash 0731     ██████████████████████████████████████████████████  50
                               └────┘
                                  │
                       Moore's Law would predict a score of ≈ 6 here
                       (starting at 3, doubling every 24 months)

The hardware stood still

The premise of this post is simple. Buy the most expensive MacBook Pro on the market. What is the smartest open-weight model you can actually run on it, measured by a fixed benchmark? Repeat every six months for a little over two years.

Three generations of flagship Max chips. RAM ceiling never moved. Memory bandwidth grew about 50 percent, which matters for decode speed, but does not change which models can fit in memory.

What changed was the models.

The six snapshots

For each timepoint I picked the smartest open-weight model that:

1. Was released by that date.
2. Fits in 128 GB at a usable quantization. Q4 is the default, but mixed Q2 schemes (IQ2_XXS for routed experts plus Q8 on attention, shared experts, and output) count as normal too.
3. Runs at 5 tokens per second or faster on the then-current top MacBook Pro.

All scores are Artificial Analysis Intelligence Index v4.1 against the full-precision hosted model, re-read on August 1, 2026. Artificial Analysis has rebased the index three times in this window, so older press release numbers — including the ones in the first version of this post — are not directly comparable.

The progression 3, 10, 9, 24, 40, 50 is not linear. There are three discontinuities.

Discontinuity 1: sparse MoE arrives (August 2025)

For more than a year, the local ceiling was 70 billion dense parameters. Llama 3 70B, then Qwen 2.5 72B, then Llama 3.3 70B. The Mac memory bandwidth wall was the bottleneck: a 70B dense model at Q4 reads about 40 GB per token, capping decode at 8 to 12 tokens per second on M4 Max.

gpt-oss-120B broke this. 117 billion total parameters, but only 5.1 billion active per token. The MoE router selects a different subset of experts for each token, so decode is bandwidth-bound on only the active path. Result: 40 to 60 tokens per second on M4 Max, while the Artificial Analysis Intelligence Index score jumped from 9 to 24.

The model also shipped natively in MXFP4, which means there is essentially zero quantization quality loss on the local copy. The hosted benchmark and your laptop run the same weights.

Discontinuity 2: small dense reasoning catches up, huge MoE fits via Q2 (April 2026)

Two things happened within two weeks of each other.

Qwen3.6 27B (Reasoning) arrived on April 22, 2026. A dense 27 billion parameter model that scores 37 on the Artificial Analysis Intelligence Index. At Q4 it occupies 15 GB. On a 128 GB MacBook Pro, that leaves 113 GB of headroom for context, KV cache, or other apps.

DeepSeek V4 Flash arrived on April 24, 2026. 284 billion total parameters, 13 billion active. At full precision it does not fit on a laptop. But antirez published a GGUF using IQ2_XXS for routed experts (the bulk of the weights) and Q8 for attention, shared experts, and output. Total: 80.8 GB. Artificial Analysis Intelligence Index at full precision: 40.

Either of these would have taken the laptop ceiling above gpt-oss-120B. DeepSeek V4 Flash takes the headline by three points, but Qwen3.6 27B is the cleaner story: a 27B dense model that comes within three points of a 284B mixture-of-experts on the same benchmark.

Discontinuity 3: the weights stopped changing (July 2026)

On July 31, 2026, DeepSeek released DeepSeek V4 Flash 0731 under MIT. Artificial Analysis scores it 50, ten points above the April V4 Flash and six points above DeepSeek V4 Pro. Among open-weight models it lands in the top three, and it sits one point behind GPT-5.6 Luna (max) — a frontier closed model.

The interesting part is what did not change. Same architecture. Same 284 billion total parameters, same 13 billion active. Same pricing on the hosted endpoint. The ten-point jump is entirely post-training: better reinforcement learning, better agentic data, and a three-level reasoning_effort knob (low, high, max) that lets you buy accuracy with tokens.

Artificial Analysis attributes most of the gain to reduced hallucination rather than raised accuracy — hallucination rate fell about 12 percentage points while raw accuracy stayed roughly flat. On GDPval-AA v2, the agentic real-work eval, Elo went from 1189 to 1559. DeepSeek's own card reports Terminal Bench 2.1 at 82.7, NL2Repo at 54.2, and Cybergym at 76.7; the coding number gets its own section below.

For a laptop this is the cleanest result in the whole two-year window. Nothing about the model got bigger. No new RAM was required, no new quantization trick was needed, no architecture had to be invented. The same weights-shaped object got ten points smarter.

Because the architecture is unchanged, the community quants arrived within a day. Unsloth's dynamic GGUFs span UD-IQ1_S at 82.5 GB up through UD-IQ3_S at 116 GB; UD-Q2_K_XL at 96.8 GB is the sweet spot on a 128 GB machine, leaving about 30 GB for context and the rest of your desktop. UD-IQ3_XXS at 104 GB fits too if you keep context short. Full BF16 is far out of reach, as it always was.

It is worth being precise about what "top" means here, because for the first time in this series it needs an asterisk. DeepSeek V4 Flash 0731 is the third-ranked open-weight model overall, not the first. Kimi K3 (max) scores 57 and GLM-5.2 (max) scores 51. Neither fits: Kimi K3 is 2.8 trillion total parameters with 104 billion active, and GLM-5.2 is 753 billion with 40 billion active. Even at one-bit-something they are hundreds of gigabytes. DeepSeek V4 Flash 0731 is the best model that fits — which is the whole question this post asks, but it is no longer the same thing as the best open-weight model.

Open-weight models that fit on a 128 GB MacBook Pro, August 2026
Sorted by Artificial Analysis Intelligence Index v4.1   (higher score better)

Model                    Quant           Size                                                      Score
───────────────────────  ──────────  ────────  ──────────────────────────────────────────────────  ─────
DeepSeek V4 Flash 0731   UD-Q2_K_XL   96.8 GB  ██████████████████████████████████████████████████   50
DeepSeek V4 Flash (Apr)  Q2-mix       80.8 GB  ████████████████████████████████████████             40
Qwen3.6 27B Reasoning    Q4             15 GB  █████████████████████████████████████                37
Qwen3.6 35B A3B          Q4             19 GB  ████████████████████████████████                     32
Gemma 4 31B              Q4             17 GB  █████████████████████████████                        29
Gemma 4 26B A4B          Q4             14 GB  ██████████████████████████                           26
gpt-oss-120B (high)      MXFP4          63 GB  ████████████████████████                             24
GLM-4.6                  Q2-mix       ~110 GB  ███████████████████████                              23
GLM-4.5-Air              Q4             57 GB  █████████████████                                    17

Models on Hugging Face: Qwen3.6 35B A3B, Gemma 4 31B, Gemma 4 26B A4B, GLM-4.6, GLM-4.5-Air.

What about local coding?

Intelligence Index is a composite, and the sub-score most people running a model on their own laptop actually care about is Terminal-Bench v2.1, which is one of the nine evaluations feeding v4.1. It is 89 curated tasks spanning software engineering, system administration, data processing, model training, and security. The agent gets a real terminal and has to satisfy a verification suite programmatically. Scoring is pass@1 averaged over three repeats, so there is no partial credit for a plausible-looking plan that does not actually work.

This is the eval that best predicts whether a local model can be your coding agent rather than your autocomplete.

The frontier sits just under 90 percent. Artificial Analysis measures GPT-5.6 Sol (xhigh) at 89.5 percent, Claude Opus 5 (max effort) at 89.1, and GPT-5.6 Sol (max) at 88.0. The official Terminal-Bench leaderboard, which pairs models with specific agent harnesses rather than testing them in isolation, tops out slightly lower: Claude Code driving Fable 5 at 83.8 percent, Codex driving GPT-5.5 at 83.1.

DeepSeek's own model card reports 82.7 percent on Terminal Bench 2.1 for V4 Flash 0731. If that number holds up, the best model you can fit on a 128 GB laptop is within about seven points of the best coding agent in the world, which would have sounded absurd twelve months ago.

Three reasons to hold that loosely:

It is self-reported. Artificial Analysis publishes Terminal-Bench v2.1 as an input to the composite index but does not surface a per-model breakdown for most laptop-class models, so there is no independent replication of 82.7 to point at.

The harness matters more than people admit. On the official leaderboard the same model moves several points depending on what is driving it — Fable 5 scores 83.8 under Claude Code and 80.4 under Terminus 2; Opus 4.7 scores 68.9 under Claude Code and 66.1 under Terminus 2. A vendor-reported number is measured with a vendor-chosen scaffold. Whatever you run locally will not be that scaffold.

It assumes reasoning_effort: max on the hosted BF16 endpoint. Your UD-Q2_K_XL copy is a different set of weights, and max effort on a laptop means waiting through a very long chain of thought for every task in the suite.

The honest summary is that local coding went from unusable to genuinely competitive inside this window, and that the gap to frontier coding agents is now small enough to be worth arguing about — but nobody has yet published a careful Terminal-Bench run of a 2-bit local quant under a normal open-source harness. That is the measurement this post most wants and cannot cite.

Compared to Moore's Law

Moore's Law as originally stated covered transistor count: doubling every 24 months. Loosely interpreted as "capability doubles every two years", it gives a reference rate for technological progress.

Local AI on a MacBook Pro went from an Artificial Analysis Intelligence Index score of 3 to 50 in 26 months. That is 4.06 doublings, or a doubling every 6.4 months. Almost four times the pace of Moore's Law.

If local intelligence had followed Moore's Law strictly, July 2026 would look like a score of 6 — below Llama 3.3 70B, which is a 2024 model. Instead it looks like DeepSeek V4 Flash 0731 at 50.

Measured from the sturdier March 2025 anchor, the gap narrows only slightly: Llama 3.3 70B at 9 to DeepSeek V4 Flash 0731 at 50 is 2.47 doublings in 16 months, a doubling every 6.5 months, against Moore's Law's prediction of 14. The two estimates agreeing is the reason I trust the shape of this curve more than any single point on it.

Even more strikingly: Moore's Law was about hardware getting faster. In this story, the hardware barely changed. All the gains came from software and model design.

Why it happened

Three ingredients did most of the work.

Sparse Mixture of Experts. MoE decouples model capacity from per-token compute. A 284 billion parameter model with 13 billion active per token reads roughly the same memory per decoded token as a 13 billion dense model, but holds far more knowledge in its weights. This is what made gpt-oss-120B and DeepSeek V4 Flash possible on consumer hardware.

Aggressive quantization as a normal practice. Q4 GGUF and MLX 4-bit became table stakes by mid-2024. The next step was mixed-precision schemes: IQ2_XXS on bulk routed experts combined with Q8 on attention and shared experts. This preserves quality much better than uniform low-bit quantization. The community now ships these by default, not as exotic experiments.

Reasoning-tuned small dense models. Qwen3.6 27B (Reasoning) at an Artificial Analysis Intelligence Index score of 37 is a dense 27 billion parameter model that comes within three points of the 284 billion parameter MoE released the same month. Better training data, better reinforcement learning recipes, and explicit chain-of-thought training pushed capability per parameter up sharply through 2025 and 2026.

Post-training on frozen architectures. The July 2026 refresh of DeepSeek V4 Flash is the purest demonstration: identical architecture, identical parameter count, same memory footprint, ten more index points. The gain came from reinforcement learning that suppressed hallucination rather than from any change you can see in a config file. For local users this is the best kind of progress — it costs no extra RAM.

What the next year might bring

Extrapolating at one doubling per 6.4 months puts the same 128 GB laptop past 100 well before mid-2027 — which is another way of saying the extrapolation breaks. The Artificial Analysis Intelligence Index is a 0-to-100 scale, the best model in the world sits at 61, and the best one that fits on a laptop is at 50. The honest reading is not "score 100 next year" but "the index gets rebased again, probably within twelve months, because the current one will have stopped discriminating at the top."

Writing this update is itself the evidence. v4.1 landed between the first version of this post and this one, and it moved almost every number downward — gpt-oss-120B from 33 to 24, Qwen3.6 27B from 46 to 37, Llama 3 70B from 10 to 3. The rebases are not noise around the trend; they are the benchmark repeatedly conceding that its old questions got too easy.

The harder constraint going forward is the 128 GB ceiling, and July 2026 is the first month it visibly bit. The top two open-weight models on the leaderboard, Kimi K3 at 57 and GLM-5.2 at 51, are both too large for any 128 GB machine at any quantization worth running. For two years "best open-weight model" and "best model that fits on my laptop" were the same sentence. They have now come apart. If Apple raises max unified memory in M6 Max, they may rejoin; if not, laptop users are riding a different, slower curve from here.

Caveats

The Artificial Analysis Intelligence Index was rebased three times in this window (v2 to v3 in early 2025, v3 to v4.0 in late 2025, v4.0 to v4.1 in mid-2026). Every score in this post is reconciled to v4.1, which incorporates nine evaluations: GDPval-AA v2, τ³-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience, and AA-LCR. Press release numbers, and the first version of this post, quote v4.0 or older values and are not comparable.

Llama 3 70B and Qwen 2.5 72B are both marked deprecated by Artificial Analysis, which means their v4.1 scores (3 and 10) come from a reduced benchmarking cadence rather than the full current harness. The May 2024 anchor is the weakest number in this post; see the second anchor in the Moore's Law section.

The May 2026 score of 40 for the April DeepSeek V4 Flash is derived, not read directly: Artificial Analysis has folded that model's page into the 0731 release and publishes only its non-reasoning score (29). The 40 comes from their own statement that 0731's 50 is "10 points above previous DeepSeek V4 Flash."

The Artificial Analysis Intelligence Index is run against full-precision hosted endpoints. Your local quantized model is typically 1 to 3 index points lower for Q4 dense. The 2-bit quantizations on DeepSeek V4 Flash have a larger hit, cushioned by keeping the sensitive layers at higher precision. The 50 quoted for DeepSeek V4 Flash 0731 is the hosted BF16 number at reasoning_effort: max; a UD-Q2_K_XL copy on your laptop will land below it, and nobody has published a careful measurement of how far below.

"Fits in 128 GB at usable quant" is shorthand. gpt-oss ships natively in MXFP4. The April DeepSeek V4 Flash uses IQ2_XXS plus Q8 in antirez's community GGUF. The July refresh is most easily run from Unsloth's dynamic quants. The shape of "fits" depends on quant tooling more than on a single bit-width number.

Parameter counts on the Hugging Face page and in the Artificial Analysis writeup differ: the repo lists 304B, Artificial Analysis says 284B total with 13B active. The gap is the attached speculative decoding (MTP) module, which most local setups load separately or skip.

reasoning_effort: max buys its accuracy with tokens. DeepSeek recommends allowing up to 384K output tokens at the high and max settings. On a laptop at 2-bit that is a long wait, and the index score assumes you let it finish.

Context length eats memory. The 5 tokens per second floor is met comfortably at moderate context. Past about 10K tokens decode falls 30 to 50 percent on M4 and M5 Max, and KV cache eats RAM headroom fast. The larger MoE models get marginal past 64K tokens.

Reproduce this yourself

Every model in this post is on Hugging Face. The Artificial Analysis numbers are at artificialanalysis.ai/models. The MacBook Pro specs are on apple.com. The antirez DeepSeek V4 Flash GGUF lives at huggingface.co/antirez/deepseek-v4-gguf, and the July refresh at huggingface.co/unsloth/DeepSeek-V4-Flash-0731-GGUF.

Pull them down. Run them locally. The numbers in this post will be a year out of date soon.