There is minimal downside to switching to open models

Hardware and Local Inference Economics

• Several commenters run open models locally (often on high-RAM Apple Silicon), finding them “fully doable” for coding only with 64–96GB+ RAM; 16GB is generally seen as insufficient.
• Apple’s unified memory is viewed as a cost‑effective alternative to high‑end GPUs, which are described as extremely expensive and power‑hungry.
• Others run the numbers and conclude local hosting of large models (e.g., GLM 5.2) is far more expensive and slower than cloud inference unless you have sustained, high utilization or strict privacy needs.
• Cooperative or shared local clusters are floated but mostly dismissed as uneconomical compared to cloud providers that benefit from scale and investor subsidies.

Quality Gap: Open vs Frontier Models

• Many say open models are “a few months behind” top proprietary models, good enough for boilerplate, refactoring, and simpler coding tasks.
• Others strongly disagree, reporting that even the best open-weight models still do not match recent frontier models for complex work, especially large software engineering tasks.
• Some highlight standout open options (e.g., GLM 5.2, DeepSeek V4, MiMo variants) as close to prior frontier releases, particularly for coding, but note instability and jagged behavior.

Pricing, Subscriptions, and Token Economics

• Hosted open-weight providers (OpenRouter, OpenCode Go, various EU routers) are compared; pricing varies widely, with some flat‑fee plans offering much more usage than equivalent frontier subscriptions.
• Debate over subscriptions vs pure per‑token billing: heavy users may prefer subscriptions; light/power users report very low API spend due to careful prompt and tool design.

Reliability, Degradation, and Vendor Lock‑in

• Multiple commenters report perceived “degradation” of proprietary models over time (possibly via quantization, routing to smaller variants, or capacity constraints), and note tracked benchmark dips.
• This drives interest in open weights as a hedge against enshittification and rug‑pulls: once weights are public, anyone can host them and you can keep using a favored version indefinitely.

Privacy, Jurisdiction, and Ethics

• Strong concerns about sending sensitive or EU‑regulated data to US companies; some insist on EU‑owned/hosted routers with strict data‑retention guarantees, even at higher cost.
• Others distrust all major US providers due to alleged involvement in military targeting and broader surveillance concerns.
• Counter‑arguments note that even EU states want surveillance; some argue true safety requires full self‑hosting.

Use Cases, Tools, and Harnesses

• Tool calling (web search, HTTP fetch, etc.) is emphasized as a harness feature, not an intrinsic model capability; open models can match frontier models here if wrapped correctly.
• Several report that for 80–95% of “grunt” or junior‑level work, smaller or cheaper open models are sufficient; frontier models are reserved for design‑heavy or high‑stakes tasks.
• Some stress that switching models is easy, but keeping complex harnesses and prompt stacks working identically is non‑trivial.

Critiques of the Article and Future Outlook

• Multiple readers note the article headline (“minimal downside”) isn’t backed by concrete evidence; the text is seen as more aspirational than proven.
• There is disagreement on trajectory: some expect open models to reach or exceed recent frontier quality within 1–3 years; others argue compute and data constraints will keep them behind without major breakthroughs or state‑level backing.
• Regulatory risk is raised: open models might be restricted or banned under security pretexts, which could entrench large proprietary vendors.