Open-Weight Frontier Parity

The supply-side half of the specialization thesis. If capability is moving from general-and-central to specific-and-local, the precondition is that a model you can actually hold reaches usable frontier capability — and 2026 is the year that stopped being hypothetical.

DeepSeek V4 (open-weight, MIT licence) ships V4-Pro at 1.6T parameters / 49B active with 1M context, reporting SWE-bench Verified ~80.6% and GPQA ~90–92% at ~$0.435 / $0.87 per Mtok — roughly one sixth the cost of Opus 4.7 or GPT-5.5, and it landed inside an eight-day April-2026 frontier window. Kimi K3 (Moonshot) went further on scale: 2.8T parameters, 896 experts with 16 active per token, 1M context, native vision — the largest open-weight model to date, topping the Arena.ai frontend-code leaderboard.

The part that complicates the thesis

Open weights are not automatically cheap weights. Kimi K3's API launched at $3/$15 per Mtok — Sonnet-tier pricing, the most expensive Chinese-lab release to date. Open-weight and low-cost came apart in the same quarter. What you get from open weights is optionality — the right to run it yourself, quantize it, tune it, and refuse a price change — not a lower sticker price from the vendor.

That distinction is the thesis's actual mechanism. The reason to care about open weights is lock-in, not headline price: a model whose weights you hold cannot be repriced, deprecated, or silently swapped underneath you.

Against the price trend

The Tiered Super-Moore analysis (3,237+ models, 2020–2026) finds a ~600× price decline with half-lives of 1.10 years at the economy tier and 1.55 at mid-tier, against a 2-year Moore benchmark — while flagship and reasoning tiers resist, holding a ~31.5× premium. Open-weight releases attack the premium tier specifically: DeepSeek V4 is a frontier-tier capability claim at economy-tier pricing.

What this concept still needs

This shelf now has two concepts (updated 2026-09-24). Until then the specialization cluster as scoped — fine-tuning, LoRA/PEFT, personalisation, distillation-for-edge, on-device inference — had no sources at all. It now has one: ALTO, a paper on the cost of LoRA fine-tuning searches, compiled on LoRA fine-tuning cost. That covers the entry fee of specialising a model; nothing yet measures a specialised model against a general one on a task, which is where our ran it seat would actually bear.

That is the clearest ingest target in the KB, and it matters more than the count: Draft Day 2026-08-08 ratifies or rejects the specialization thesis, and the evidence base for the brand's central claim was one page about model releases until 2026-09-24; it is now two, and neither measures a specialised model doing a task.

Key Claims

• DeepSeek V4-Pro: ~80.6% SWE-bench Verified, ~90–92% GPQA, at ~$0.435/$0.87 per Mtok — ~1/6th frontier cost. Evidence: weak (news-tier report of vendor figures) (DeepSeek V4)
• Kimi K3: 2.8T params, 896 experts / 16 active, largest open-weight model to date — but priced at $3/$15 per Mtok. Open weights ≠ cheap tokens. Evidence: weak (analysis) (Kimi K3)
• ~600× token-price decline 2020–2026; economy-tier half-life 1.10yr vs Moore's 2yr; flagship tier resists at a ~31.5× premium. Evidence: strong (systematic empirical analysis, 3,237+ models) (Tiered Super-Moore)