Ling-3.1-flash-1019 Ties Kimi K3 at 44 — and There Is Still Nothing to Download

The number arrived before the model did. On 22 October the Ling team at Ant Group put Ling-3.1-flash-1019 on its own account with a score and a promise attached: 44 on the Artificial Analysis Intelligence Index, up from 41 for Ling-3.1-flash a week earlier, and level with Moonshot's Kimi K3. The same post says API access and open weights are "coming soon." Artificial Analysis's model page agrees with the first half — 44, ranked third of 174 models it tracks — and with the second: no publicly available API, an undisclosed parameter count, and a licence status of proprietary. So the accurate sentence is not "Ant released a K3-class model." It is: Ant shipped a measured score for Ling-3.1-flash-1019, and as of today nobody outside Ant's own evaluation pipeline can call the build that earned it. This is a what-we-know-so-far readout of a model that is one number and one promise deep, and every figure below carries either a source or an explicit label saying it has none.

The shape of the thing, for anyone who wants it in one breath: Ling-3.1-flash-1019 is a date-stamped checkpoint of the same line as Ling-3.1-flash — a roughly 560-billion-parameter sparse mixture-of-experts with about 25 billion parameters active per token, a hybrid linear-attention stack, and a claimed 1M-token context — and "1019" is a build date, not a new architecture. What is new is the score, the cost of getting it, and the gap between the two labs this comparison now implicates: Ant, which announced rather than shipped, and DeepSeek, whose own dated flash build, DeepSeek V4.1-Flash-1017, went out five days earlier with weights and an unchanged rate card. The distance between those two release styles is most of what this article is about.

What 44 is, and what the page does not say

The Artificial Analysis Intelligence Index is a composite, not a single test. Version 4.3.2 weights ten evaluations: AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience and AA-LCR v1.1. A composite that blends office work, terminal use, competition maths, long-context retrieval and hallucination behaviour will produce a number for almost any model, and that number is only as informative as the rows under it.

For Ling-3.1-flash-1019, the rows are not there yet. The page publishes 44 and ranks it third overall; it lists a 1M-token context, text in and text out, a reasoning model, an undisclosed parameter count, and — the part that matters most — "No publicly available API" wherever per-evaluation results and speed belong. SciCode is marked under review. There is no output-speed figure, no latency figure, and no way to reproduce a single row without the weights or an endpoint. That is not a criticism of Artificial Analysis, which can only measure what it can reach; it is a description of how thin the evidence is at this moment. The model detail below the composite has been measured, in all likelihood on a vendor-hosted or vendor-supplied endpoint, and the details of that arrangement are not public.

Two numbers on that page do more work than the headline, because both are costs rather than scores. The first is output tokens consumed during the index run: 350 million for the 1019 build, against 220 million for the build that scored 41. The second is cost per Intelligence Index task: $1.49, against $1.00 for its predecessor. Read together, they say something a score cannot: the three extra points were bought by thinking roughly 59% longer at the same nominal rate card, and the bill per task rose about 49%. A lab that gains index points by making a model reason longer is not wrong to do so — reasoning length is a legitimate lever — but it does mean the 44 is not free, and that a buyer comparing on index alone is comparing a number whose cost per point has moved in the wrong direction.

The rate card itself is unchanged. Artificial Analysis lists $0.30 per million input tokens and $0.90 per million output for both builds, an 80% cache discount, a blended $0.19 per million at its standard mix, and a 1M context window. Those are the parent's published prices carried over, which is the reasonable assumption when a vendor has announced no new pricing, but it is an assumption. Ant's promotional pricing on the live build — more on that below — runs well under the list rate.

Where the three points came from

Ant has not published a per-evaluation table for the 1019 build. The sub-scores that have circulated come from Chinese-language coverage of the announcement, which reports two movements attributed to Artificial Analysis: AA-Briefcase climbing from roughly 1,400 Elo to 1,524, and AutomationBench-AA from about 62% to 68%. Treat both as second-hand until the evaluation rows appear on the model page — they are reported numbers, not ones this article verified against a primary table.

They are at least internally consistent with what the parent build measured. When we read the Ling-3.1-flash page on 7 October, AA-Briefcase stood at 1,400.35 and AutomationBench-AA at 0.617 — the same values the coverage rounds to "roughly 1,400" and "about 62%." So the direction of travel is plausible, and it points at a specific claim: the update is an agentic and office-work improvement, not a general capability lift. AutomationBench and AA-Briefcase are the two rows in this index that most directly reward multi-step tool use and long-horizon task completion. If those moved and GDPval, SciCode, HLE and CritPt did not, the correct description of Ling-3.1-flash-1019 is "better at finishing tasks," not "smarter."

That distinction has a practical edge. A model that gains on agentic rows is worth testing against your own tool-calling workload; a model that gains evenly is worth testing everywhere. Until the rows publish, the 44 cannot tell you which of those you are holding, and the reasonable posture is the one the account that flagged this build took: show the full scores. The composite is a headline. The rows are the decision.

The Kimi K3-Flash framing, priced

The framing that put Ling-3.1-flash-1019 on a lot of timelines came from a widely followed account on 22 October, which described the build as, in effect, a Kimi K3-Flash — K3-class intelligence at flash-tier economics — and drew a parallel between Ant's progress on this line and DeepSeek's on V4.1-Flash. Neither "Kimi K3-Flash" nor the parallel is a product claim; both are shorthand, and it is worth separating what the shorthand gets right from what it glosses over.

What it gets right is the price-to-index arithmetic, and that arithmetic is stark. Kimi K3 sits at 44 on the same index in its Max configuration. It is also a 2.8-trillion-parameter model with 104 billion active per token, priced by Moonshot at $3.00 per million input and $15.00 per million output, with cached input at $0.30. Ling-3.1-flash-1019 is a 560-billion-parameter line with roughly 25 billion active, priced at $0.30 and $0.90. Same composite score, one-tenth the input rate, one-sixteenth the output rate, and a quarter of the active parameters per token. If all you need is the index, that is not a marginal win; it is a different cost structure for the same headline.

What the shorthand glosses over is everything the index does not measure. Kimi K3 takes image input; Artificial Analysis lists Ling-3.1-flash-1019 as text in, text out. Kimi K3 has been downloadable under its own licence since late July and callable from multiple providers since then; Ling-3.1-flash-1019 has neither weights nor an endpoint. Kimi K3's evaluation rows are published and its speed is measured — 44 is a composite over a visible table. Ling's 44 is a composite over a table that is not yet visible. And the parameter counts are not comparable in the way the word "560B" invites: K3 spends 104 billion parameters per token, this build spends 25 billion.

• Index — 44 for both, on Artificial Analysis Intelligence Index v4.3.2.

• Price — Ling-3.1-flash-1019 $0.30 in / $0.90 out per 1M versus Kimi K3 $3.00 / $15.00.

• Active parameters — about 25B per token versus 104B per token. Total: about 560B versus 2.8T.

• Modality — text only versus text and image input.

• Availability — announced, no API, no weights, versus downloadable and served since July 2026.

• Evidence — a composite with no published rows, versus a composite with a published table and a measured output speed.

Set out that way, "Kimi K3-Flash" is a fair description of the intent and an unfair description of the present tense. The intent is real: if Ant ships this build under the parent's MIT licence at the parent's price, it becomes the cheapest route to K3-level composite performance on the market, by a wide margin. The present tense is that it is a score with no artifact behind it, and the models you can actually rent at 44 are priced nothing like it.

Two labs, one playbook: the date-stamped checkpoint

The more interesting pattern is not the score. It is the form of the release. Both of the labs this comparison touches shipped mid-cycle refreshes this month under the same naming convention — a four-digit month-and-day suffix bolted onto an existing product name — and both asked you to read the suffix as the version.

DeepSeek's entry is DeepSeek V4.1-Flash-1017, dated 17 October, five days before Ant's. Its changelog called the update a "performance-cost Pareto improvement," said the API calling method and pricing were unchanged, and pointed existing calls at the new build. Its weights went up on Hugging Face the same morning under MIT, at 763 billion parameters, with nothing gated. On the index, coverage citing Artificial Analysis places it at 45, one point above Ling's 1019 — a six-point climb from the 39 that the V4.1-Flash (Max) configuration was measured at in early October. Six points in ten days, weights included.

Ant's entry is Ling-3.1-flash-1019, dated 19 October by its suffix and announced on 22 October. Three points in about a week, weights not included, no date given for their arrival. Same playbook, different execution: DeepSeek treated the checkpoint as a ship, Ant treated it as a scoreboard update with a ship to follow. That is not a scandal, and it is not unusual — labs announce ahead of weights constantly — but it changes what the reader can do today, which is the only test that matters for a model you might route traffic to.

There is an operational consequence buried in the naming convention, and it is the reason this pattern is worth tracking beyond one release. When a vendor refreshes a model by changing the suffix and leaving the API model string alone, "the flash model" stops being a stable target. You ask for the model by name, the provider points the name at whatever build it shipped most recently, and your evaluation from last week no longer describes what answered this morning. DeepSeek was explicit about this — the changelog says calls to the existing model string now reach the 1017 build.

Pinning is the fix, and it is a routing concern rather than a benchmark concern. It is also where our own catalogue happens to be useful as an illustration: it lists both deepseek/deepseek-v4.1-flash and deepseek/deepseek-v4.1-flash-1017 as separate model IDs, and kimi/kimi-k3 alongside them, so a caller can name the build rather than the family. Ling-3.1-flash-1019 is not on our catalogue and this article will not pretend otherwise — there is no endpoint for it anywhere we can verify, which is the whole point of the section above.

What is actually open, and what "coming soon" is worth

The build being announced is not the build that is open. The open one is Ling-3.1-flash, whose weights Ant published on 15 October under MIT on Hugging Face and ModelScope. Its card is specific in a way the 1019 announcement is not: 560 billion total parameters, 25 billion active per token, a 572-billion-parameter downloadable checkpoint, 63 KDA layers plus 9 gated MLA layers in a 7:1 ratio and 4 dense layers, 512 routed experts with 8 activated and 1 shared, a vocabulary of 157,184, and a 1M-token context.

The serving story on that parent build is the reason a 560B model is interesting at all. Ant's release notes describe a mixed-precision scheme — MXFP4 for the routed experts, block-FP8 for the dense layers — that cuts the weight footprint from roughly 534GB to 292GB and, in the tested configuration, drops a Hopper deployment from eight GPUs to four. There are recipes for SGLang and vLLM, a long-context parallel scheme Ant calls PCP/DCP, and improved JSON Schema support for tool calls. The measured serving numbers Ant published are unremarkable in the good way: on eight Hopper GPUs, a mean time to first token of 604ms and inter-token latency of 3.28ms at a 76% cache hit rate; on four, about 1.36 queries per second with a P50 time to first token under three seconds.

None of that is confirmed for the 1019 build. It is a reasonable expectation — a date-stamped checkpoint on the same line, with the same parameter counts going unmentioned, is unlikely to have rebuilt its quantization story — but expectation is not documentation, and the two things that would decide whether this release matters at the operational level are both missing. The first is the licence. Ant's own 15 October post describes the parent as open-source and names no licence at all; the Hugging Face card says MIT. Those are different artifacts with different authority, and for the 1019 build neither exists yet. The second is the checkpoint. A 25-billion-active-parameter model that runs on four Hopper GPUs is a genuinely different proposition from one that needs eight, and the entire four-GPU claim rests on a quantization recipe that has not been attached to this build.

So the fair reading of "coming soon" is that it is a scheduling statement, not a technical one. It tells you Ant intends to ship. It does not tell you under what terms, at what size, or whether the deployment maths carries over.

What you can run today

The callable build is the parent, Ling-3.1-flash, and the caveats on it are the kind that only surface after you try to use it. The served context has been 262,144 tokens rather than the 1M the specification claims, with a 32,768-token output cap on the hosted endpoint, and the hosted price has been promotional rather than list. Ant's 15 October notes put the promotion at roughly 75% off: $0.075 per million input, $0.225 per million output and $0.015 per million cached reads on the international side, and ¥0.5 / ¥1.5 / ¥0.1 on Ling Studio. Those are real numbers for the parent, not for the 1019 build, which has no published rate of its own.

If your question is "should I move a production path onto this," the answer is no, and not because the model is bad. A build with no endpoint, no weights, no licence and no published per-evaluation rows cannot be evaluated, only anticipated. If your question is "is this line worth watching," the answer is yes, and the cheap way to act on that is to benchmark the thing it claims parity with. Kimi K3 and DeepSeek V4.1-Flash-1017 are both reachable through one OrcaRouter key at provider list price with 0% markup, so running the same harness across both tells you what 44 and 45 actually cost you per task before Ant ships anything. Automatic failover matters here for a mundane reason: when you are pointing a fresh evaluation harness at an unfamiliar build, you want a bad endpoint to cost you a retry rather than an afternoon. The routing DSL will put either model behind a single endpoint if you would rather not rewrite the client twice.

What would settle it

Four things, in descending order of how much they would change the picture.

• The evaluation rows. A published per-evaluation table on the Artificial Analysis page would turn 44 from a headline into a profile, and would tell a reader whether this is an agentic improvement or a broad one. SciCode coming off "under review" matters more than usual, because coding is where a cheap model displaces an expensive one first.

• The weights, with a named licence. MIT would make the four-GPU claim portable and turn this into the cheapest K3-class model anyone can self-host. A restricted or bespoke licence would make the 44 a marketing number with a contract attached, and would be its own story.

• An API, so that speed and latency exist. Every model on the index that has an endpoint has a measured output speed; this one has none, which means the efficiency half of the "flash" claim is currently untested. A flash model that reasons 59% longer is a different product from one that reasons the same amount faster.

• A single harness across all three. No published benchmark yet runs Ling-3.1-flash-1019, DeepSeek V4.1-Flash-1017 and Kimi K3 on the same evaluation, and until one does, every comparison in this article — including the one in the post that started it — is a comparison of composites computed at different times on different configurations.

Until at least the first of those lands, the accurate summary of Ling-3.1-flash-1019 is this: a real score, from a real lab, on a model that does not yet exist as an artifact. Ant has now announced two Ling-3.1-flash builds before either was fully available — the September launch with no weights, and this one with neither weights nor an endpoint — and the pattern is legible enough to plan around: treat an Ant suffix as a signal that something is coming, and treat DeepSeek's as a signal that something is here. The 44 is worth knowing. It is not yet worth routing.