Kimi Linear Is Spreading: Every Model We Can Verify Actually Runs KDA

"Wow! Kimi-Linear is gaining adoption! This is like the third external model I see." That was the entire post — one line from @teortaxesTex on August 5, 2026, with a screenshot attached and not a single model named. The shortened link in it resolves to the image, not to a repo, so there is nothing to click through to and nothing to corroborate. The claim is checkable anyway, and it matters more than a one-liner suggests: if labs other than Moonshot AI are now building on the attention design behind Kimi-Linear-48B-A3B-Instruct and Kimi K3, then Kimi Delta Attention has stopped being one lab's in-house trick and started becoming an ingredient other people buy into. So we went looking for the models instead of the screenshot. The short answer: the strongest case is Ling-3.0-flash, whose own model card describes a 5:1 stack of KDA and MLA layers; the most useful case is a US lab's research checkpoint that quantifies what pure KDA costs; and at frontier scale, outside Moonshot, nobody has shipped it.

What the claim is, and what it is not

One practitioner, one screenshot, no model named, no corroboration. That is the whole evidentiary basis for "gaining adoption," and it should be read as a prompt to go check rather than as news. We could not reproduce the screenshot, so nothing below is a confirmation of it — everything that follows is what public model metadata showed on August 5, 2026, when we swept for it, plus what each lab says in its own model card. Where a number comes from a vendor's own measurement, it is labeled that way, because in this particular story almost every headline number does.

One screen on Kimi Linear, because "adoption" means adopting this

Kimi Linear is an attention architecture, published by the Kimi Team as arXiv 2510.26692 on October 30, 2025, not a product. Its core is Kimi Delta Attention (KDA), which takes Gated DeltaNet and replaces its coarse forget gate with fine-grained, per-channel decay, so the recurrent state learns what to keep channel by channel rather than wholesale. The update is restructured into chunked, Diagonal-Plus-Low-Rank form specifically so it lands on tensor cores instead of leaving them idle. That hardware framing is the point of the design: linear attention has always been cheap in theory, and usually disappointing in practice.

The released checkpoints — Kimi-Linear-48B-A3B-Base and Kimi-Linear-48B-A3B-Instruct — are 48B total parameters with 3B active, a 1M-token context window, and an MIT license. Layers alternate roughly 3:1, twenty KDA layers to seven Multi-Head Latent Attention layers, so three out of four layers are the cheap recurrent kind and every fourth keeps exact global attention.

What Moonshot reports, all measured against a full-MLA baseline trained on an identical recipe — vendor numbers, not independently reproduced:

• Decoding throughput — up to 6x faster time-per-output-token at 1M context versus full MLA

• KV cache — up to 75% smaller, which is where the throughput comes from

• Long context — RULER at 128k: 84.3 for Kimi Linear with a 3.98x speedup, against 81.3 for the MLA baseline

• Short context — MMLU-Pro at 4k: 51.0 against 47.2 for the MLA baseline and 47.9 for a Gated DeltaNet hybrid, at roughly the same speed as full attention, so the design is not buying long-context speed by giving up short-context quality

• Pretraining ablation — the 3:1 hybrid reaches 5.65 validation perplexity where full attention lands at 5.77

The honest limit on all of that: the matched-baseline evidence stops at 48B. Nobody built a full-attention 2.8T twin of Kimi K3 to compare against, and as of a late-July 2026 fact-check of the K3 architecture claims, no independent shortcut-free long-context recall probes had been published for either model. The efficiency story is well supported. The "outperforms full attention" story is supported at research scale by the lab that wrote the paper.

Traction, so far, has looked like downloads rather than like other people's architectures: 98,164 downloads in the last month, 570 likes, one inference provider listed on Hugging Face, and a model tree of 2 adapters, 8 finetunes and 24 quantizations. Popular, clearly. Copied is a different question.

The strongest adoption case: Ling-3.0-flash

Ling-3.0-flash is the one that makes the claim real. Its Hugging Face card describes a native hybrid linear attention architecture built from a 5:1 alternating stack of Kimi Delta Attention and MLA — 35 KDA layers to 7 gated MLA layers — on a 124B-parameter Mixture-of-Experts model with 5.1B active per token, a 256K context trained in an 8K to 32K to 256K progression, and an MIT license. That is not a research toy from a hobbyist; it is a shipping model from an established lab, and it names Moonshot's attention module in its own architecture description.

It is also more aggressive about it than Moonshot is. Moonshot keeps one exact-attention layer in four; Ling-3.0-flash keeps one in six. If a design is a liability you hedge it; you do not spend fewer global layers than the people who invented it. Read alongside the earlier generation, that is a switch: the February 2026 architecture surveys that catalogued this class of model had the previous Ling flagship on Lightning Attention paired with MLA at a 7:1 ratio. The mechanism changed between generations, and the direction it changed in was toward KDA.

Two caveats we will not paper over. This is card-reported: we read the repository's own architecture description and its config, which declares a custom bailing_hybrid model type with a BailingMoeV3 implementation — we did not audit the kernels to confirm the math is KDA rather than KDA-inspired. And the repo carries no KDA tag at all; what surfaces in a tag search is a third-party GGUF conversion someone else labeled. Which is a useful warning about method, and leads directly to the next two sections.

Arcee AI ran the experiment Moonshot didn't

The most informative external use of KDA is not a product at all. Around six weeks after the Kimi Linear paper, in mid-December 2025, Arcee AI published two Apache-2.0 research checkpoints — AFM-4.5B-Base-KDA-Only and AFM-4.5B-Base-KDA-NoPE — under their own ArceeKDAForCausalLM implementation, tagged, explicitly, kimi-delta-attention. Their question was the one Moonshot's hybrid design carefully avoids: can full attention be replaced entirely? So they converted all 24 attention layers to KDA, with no global attention layers left, and distilled the result from the original AFM-4.5B-Base as teacher at 32k sequence length.

Their reported results, teacher versus pure-KDA student:

• MMLU — 63.1 to 55.8, a real but survivable drop

• GSM8K — 52.1 to 26.8, roughly halved

• HellaSwag — 78.0 to 74.3, close to intact

The shape of that damage is the interesting part. Broad knowledge and commonsense continuation mostly survive being pushed through a fixed-size recurrent state. Multi-step arithmetic, which needs exact retrieval of intermediate results the model wrote several steps ago, does not. Arcee also reports that long-context degradation is graceful, with no sharp cliff. Taken together, it is the clearest public evidence for why every serious KDA deployment is a hybrid: Moonshot's one-in-four global layers and Ant's one-in-six are not timidity, they are the part that keeps the arithmetic.

What it is not: a verdict on KDA at scale. This is a 4.5B distillation, not a pretraining run, and distillation into a changed architecture loses things that pretraining from scratch would not. Read it as the ablation that a vendor had no incentive to publish — from an independent lab that had no stake in the answer.

The long tail: a cluster of tiny KDA repos in a single week

Beneath the two serious cases sits a scatter of small ones, and the dates are what make them worth mentioning. NEXIVON-1B-BASE, uploaded July 31, 2026, declares its model type as, literally, kda — Apache-2.0, 285 downloads, no likes. qingyi-kda-0.6b, the same week, is a Qwen3-0.6B-Base finetune shipping a custom qingyi_kda implementation. Scrapegoat-Tiny-Coder, also that week, combines a kda tag with a "dual-track parallel attention" design of its own. Two more, maccy-106m-base and maccy-96m-16k-base, were tagged kimi-delta-attention on July 27 and 28.

None of these is important on its own — single-digit likes, unknown authors, research-scale parameter counts. Collectively they are probably what the "third external model I see" tally is counting, and we cannot confirm which three, since the post named none. The signal in them is not the models, it is the ten days: four independent small training runs reached for KDA inside a week and a half, which happens when a kernel stops being a research artifact and becomes something you can drop into a training script over a weekend.

What the sweep did not find

We queried Hugging Face for every repository carrying the kimi_linear model type. There were 47. All but three have "Kimi" in the name, and the ones that are not Moonshot's own are derivatives rather than adopters: AWQ, GPTQ, FP8, NVFP4, SINQ, GGUF and MLX quantizations at every bit depth; REAP expert-pruned 35B variants from Cerebras; and FlagRelease FlagOS builds of the Instruct checkpoint for NVIDIA, MetaX and Hygon accelerators. That last group is genuinely interesting for a different reason — someone did the work to make KDA run on domestic Chinese silicon — but none of it is another lab designing another model.

No frontier-scale model outside Moonshot declares KDA. Ling-3.0-flash at 124B total and 5.1B active is the ceiling of external adoption right now.

And the method has a hole worth naming, because it cuts against our own negative result. A lab that reimplements KDA registers its own model type — arcee_kda, qingyi_kda, bailing_hybrid — so tag sweeps systematically undercount exactly the adopters we are looking for, and a model can use the mechanism without ever writing "KDA" in its card. Absence from a tag is weak evidence, not proof. If there is a fourth or fifth quiet adopter, this is precisely how it would stay invisible.

Everyone else picked a different linear attention

The reason "Kimi Linear is gaining adoption" is a story at all is that, for most of 2026, it wasn't. Hybrid linear attention swept the open-weight field — but not this variant of it. Per the architecture surveys published in February 2026: Qwen3-Next 80B-A3B and Qwen3.5-397B-A17B both pair Gated DeltaNet with gated attention at 3:1, meaning Qwen3.5-397B-A17B runs 45 recurrent layers against 15 full-attention ones out of 60. The previous Ling flagship used Lightning Attention with MLA at 7:1. Nemotron 3 Nano 30B-A3B and Super 120B-A12B are Mamba-2 hybrids, with just 6 attention layers in a 52-layer stack and 8 in an 88-layer stack respectively. GLM-5 stayed with MLA and added sparse attention on top. MiniMax-M2.5 went the other way entirely and kept plain multi-head attention, reportedly for reliability. And Kimi K2.5, Moonshot's own flagship before K3, was MLA — the lab published KDA in October 2025 and did not put it in its frontier model until July 2026.

So the category won and the variant did not. Everybody agrees three-quarters of your layers can be recurrent; nobody agreed on which recurrence. KDA's differentiator is the per-channel decay gate, and its cost is that you need its custom kernels and its prefix-caching plumbing rather than the Gated DeltaNet path that half the ecosystem already had. Until this month, one lab paid that cost.

The adoption that already happened is in the serving stacks

Architectures do not spread because they are elegant; they spread when the infrastructure makes them cheap. That part is already done for KDA, and it happened faster than the model-card adoption. Both vLLM and SGLang shipped day-zero support when Kimi K3's weights landed on July 27, 2026, with Moonshot upstreaming its KDA prefix-caching implementation into vLLM itself rather than maintaining a fork. SGLang shipped fused KDA and Gated DeltaNet kernels and reported logical KV capacity rising from 1.5M to 12.2M tokens on identical hardware — its own measurement, and the most concrete third-party efficiency figure in this whole story. The reference kernels live in fla-core, which any small training run can import.

That is the difference between Arcee needing a deliberate research effort in December 2025 and four anonymous repos casually declaring KDA in one week of July 2026. The mechanism became a dependency you install. Whether the frontier follows is a separate question, but the friction argument against KDA has largely evaporated.

What it changes if you are only buying tokens

Nothing about your code. You never call KDA; you experience it as what a million-token context costs and how long the first token takes. Kimi K3 is the model where that shows up commercially today — on OrcaRouter it runs $3.00 per million input tokens and $15.00 per million output, with the full 1M-token context and a measured p50 time-to-first-token of 8.88 seconds over the last seven days. Those economics are, substantially, what the 3:1 KDA hybrid buys: serving a million-token context at a price that is merely expensive rather than prohibitive.

The research checkpoint is a different proposition. Kimi-Linear-48B-A3B-Instruct is MIT-licensed with one inference provider listed on its Hugging Face page, and it is not on OrcaRouter — if you want to run it, that means self-hosting or that provider. The arithmetic for self-hosting is friendlier than the parameter count suggests: 48B weights in bf16 is roughly 96 GB, so two 80 GB cards, while the 4-bit MLX and GGUF conversions land near 25-30 GB and fit a single card or a well-specced Mac. Ling-3.0-flash is likewise not on OrcaRouter today. We are not going to pretend otherwise to make a point about routing.

Where routing does matter here is the cost of being wrong about an architecture bet. Hybrid linear attention models are exactly the class where the vendor's benchmark table and your workload can disagree — a fixed-size recurrent state fails on retrieval patterns that no aggregate score exposes, which is the lesson of that halved GSM8K number. Running the comparison through one API key across 200+ models means the test is a changed model string rather than a procurement cycle, at provider list price with 0% markup on our side, and with automatic failover so a long-context model you are still evaluating is not a single point of failure in a production path. Try it on your own long-context traffic for a day. That is a cheaper answer than any benchmark table, including ours.

Questions worth actually asking

Is Kimi Linear the same thing as Kimi K3?

No, and conflating them is the most common error in coverage of both. Kimi Linear is the architecture paper plus a 48B-total, 3B-active research model with a 1M context, released under MIT in late 2025 to demonstrate that a KDA-heavy hybrid beats full MLA at matched training. Kimi K3 is Moonshot's 2.8-trillion-parameter flagship from July 2026 — 104B active, natively multimodal, 1M context — which took the KDA-at-3:1 idea to frontier scale and added Attention Residuals, a separate trick that the lab reports buys around 25% training efficiency for under 2% extra cost. Shared mechanism, entirely different scale, capability and price. Benchmarks from one tell you very little about the other.

Why would a lab choose KDA over Gated DeltaNet, given that most chose Gated DeltaNet?

KDA is a strict refinement of Gated DeltaNet, so the question is whether the refinement is worth the switching cost. The refinement is the gate: Gated DeltaNet decays its recurrent memory with one scalar per step, KDA learns a decay per feature channel, which lets the state hold a variable name across ten thousand tokens while flushing the sentence it appeared in. Moonshot's ablation puts that at roughly 0.12 validation perplexity at 48B, and its chunked DPLR formulation was written to keep tensor cores busy, so the extra expressiveness does not cost the throughput it earns. Against that, Gated DeltaNet already had broad kernel and framework support before either was fashionable, which is worth real engineering time. The 2026 answer was mostly "not worth it." Ling-3.0-flash is the first production-scale bet the other way.

Should any of this change what I deploy this quarter?

Only if you are running very long contexts. If your prompts are under about 32k tokens, hybrid linear attention is an implementation detail with no bearing on your model choice; pick on quality, price and latency as usual. If you are pushing 128k and beyond, it is worth knowing that the models keeping your costs sane at that length are increasingly KDA hybrids, and that their published long-context scores are aggregate benchmarks rather than adversarial recall probes. Test retrieval at your real context length rather than trusting a RULER score. That advice would be identical if the mechanism were Gated DeltaNet or Mamba-2.

Where this leaves the claim

Directionally right, and smaller than it sounds. What is verifiable on August 5, 2026 is one production model from an established lab, one research pair from an independent US lab that published the honest downside, a handful of sub-1B repos from the past ten days, and a serving ecosystem that has already absorbed the kernel. That is more than "one lab's trick" and much less than a standard. The number to watch is not three external models — it is whether the next frontier open-weight release from a lab that is not Moonshot ships per-channel gating, and whether anyone outside Moonshot ever publishes a recall probe that tests what the fixed-size state actually forgets. Both would tell you more than another screenshot.