Announcement to Weights: The 2026 Open-Weight Gap Ledger

The announcement-to-weights gap is the most revealing number in open-weight AI that nobody tracks. Every 2026 model launch produces two dates — the day the vendor said “open weights” and the day a checkpoint actually appeared on Hugging Face — and the distance between them varies from zero to open-ended, sometimes inside the same lab’s own release history.

The spread is the story. Moonshot AI shipped Kimi K2.5’s weights the same day it announced them in January, then ran a ten-to-eleven-day gap on Kimi K3 in July. Z.ai put GLM-5.1 on Hugging Face the day it was announced in April, then announced GLM-5.3 on August 14 with no weights at all — and as of this ledger’s August 23 dateline, they still have not shipped. NVIDIA announced a 550B-parameter model from a keynote stage and closed the gap in three days.

This ledger lines up every notable 2026 open-weight release — announcement date, weights-live date, and the gap in days — with a confidence flag on every row. Where two credible sources disagree on a date, the row shows a range or “not established” instead of a number we cannot defend. That restraint is deliberate: a ledger you can cite is worth more than a ledger that looks complete.

The ledger measures exactly one thing: the number of days between a vendor’s own public announcement of an open-weight model and the first date the weights were verifiably available on Hugging Face or an equivalent public repository. The “announced” column comes from the vendor’s blog post, keynote, or launch coverage dated to the announcement itself. The “weights live” column requires either a model-card commit date or at least one independent press report explicitly stating the weights were downloadable.

That two-source discipline matters because the press routinely conflates the two events. Coverage of Kimi K3 described it as “open-weight” from July 17 — when what existed on July 17 was a licence commitment, not a download link. The weights arrived roughly ten days later. A ledger that took launch-day language at face value would erase the very gap it exists to measure.

Three honesty rules govern every row. First, where two credible sources give dates a day or two apart, the row shows the range — Kimi K3 reads “10–11 days”, not a single integer. Second, where the weights-live date could not be confirmed at all, the gap column reads “not established” rather than an estimate. Third, licence terms are a secondary column only — our licence audit of 30 open-weight models owns that story, reading each licence text directly; this ledger just records the tag and links out.

One methodological caveat cuts the other way: sometimes weights are effectively public before the announcement. Meituan’s LongCat-2.0 circulated on OpenRouter as the anonymous “Owl Alpha” listing before the company’s formal reveal — so for some users, first access predated the announcement date entirely. The ledger measures the vendor’s public commitment against the vendor’s public delivery; stealth aliases are noted, not counted.

The table below is the primary asset. Rows are grouped by outcome — same-day ships, multi-day gaps, and rows that are still open — and each carries a confidence flag. The licence column records the tag only; for what those licences actually permit, see the full licence audit. Dates are ISO-format, 2026 unless noted.

Two reading notes. The “see repo” licence cells mark rows where this pass recorded the release but not the licence tag — the licence audit is the authority there. And the July rows are ledger entries only: the fuller story of that month’s release cluster lives in July’s open-weight momentum tracker, which this page deliberately does not retell.

Seven rows ship at gap zero, and they span the whole field: a 1T-parameter multimodal model (Kimi K2.5, January 27), a 754B agentic MoE (GLM-5.1, April 8), Google’s Gemma 4 family (April 2), both of Tencent’s Hy3 drops (the region-restricted April preview and the Apache 2.0 full release on July 6), DeepSeek’s V4 Preview pair (April 24), and NVIDIA’s Nemotron 3.5 Lightning (August 11), which hit Hugging Face, ModelScope, and OpenRouter the day it was announced.

Same-day shipping is not a coincidence of logistics — it is an operational stance. A vendor that uploads weights before the announcement is treating the download link as part of the announcement, which means every launch-day claim can be checked within hours. DeepSeek went further still, shipping V4-Pro and V4-Flash simultaneously across weights, API, and its chat surface — the pattern we documented in our V4 Preview launch coverage.

Two rows sit near this cohort but do not earn a hard zero. DiffusionGemma is widely reported as a June 10 same-day drop with day-zero inference support across vLLM, Transformers, MLX, and SGLang — but this pass could not verify the date against Google’s own blog, so the row reads “reported ~0; unverified”. GLM-5 and GLM-5.2 both appear to follow Z.ai’s same-day pattern, but neither February nor June produced a single authoritative day this pass could pin down, so both gaps stay “not established” rather than being forced to zero.

The multi-day cohort splits into two very different patterns. The short gaps are event-calendar artifacts: NVIDIA announced Nemotron 3 Ultra (550B total / 55B active) from the Computex keynote stage on June 1 and had weights live on June 4 — a three-day gap where the announcement date was fixed by a keynote schedule, not by upload readiness. Meituan’s LongCat-2.0 (1.6T total / ~48B active) ran four days from its June 30 announcement to a July 4 Hugging Face upload.

The long gaps are staged releases. Moonshot unveiled Kimi K3 — 2.8T total parameters, 16 of 896 experts active per token, roughly a 50B-parameter active compute cost — at WAIC on July 16, and the weight dump landed ten to eleven days later. Launch coverage at the time reported K3 placing 4th on the Artificial Analysis Intelligence Index and 1st on Arena.ai’s Frontend Code Arena — reported-at-announcement figures, noted here only as colour. The same July window also saw the White House publicly accuse Moonshot of distilling Anthropic’s Fable model and Treasury Secretary Scott Bessent threaten sanctions — a one-line reminder that weight-release timing on Chinese models now carries political freight beyond logistics.

Alibaba ran the most elaborate staging of the year. It previewed “Qwen3.8” on July 19 — API-only, no benchmark table, no repo, no licence file — with the release language “Qwen3.8 is launching and going open-weight soon.” The full Qwen3.8-Max announcement followed on August 3, and the 2.4T-parameter open checkpoint went live on Hugging Face around August 12–13: roughly 9–10 days after the full announcement, 24–25 days after the preview. A smaller Qwen3.8-27B under Apache 2.0 followed around August 13–14.

The staging carried a substantive catch: the open checkpoint is narrower than the hosted product — text-only, no vision, and without the API version’s native 1M-token context. That is exactly the kind of announcement-versus-artifact drift this ledger exists to surface: the thing announced on August 3 and the thing downloadable on August 13 are not the same thing.

One row belongs in this cohort but cannot be plotted. MiniMax announced its M3 flagship around June 1; one source states weights were live on Hugging Face by June 7, while another framed a similar-length window as merely expected. Those two claims were never independently reconciled, so M3’s gap stays “not established” — and its 59.0% SWE-Bench Pro score is a vendor-stated figure, for the record, not an independent benchmark.

The most instructive rows in the ledger have no gap number at all — and they got there by three entirely different mechanisms. One is a deliberate, safety-cited hold. One is a tease that never produced a public artifact. And one is a release that was never formally announced in the first place.

The GLM-5.3 row is the cleanest anchor for the whole ledger’s thesis, precisely because of Z.ai’s own history: GLM-5.1 shipped weights same-day in April, and GLM-5 and GLM-5.2 both appear to follow the same pattern. GLM-5.3 broke it — the announcement came with API access and a stated intent to publish weights after safety evaluation, and as of this dateline the row still reads “pending”. Same lab, same model family, a completely different release posture. When the weights do land, this row gets its number; until then, the honest entry is the open one.

The Llama 5 row deserves its own footnote on reporting hygiene. The neutral fact is simple: no Llama 5 weights have ever shipped and no release date has ever been announced, a trajectory we traced in our H1 2026 open-weight retrospective. Yet the reporting around it is where this space gets murky.

Read across all twenty rows, the pattern is too consistent to be noise. Labs that treat open weights as the product — DeepSeek, early-2026 Z.ai, Google’s Gemma line, Tencent by July — ship at gap zero, because the upload is finished before the press release goes out. Labs that announce from an event stage run short, calendar-driven gaps: NVIDIA’s three days from a Computex keynote is the disciplined version of that pattern. And labs that stage releases — announcing first, uploading later — run gaps of nine days and up, during which the announcement does marketing work that no downloadable artifact can yet verify.

Looking forward, we expect the GLM-5.3 pattern to become more common, not less. As open models post stronger agentic and cybersecurity results, deliberate safety-review holds between announcement and weights are a rational vendor response — which means the gap will increasingly encode a lab’s governance posture rather than its upload bandwidth. That makes this ledger’s axis more informative over time: a widening gap on a vendor’s next release is signal, and an announcement with no dated weights commitment is a row you should treat as open until proven closed.

For teams, the operational translation is straightforward: gate any self-hosting or fine-tuning decision on the weights-live date, not the announcement date — and on the shipped checkpoint’s actual capabilities, not the launch post’s. Our guide to which open-weight coding models you can actually self-host picks up exactly where this ledger stops. And if you are deciding whether an open-weight model belongs in your production stack at all, that evaluation is the first step of our AI transformation engagements.