Ant's LingDT and Telecom's Xing4.0: Two Decision Models Launched This Week, Zero Independent Benchmarks

Verdict first: the "decision model" category went from curiosity to crowded in a single week — Jev, Microsoft Decision-1, and now two Chinese launches within 48 hours. The models are real; the benchmarks backing them are not yours to trust. For developers outside China, there is nothing to buy yet. Watch the category, ignore the press-release numbers.

On October 22, Ant Group's digital-tech arm quietly launched LingDT-Decision-1.0, a pair of small models (4B and 27B) built for one job: making fast, structured decisions instead of generating prose. Hours earlier the same day, China Telecom's Xingchen lab open-sourced Xing4.0-29B-Intent, an intent-understanding model in the same weight class — and unlike Ant, the weights are real and downloadable (more on that below, including a correction to what we first published). The two announcements together reached the English-speaking web as one news brief and a shrug.

What actually launched

The Ant story arrived as an authorized press release carried by QbitAI (the article footer says so explicitly), which shapes everything below. The claims, all self-reported by Ant: on 23 open benchmarks, the 27B version posts the top composite score among decision models, ahead of Microsoft Decision-1 and Jev 1.13.0; versus general-purpose LLMs in Ant's own customer-service and risk-control workloads, token consumption drops over 90% with roughly 10x faster response. The 4B and 27B models are live on Ant's DTMaaS platform with "limited-time free" access. No pricing sheet, no Hugging Face weights, no open-weights claim at all — my searches for LingDT on Hugging Face and GitHub returned nothing but noise. Single source, vendor-authored. Treat every number as marketing until weights or an API show up.

China Telecom's Xing4.0-29B-Intent got a stronger paper trail — and after publication we verified it end to end. The weights live at Hugging Face org XingChen-AGI (repo created Oct 15, Apache-2.0, 41 safetensors shards totaling 62.43 GB, mirrored on ModelScope), alongside a 2.75M-example training corpus and the 29B-A4B MoE base model it's built on. Full transparency: our first version of this article said the opposite. I searched HF for "Xing4.0", "xing4", and "TeleAI", got zero hits, and reported the open-source claim as unverifiable. That was a search-method failure — the org name is XingChen-AGI, which none of my queries matched. The correction matters because it changes the model's story: this is the only one of the week's launches with verifiable weights, a permissive license, and a config you can audit. The pitch is "multiple choice plus fill-in-the-blank" — intent classification plus parameter extraction in one pass, outputting strict JSON, averaging a few dozen tokens per response with sub-second latency. The config I pulled confirms a 256K context window (262,144 positions) and a 40-layer MoE routing 4 experts per token — the "29B total, ~4B active" framing from the release checks out against the file sizes.

Why the category is suddenly hot

The logic is genuinely appealing, which is why four players moved inside a week. Intent routing, tool selection, ticket triage, risk flags — these tasks need a judgment, not an essay. A specialized small model that answers in 30 tokens will beat a 400B generalist on latency and cost every time, and its output parses deterministically because the schema forces it. Jev (TypeSafe) proved demand; Microsoft followed with Decision-1 at $0.042 per million input tokens with output free — a price that only makes sense for a model that emits almost nothing. That pricing was verified against Microsoft's official page when we covered it; it remains the only public price anchor in this category. Ant and China Telecom both skipped pricing entirely: one hides behind "limited-time free," the other behind open-weights you can download but can't run anywhere convenient yet.

Now the skepticism, because it matters more than the specs. Both Chinese launches were press releases, not papers. Ant's headline benchmark — "beats Decision-1 and Jev across 23 benchmarks" — is a home game: the composite is 90% single-choice score plus 10% multi-choice, a weighting Ant chose, on a benchmark set Ant chose, measured by Ant. That's not a standardized comparison; it's a slide deck. Xing4.0's materials lean on deployment anecdotes (telecom call centers, smart speakers) rather than external rankings. If you've read this site before, you know the rule: never procure off a vendor's self-selected exam — the same reasoning we applied to Google's Harness² benchmarks yesterday.

What an English-speaking developer should actually do

Honest answer: nothing, this week. DTMaaS has no international sign-up flow I could find and the models aren't on OpenRouter; Xing4.0's weights are on Hugging Face and ModelScope, but at 62.43 GB of BF16 there is no official GGUF quant yet, so llama.cpp and consumer-VRAM setups are out for now — a 62 GB model wants roughly 64 GB+ of VRAM or aggressive offloading, which most intent-routing buyers don't have. The actionable takeaway is directional. If your stack routes intents, classifies tickets, or picks tools with a frontier model today, the category — not these specific models — is telling you that line item is about to get 10x cheaper. Microsoft's pricing is already public; when Ant publishes real numbers or a verifiable API, the arbitrage becomes concrete. My prediction, on the record: within a quarter, at least one of these Chinese decision models lands on an international aggregator with public pricing, and the press-release benchmarks will not survive contact with independent evaluation. If they do survive, I'll gladly eat this paragraph.

Dated caveat: correct as of October 23, 2026. Ant's LingDT claims rest on a single vendor-authored release (via QbitAI). Correction log: v1 of this article (Oct 22) reported Xing4.0-29B-Intent's open-source status as unverifiable after failed Hugging Face searches; v2 verified the weights at HF org XingChen-AGI (created Oct 15, Apache-2.0, 62.43 GB across 41 shards) via direct API pulls. Prices and availability will move.

Sources: QbitAI (Oct 22, Ant press release, authorized reprint); QbitAI + Zhidx (Oct 19–22, Xing4.0-29B-Intent); Hugging Face / hf-mirror API pulls, Oct 23 — XingChen-AGI/Xing4.0-29B-Intent (files, license, config.json) and XingChen-AGI/Xing4.0-29B-A4B (62.43 GB total); OpenRouter model index (no LingDT or Xing listings); our Microsoft Decision-1 pricing breakdown (Oct 11, price verified against Microsoft's official page).