This is the working register behind “The open-weight map, mid-2026.” The essay makes a buying argument, and a buying argument that cannot show its sources is marketing. What follows is the audit trail: how sources were graded, the five contradictions that had to be resolved before anything could be charted, and the claims that were cut.
How sources were graded
Three tiers, applied before drafting. First-party and rigorous: model cards, vendor announcements, published research, and measurement organisations reporting their own methodology — Epoch, Stanford HAI, OpenRouter, Artificial Analysis, Hugging Face. Named analyst: identified authors publishing under their own name with visible method. Corroboration only: unattributed vendor blogs and SEO content.
The operative rule was that a third-tier source never carries a number on its own and never reaches a chart. Several widely circulated specifications for recent model releases were reachable only at that tier, and the essay either omits them or states the family-level claim without the specific figure. Where the register says a claim rests on thin ground, the piece says so in the prose too.
Three market-share numbers, three denominators
The single most misused fact in this subject area is open weights' market share, because at least three defensible numbers are in circulation and they disagree violently. They disagree because they are measuring different populations.
OpenRouter's routing data shows open weights crossing from minority to majority of routed tokens across roughly eighteen months, with 72.4 percent of the tokens going to the platform's top ten models. Menlo Ventures' enterprise survey shows open models' share of enterprise LLM usage falling from 19 percent in 2024 to 11 percent. AI.cc's infrastructure report puts open weights at 38 percent of API tokens in the first quarter of 2026.
Developer routing, enterprise workload share, and API traffic are three different denominators over three different populations. All three measurements appear to be honest. Citing any one of them as “the market” produces a confident and useless story, which is why the essay carries all three and spends a section on what sits in the gap. The resolution the evidence supports is production conversion: open models reach production at 53 percent against closed at 63 percent, and the closed figure climbs with company size while the open figure stays flat.
Five contradictions, resolved
Five conflicts in the source set were material enough to gate the charts. All five resolved; none required omitting a finding.
A capability index that changed methodology mid-year. Two scores circulate for the same NVIDIA model, roughly 48 and 38. Both were correct at the time of writing: the higher figure is from the index's earlier version, the lower from the current version, which shifted weighting toward agentic workloads. The current figure is used throughout, and any use of the older one requires the version and date attached. This is the general hazard with capability leaderboards — the index is a moving instrument, and a year-on-year comparison across a methodology change is not a comparison.
A share-of-tokens figure with two denominators. Chinese-origin models' share of routed tokens appears as roughly 73 percent and as roughly 46 percent. The first is the share of open-weight tokens only, as of January 2026; the second is the share of all routed tokens at mid-2026. Plotting both on one trend line would depict a collapse that did not occur. No open-only figure exists for mid-2026, so the earlier number is cited with its date attached and no trend is drawn between them.
A load-bearing enterprise number that is seven months stale. The 11 percent enterprise-usage figure comes from a report dated 31 December 2025, based on 150 technical leaders, with no 2026 refresh. It is labelled with its date everywhere it appears. The staleness is itself reported in the piece rather than hidden, because a widely cited number with no refresh is a fact about the evidence base that buyers should know.
A coding benchmark its own publisher has retired. OpenAI published in February 2026 that SWE-bench Verified no longer measures frontier coding capability, having found that frontier models reproduce gold patches from the task identifier alone — pre-training contamination. A separate finding in the same work is that a majority of one frequently-failed subset had flawed test cases rejecting correct patches. These are distinct problems and the register keeps them distinct; merging them into a single contamination percentage is a misreading. The consequence is that the strongest open coding score in the pack is hedged with this context wherever it appears.
Frontier leaderboard standings that depend on reasoning effort. Top-of-table scores shift by several points depending on the reasoning-effort setting a model is evaluated at. The essay fixes on one index version at maximum reasoning and stays inside it, and the gap it reports between the strongest open model and the closed frontier should be read with that setting attached.
The number the whole sovereignty argument rests on
One finding does more work than any other in the closing section: that 56 percent of sovereign large language models are adapted models built on existing open bases. If that is right, national AI programmes are not an alternative to the open-weight ecosystem — they are its largest downstream consumer, and sovereignty at the model layer is a fine-tuning exercise on someone else's pre-training run. The essay leans on this and therefore names its source in the prose rather than burying it in a citation list.
What was cut
A claim that open models trail closed ones by 25 to 40 percent on coding benchmarks was cut as irreconcilable with a first-party score more than twice that high; the discrepancy traces to stale figures and different harnesses, and no honest reconciliation was available. Detailed specifications for several 2026 model releases were cut where they existed only at the corroboration tier — the families are discussed, the unverifiable parameter counts are not. Announced-but-unshipped models were excluded entirely; a release target is not a release. And US proposals to restrict open-weight distribution are described as proposals, because that is what they are — the essay states this precisely rather than implying law where there is none.
One gap is worth declaring rather than papering over. A US government commission report on China's open-weight strategy is the highest-value primary source identified for the geopolitical section and was not read in full before publication. The section's claims are sourced elsewhere and stand without it, but a reader who wants the strongest version of that argument should go to it directly.