Ant Group's InclusionAI division has put the weights for Ling-3.0-tiny on Hugging Face: a 7.9B-parameter mixture-of-experts model that activates just 1.3B parameters per token, with a 256K context window, switchable Thinking/Instant modes, and BF16, FP8, and INT4 checkpoints for everything from datacenter to laptop. The API listing appeared quietly on August 6 — free on OpenRouter and Vercel — and the weights followed about five days later. If that release choreography sounds familiar, it should: Ling-3.0-flash did exactly the same thing in July, API on the 23rd, MIT-licensed weights on the 26th. This is now a pattern, and it's worth naming.
The architecture: Kimi's attention, Ant's model
The technically interesting part of the whole Ling 3.0 generation is the attention stack. Tiny interleaves KDA — Kimi Delta Attention, Moonshot's linear-attention design — with MLA (multi-head latent attention) in a 3:1 ratio: three linear-attention layers for every one full-attention layer, per four-layer block. Flash, the 124B sibling, uses the same recipe at 5:1. That hybrid is what makes a 256K context plausible on a model this small: linear attention keeps the long-context memory cost from exploding, the periodic MLA layers keep retrieval quality from collapsing — the same wager Mamba hybrids have been making, with the same known failure mode to watch for on precise long-range recall.
Note what's happening between labs here: one Chinese lab's attention innovation is now load-bearing infrastructure in another Chinese lab's model line, within months of publication. KDA is a Moonshot (Kimi) design; Ant adopted it across the entire Ling 3.0 family. That's the open-ecosystem flywheel the trends post keeps pointing at — Asia's 56% open-weights share isn't just a release statistic, it's the reason architecture ideas compound across labs there at this speed.
The sparsity is the other headline: 128 routed experts, 8 routed plus 1 shared active per token, for a 6:1 total-to-active ratio at a size where MoE barely existed a year ago. A 1.3B-active model with INT4 checkpoints is aimed at genuinely modest hardware — this competes head-on with the LFM2.5 and TwiL-LM local-agent tier, not with anything on a leaderboard.
And unusually, the card prices that claim in hardware terms rather than vibes: validated on NVIDIA's DGX Spark, Apple-silicon MacBooks, and a Mac mini, with FP8 throughput of 100–105 tokens/s on DGX Spark and 86–90 tokens/s on an M4 Pro MacBook, at roughly 8.34 GiB peak memory for an 8K context. Artificial Analysis independently measures 160+ tokens/s served, with ~18 seconds end-to-end for a 500-token response including thinking time. Those are the numbers a local-agent buyer actually needs — tokens per second on the machine they own, and whether the thinking mode's latency tax fits an interactive loop — and most cards in this tier don't publish them.
What we can't say yet
Two honest gaps. First, capability was unpriced at launch: through the API-only week the only published numbers were composite scores — Artificial Analysis Intelligence Index 25, Agentic Index 16 — with no per-benchmark table behind them. That gap has since closed: the Hugging Face card now carries a full 15-benchmark table, decoded in the update below. Second, the license on tiny's weights needs reading: flash shipped MIT, prior Ling releases were MIT, and there's no signal Ant is changing course — but "the sibling was MIT" is an expectation, not a license. Given how much of this month's coverage has turned on exactly what an "open" release lets you do commercially, the license file matters more than the announcement.
Worth flagging the reporting lag too: most of the secondary coverage this week still says tiny's weights are "promised but not delivered" — the Hugging Face repo going live appears to have outrun the reviews. The primary source beats the aggregators by days, again.
Update (August 11): the card's own table, decoded
The model card now publishes a 15-benchmark table, all models in Thinking mode, against Qwen3.5-4B, Qwen3.5-9B, Gemma-4-E4B-it, and Gemma-4-12B-it. Read as a scoreboard, tiny loses: it leads on 4 of 15, Qwen3.5-9B leads on 4, and Gemma-4-12B leads on 6. Ant published a table its model doesn't top — which, as with Meta's Glimmer card, is exactly why the table is worth reading. The profile it draws is coherent:
- The wins are the agentic ones — and they're not close. GDPval v2-AA at 772 versus 645 for the next best (Gemma-4-12B), and TAU3-Banking-AA at 20.8 versus 8.7 — 2.4× the runner-up, though note the low absolute numbers: every model in the row fails most of that benchmark. Add BFCL-v4 function-calling (62.72) and Terminal-Bench 2.1 (27.7), both second only to a Qwen with 7× the active parameters, and the "executor inside an agent loop" positioning stops being marketing language and starts being what the table actually shows.
- Knowledge is the sacrifice, and it's honest about it. AA-Omniscience accuracy of 8.52 is dead last — a 1.3B-active model simply cannot store what a 12B dense one can. But its non-hallucination rate is 69.54, against 13–19 for the Qwens and Gemma-12B: tiny knows very little and says so, where its rivals know more and confabulate the rest. For a chatbot that trade reads badly; for an executor whose harness will supply facts through tools, calibrated ignorance is arguably the more useful spec.
- The long-context caveat materialized on cue. AA-LCR comes in at 58.7 — behind both Qwens and Gemma-12B — despite the 256K context claim. That's the linear-attention recall weakness we flagged above showing up in the vendor's own table: the KDA-heavy stack buys the context window cheaply, and pays for it in retrieval quality across it.
- Math splits oddly: a genuine win on IMO-AnswerBench (71.03) next to third place on HMMT — consistent with contamination-sensitive variance between competition sets more than with any clean "good at math" story.
Usual caveats, sharpened: the comparison set is the vendor's, and the framing flatters — pitting 1.3B active parameters against 4–12B dense rivals makes every near-tie a win on efficiency grounds, which is fair as far as it goes but is also precisely the axis Ant chose to fight on.
One methodological footnote on the card deserves credit, though: the Terminal-Bench 2.1 score comes with its full harness disclosed — the default Terminus 2 harness under the Artificial Analysis protocol, a 2-hour timeout, preserve-thinking JSON parsing, 3 runs averaged, 32K generation budget in a 256K window. After a month of harness swings moving agentic scores more than model generations do, a vendor stating which harness produced the number is exactly the disclosure agentic benchmarks have been missing — a small thing that should be table stakes, and mostly isn't.
Why the choreography matters
API-first, weights-later is becoming the default release pattern for Chinese labs, and it's a rational one: the free API week generates usage data and buzz while the weights are prepared, and the weights release then converts attention into ecosystem adoption. But it also means the announcement date and the open date are diverging — flash was "released" July 23 and open July 26; tiny was "released" August 6 and open around the 11th. For anyone tracking open-weights share over time (as our dataset does), the choreography is a measurement hazard: a model can spend its whole news cycle closed and still end up an open release. We've dated tiny to its August 6 launch and corrected flash's entry, which our arcade had honestly marked "weights not yet confirmed" at launch — they since arrived.
The strategic read is the same one the trends post's bear case keeps circling: the sub-10B agent-executor tier now has entries from Liquid (revenue-gated), webAI (non-commercial), NVIDIA (open-with-conditions, as of this morning), and Ant (MIT, if the pattern holds). Four labs, four license philosophies, one converging product: the cheap local executor inside someone's agent loop. When capability differences are this hard to measure — composite indices, vendor tables, no common harness — the license may be the most legible spec on the sheet.
What to expect next
- Watch the license file on the Hugging Face repo. MIT would make tiny the least-encumbered model in its tier and the obvious default for commercial local agents. Anything else would be news precisely because it breaks the Ling pattern.
- Watch for independent replication of the agentic wins. The card's GDPval and TAU3-Banking leads are the whole thesis of this model, and they're vendor-run; a third-party harness bake-off against Qwen3.5-4B would settle whether the executor positioning survives neutral ground.
- Watch the KDA diffusion. Kimi Delta Attention is now in Moonshot's own models and Ant's entire 3.0 line. If a third lab adopts it this quarter, it's the de facto linear-attention standard for the open ecosystem — decided by adoption, not by a paper's benchmark table.
- Watch whether tiny's Thinking mode survives contact with real harnesses. Switchable reasoning modes at 1.3B active is an aggressive claim; the failure mode (thinking tokens that burn latency without buying accuracy) will show up in agent-loop completion times first.
References: inclusionAI/Ling-3.0-tiny on Hugging Face · OpenRouter — Ling 3.0 Tiny · AIToolsReview — Ling 3.0 Tiny · Data Science in Your Pocket — Ling 3.0 Flash architecture · related coverage: Nemotron 3.5 Lightning · LFM2.5 local agents · Levels of Open · What Mamba Forgets