By end of 2027, a model that runs well on 64GB Apple Silicon will match today's closed frontier on coding and daily-driver work.
confidence 0.45 · horizon 2027-12-31 · status open
Falsifiers
- Best local-runnable model still more than 15 SWE-bench Verified points behind the 2026-08 frontier on 2027-12-31.
- Open labs stop releasing sub-100B dense or small-MoE variants of frontier lines.
Evidence
- +1 Practitioner demonstrates a fully offline coding agent using open-weight Qwen3.5 9B — A practitioner demonstration of a fully offline coding-agent workflow using open-weight Qwen3.5 9B, with published scores ahead of the smaller Qwen3.5 4B variant, is a mild data point toward local-runnable models closing the gap with frontier agentic coding capability, though the cited benchmarks are lab-reported via Qwen's own model card rather than independently verified.
- -1 Moonshot launches Kimi K3.1 Pro, scoring near the closed frontier ahead of a promised Nov 4 open-weight release — At 2.8T total / 104B active parameters, Kimi K3.1 Pro extends the pattern of the open frontier moving to sizes far beyond any 64GB local-hardware deployment.
- +1 Meta partners with Apple to run its new open-source Muse Spark model locally — A reported Meta-Apple partnership to run an open model locally would raise the local-hardware ceiling relevant to this thesis, though it is single-sourced via an opinionated outlet with no primary confirmation yet.
- +1 Qualcomm launches new smartphone chips that run a 30B MoE model locally — A new phone chip reportedly running a 30B MoE model locally raises the on-device capability ceiling, a data point toward local-runnable models closing the gap with frontier, though it is mobile silicon rather than the owner's 64GB Mac scenario and unverified independently.
- -1 DeepSeek ships V4.1-Flash with a novel prefill/decode-split MoE architecture, retiring old Flash naming — At 763B total parameters, V4.1-Flash's efficiency gains don't bring it within reach of 64GB local hardware, extending the pattern of the open frontier moving to sizes that won't run locally.
- +1 Mystery model 'Ox Alpha' confirmed as GLM-5.3-Flash; independent quantization and serving benchmarks emerge — 4-bit quantization reportedly retains 93% accuracy and fits a 256GB Mac or two DGX Sparks, showing frontier-adjacent open models moving toward consumer/prosumer-hardware feasibility, though above the owner's 64GB threshold.
- +1 Apple debuts new Mac Mini and Studio with AI-dedicated chips — New AI-dedicated Apple silicon in the Mac Mini/Studio refresh could raise the local-hardware ceiling relevant to the 64GB local-fast/local-heavy verdicts, though no benchmarks are available yet.
- -1 DeepSeek releases V4 open weights under MIT license — The open frontier is moving to huge MoE sizes that will never fit locally.
- +1 Qwen releases open agent runtime tuned for its coder models — Scaffold-level gains accrue to local models too.
Confidence history
- 2026-05-01 → 0.4 — Initial
- 2026-08-05 → 0.45 — Qwen agent runtime closes part of the scaffold gap