ModelBrief · Hy4 · Topic Page
Last updated Nov 20, 2026

OrcaRouter Checks Hy4 Flash: 1/6 KV Cache, No Score Tables

Published Nov 16, 2026 · Last updated Nov 18, 2026

The launch announcement says "95% cheaper". OrcaRouter's November 16 check asks a different question: how that number is calculated, and what is still missing from the release. The answer is two-sided — the efficiency gains are real, but the evidence chain is incomplete.

The efficiency bill: one-sixth the KV cache, over 150 tok/s decoding

The check confirms two efficiency figures beyond the announcement: Hy4 Flash's KV cache is 1/24 of Hy3's and 1/6 of Hy4's, and decoding runs above 150 tokens per second (megakernel, 4-bit quantisation, DFly inference). A smaller cache means more concurrent sessions on the same hardware — one of the core reasons Flash can be cheap. One pricing detail needs separating: TokenHub's 50% discount has no stated end date, which is different from WorkBuddy being free until November 30; do not merge the two offers into one.

"18x" is task cost, not price

The "18x" in the announcement is the easiest figure to misread. OrcaRouter notes it is a task-cost figure, not a price ratio: the pay-as-you-go price gap itself is 12x on input and 9x on output. Task cost falls further because Flash uses fewer tokens per task — price and usage stack to produce the large drop. The "close to Haiku 5.5" wording also has two versions: the official account says "matches", the research page says "close to"; the 85-person blind test scored 2.79 against Haiku 5.5's 2.81, which is close, not tied.

Product-side figures: Tencent's own claims

The check also collects the product-side figures from the announcement, all Tencent's own claims: WorkBuddy task time down 70% and cost about 1/15 of Hy4; Yuanbao general chat up 2.52% (against GLM-5.3-Flash); WeChat Pay agent flows down 29% in output tokens and 6% in steps; WeCom output 88% faster with 60% fewer thinking tokens; Honor of Kings root-cause analysis requests down 22–39%. None of these has independent confirmation; cite them with the "Tencent says" qualifier intact.

What is missing: a checklist

The most useful part of the check is the missing list: no model card, no architecture paper, no license, no weight repository, no independent benchmark score. The seven benchmark figures (WorkIndex 42.1, Terminal-Bench 4.0 19.2%, OSWorld 2.1 68.1% and the rest) exist only as PNG images on the official site, with no machine-readable table; Artificial Analysis has not scored Hy4 Flash. On launch day, the only checkable evidence is the vendor's own images — not even the raw data behind the scores. iFanr's same-day independent test is the first third-party number, but it is a single sample inside WorkBuddy, not a benchmark reproduction.

Update, Nov 18, 2026: the model card, license and weights that were missing at launch have now arrived with the Hy4 Flash open-weights release: three variants at once, under Tencent Hy License 1.0 (not Apache 2.0), with the standard weights at 440.5 GB. See our open-weights article. The "no model card, no license, no weights at launch" wording above is the launch-day picture and stands as written; what is still missing is an architecture paper and independent benchmark scores.

Our take: The efficiency story is credible — one-sixth the KV cache, decoding above 150 tok/s and a much lower task cost all have concrete mechanisms behind them. But headline figures like "95% cheaper" and "close to Haiku 5.5" should be read in their task-cost and blind-test senses. Until a model card, weights and independent scores appear, the model's true capability tier is unverified; everyday high-frequency work can follow iFanr's test and be tried with confidence, while selection for critical work should wait for independent scores.

Reference source: this article is an original rewrite produced by this site; facts and figures were checked against OrcaRouter — Tencent Hy4 Flash Launch (Nov 16, 2026)

← Back to the Hy4 topic page