The Vibe Father
Analysis
"Should I run my coding model locally or in the cloud?" is a question that usually gets answered with vibes — privacy people say local, speed people say cloud, and nobody does the math. This post does the math, or at least the honest version of it. The short answer: cloud flagships still win raw capability by a clear margin, local wins on privacy and on fixed cost at scale, and the right choice depends on which of those you are actually optimizing for. Let us take the axes one at a time.
Capability: cloud still wins the ceiling
This is the axis people most want to be a tie, and it is not one yet. The best models you can run locally are open-weight, and the best open-weight SWE-bench Verified score on our board at /benchmarks is GLM 5.2's 78.7, with DeepSeek V4 Pro (77.6) and Qwen3.7 Max (77.3) just behind. Those are genuinely strong. But the cloud-only closed flagships sit above them: Claude Opus 4.8 at 88.6 and Claude Fable 5 at 95.0. That is a ten-to-sixteen-point gap on the benchmark closest to real repository work, and on hard multi-file tasks you feel it. If your work lives at the difficult tail, the capability ceiling alone points at cloud.
Latency: it is not the layup you think
The intuition is "local has no network round-trip, so it is faster." Sometimes. But throughput on a self-hosted open model is bounded by your hardware, and a cloud provider's optimized serving stack often streams faster than a box you provisioned yourself. On our board, cloud-served models like Qwen3.7 Max hit 204 tokens per second and Gemini 3.5 Flash 167 — numbers most self-hosted setups will not match without serious GPUs. Local can win first-token latency and removes network variance, which matters for tight interactive loops; it does not automatically win sustained throughput. Measure your actual setup before assuming.
Privacy: local wins, cleanly
This is the axis where local is unambiguously ahead. Run the model inside your own perimeter and no code, no prompts, and no proprietary context ever leave the building — no vendor sees your repository, and there is no data-processing agreement to negotiate or trust. For teams under a compliance regime, or working on code that contractually cannot touch a third party, this is not a preference, it is a requirement, and it is the single most common reason we see teams self-host. Cloud vendors offer strong contractual protections, but "the data never left" is a categorically stronger guarantee than "the data left and was handled correctly."
Total cost: the calculation people skip
Cloud pricing is per-token and visible: DeepSeek V4 Pro is $0.435 in and $0.87 out per million; MiniMax M3 is $0.30/$1.20; Opus 4.8 is $5/$25. Local pricing looks like "free after hardware," which is the trap. The real local cost is hardware amortization plus power plus the operations burden of keeping an inference stack healthy — and that ops time is the line item people forget.
| Factor | Local (self-hosted open) | Cloud (hosted API) |
|---|---|---|
| Capability ceiling | strong (open ~78 SWE) | highest (up to 95.0 SWE) |
| Sustained throughput | bounded by your GPUs | up to 200+ tok/s |
| Privacy | data never leaves | contractual, not physical |
| Cost shape | fixed (hardware + ops) | variable (per token) |
| Scales best when | volume is very high | volume is low or spiky |
The crossover is about shape, not just size. Per-token cloud pricing scales linearly with usage — cheap when you are light or spiky, expensive when you run agents flat-out all day. Fixed local cost is the reverse: painful to stand up, but the marginal token is nearly free, so it only pays off once your volume is high enough to amortize the hardware and the ops time. Below that line, cloud is cheaper and far less hassle. Above it, local's fixed cost wins. Most individuals and small teams are below the line; a few high-volume shops are decisively above it.
The honest recommendation
For most people, most of the time, cloud is the rational default: highest capability, no ops burden, and per-token pricing that is genuinely cheap for open models served by competing providers. Reach for local when a specific requirement forces it — data that cannot leave your perimeter, a compliance regime, or volume high enough that fixed cost beats per-token — and accept the capability ceiling that comes with it. The mistake in both directions is treating this as identity: cloud people overpay for privacy they do not need, local people ship on a weaker model to save money the math does not support.
The better framing is that you do not have to choose globally. Route privacy-bound and high-volume work to a local open model, send the hard tail to a cloud flagship, and keep both available. That is exactly how The Vibe Father is built — local and cloud models side by side, each job going to the right one, on your own keys. Whether self-hosting actually pays for your workload is its own calculation, and we work it through in detail in our self-hosting guide.
Bottom line
Local versus cloud is four separate trade-offs wearing one question. Cloud owns the capability ceiling and hassle-free throughput; local owns physical privacy and fixed cost once volume is high. Do the total-cost math including ops time, be honest about the SWE-bench gap, and route by requirement rather than by tribe. Live scores stay current at /benchmarks.
