Kimi K3 & Qwen 3.8: Open Weights You Can't Run (2026)

Two trillion-parameter “open” models got announced in ten days. One of them has now actually shipped its weights, on time and under a real license, and you still can’t run it.

That’s the update, and it’s the more interesting version of the story I published here on July 19. Kimi K3 landed July 16 at 2.8 trillion parameters with weights promised for July 27. Alibaba announced Qwen 3.8 three days later at 2.4 trillion, weights “soon.” I wrote then that both were running ahead of reality. Eight days on, one of them has proved me half wrong in the best way, and the other has gone quiet exactly as its own track record predicted.

Update, July 27: Kimi K3’s weights shipped

Moonshot said July 27 and delivered July 27. The weights are up at moonshotai/Kimi-K3, with a separate MXFP4 repo alongside. No slip, and no closed Max tier wearing the open name. Credit where it’s due: this is what keeping a promise looks like, and it’s worth saying plainly, because the rest of this piece is about why it still doesn’t help you.

The license needs describing carefully, because a lot of coverage got it wrong in both directions. The repo calls it the Kimi K3 License. It is MIT-derived, carrying the MIT grant and the AS IS disclaimer verbatim, with three additions:

Standard attribution, plus a requirement to comply with applicable law.
A Model-as-a-Service clause. If you run a MaaS business with aggregate revenue above $20M over any consecutive 12 months, you need a separate agreement with Moonshot before commercial use. MaaS here means giving third parties inference or fine-tuning access with meaningful control over inputs, parameters, or training data. It explicitly excludes end-user products with embedded model features, and excludes merely relaying requests.
An attribution clause for scale. Commercial products above 100M monthly active users or above $20M in monthly revenue must display “Kimi K3” prominently in the UI.

Internal use is exempt, as is use through Moonshot’s own products or certified inference partners.

So for you, and for basically any individual or small team: download it, fine-tune it, deploy it, sell what you build with it. Effectively unrestricted. It is open weights, not OSI open source — revenue-gated conditions fail the no-discrimination test, so the label matters if you care about the distinction. What it isn’t is the flat “Modified MIT, commercial use permitted” that went around, which drops the thresholds entirely, or the “Apache 2.0” that several outlets printed and the repo does not say.

104B active, 2.8T resident: the whole problem in two numbers

Here is the architecture, from the model card rather than the recaps: 2.8T total parameters, 896 experts with 16 selected per token plus 2 shared, 104B activated per token, a 1,048,576-token context, and native vision. Attention is Kimi Delta Attention (KDA), which is why serving it efficiently needs an engine with KDA support rather than whatever you have lying around.

Watch what those two numbers do. 104B active means that per token, K3 does roughly the compute of a 100B dense model. On hardware that can hold it, that’s quick. It is not a 2.8-trillion-parameter-sized wait.

But all 2.8T have to be resident before it generates a single token, and that’s where it ends for consumer hardware. Moonshot’s own tech blog points at “supernode configurations with 64 or more accelerators.” Download-size figures circulating in coverage disagree with each other — roughly 594 GB in some write-ups, roughly 1.4 TB in others — and the repo is gated, so I haven’t verified either firsthand. It doesn’t matter which is right. The smaller of the two is still about 25 times a 24GB card and roughly nine times a 64GB desktop’s entire system RAM. The “8× H100 minimum” line making the rounds doesn’t survive arithmetic on the larger figure either: eight 80GB cards is 640GB, and you cannot load 1.4 TB into 640 GB.

I can’t give you firsthand K3 numbers. We measure what we can put on a bench, and nothing on our bench comes within two orders of magnitude of this.

This is the active-versus-total rule at its most extreme, and it’s the rule our own measurements established. On a 12GB RTX 3060 we ran Qwen3.6-35B-A3B at 38 tok/s with most of its experts living in system RAM, because only ~3B parameters fire per token. That same page shows a dense 14B collapsing 6.3x under the same treatment. Active parameters set your speed. Total parameters set whether you can start.

One limit, learned since: that’s a guide to whether a model loads and roughly how fast, not a way to rank two of them against each other. Only the routed-expert share of those active parameters crosses the bus when experts live in RAM, and it varies enough that Trinity-Mini and Qwen 3.6-35B-A3B, both advertising about 3B active, landed roughly a third apart on one 3090. The MoE primer has the measurements.

K3 is what happens when you push both dials to the end of the board. The active count is friendly. The total count is a datacenter.

And the specific technique that rescues MoE models on small cards — --n-cpu-moe, pushing idle experts out to system RAM — does nothing here. It works by trading VRAM for ordinary RAM. K3 needs more RAM than a consumer motherboard physically accepts. There’s no overflow tier left to spill into.

One genuinely interesting detail for anyone tracking quantization: K3’s MXFP4 weights come from quantization-aware training applied from the SFT stage onward, with MXFP8 activations. The 4-bit weights are the model, not a lossy copy made afterward. That’s the same shift we covered in the FP4 inference guide, now shipping at the largest scale anyone has attempted.

On quality, it earns the attention: K3 debuted at #1 on Arena’s Frontend Code leaderboard at 1,679, ahead of Claude Fable 5 (1,631) and GPT-5.6 Sol (1,618) — the first open-weight model to take that top spot. On the Artificial Analysis Intelligence Index it scores 57.11, in the same band as Opus 4.8 and GPT-5.5. This is a frontier model by any honest reading. It’s just a frontier model you’ll be renting, not hosting.

Update, August 4: the fork resolved, and a 27B has a name

Qwen3.8-Max launched August 3: 2.4 trillion parameters, a million-token context, text and image and video in, hosted API only. Alibaba says the open weights ship “next week.” No date, same as the July 19 post, now with a shorter fuse.

The part this page has been waiting on: Alibaba named a second checkpoint, Qwen3.8-27B, and said it is going open-weights too. That is the signal the July 27 update said to watch for, a named small variant. If it lands, 3.7 reads as one generation held back rather than the start of a permanently closed Max tier.

Nothing has shipped. I checked the Qwen Hugging Face org on August 4: newest repos are still Qwen3-ASR and the AgentWorld MoE. No Qwen3.8 of any size, no 27B, no model card. The license is unstated. Qwen 3.5 and 3.6 were Apache 2.0, but precedent is not a commitment and there is no 3.8 license text to read. The 27B’s parameter count, context, and quantized size are all unpublished, so I can’t tell you what card it fits.

One figure to handle carefully: a 95B active-parameter number is circulating for Max, and it isn’t Alibaba’s. It traces to a third-party summary, and MarkTechPost reports the activated count as undisclosed. This page’s own rule applies to this page. Watch the org, not the tweets.

None of this moves the thesis. 2.4 trillion parameters is a server-room object under any license, and open-weighting it does not put it on your GPU. What changed is that a runnable sibling now has a name and not yet a spec sheet.

Qwen 3.8: still nothing, eight days later

Moonshot’s delivery makes the comparison sharper, so here is where Alibaba stands. What it posted on July 19, from the official @Alibaba_Qwen account: “Qwen3.8 is launching and going open-weight soon!” With, it says, 2.4 trillion parameters, “second only to Fable 5.”

Read that carefully, because two different things are wearing one name.

The thing that’s live right now is Qwen3.8-Max-Preview. It’s a closed, paid API tier — routable through Alibaba’s Token Plan and Qoder, same subscription bucket as qwen3.7-max and glm-5.2. No model card, no per-token price sheet, no benchmark table, no weights. You can pay to call it. You cannot download it.

The thing that would actually matter to you, open weights, still does not exist. I rechecked Hugging Face’s Qwen org on July 27: the newest repos are Qwen3-ASR and an AgentWorld MoE, both from weeks back. No Qwen3.8 of any size, and no small variant has been announced even as a plan. The tweet promised “soon” with no date, and eight days later “soon” is still all there is.

Even when it lands, a 2.4-trillion-parameter model is a datacenter object no matter how hard you quantize it, the same shape as Inkling’s 975B whose smallest 1-bit quant is 270GB. K3 just demonstrated exactly what that means in practice: fully open, fully licensed, fully undownloadable by anyone without a server room.

And “second only to Fable 5”? That’s Alibaba’s own claim — the tweet literally says “we believe.” It isn’t a third-party result. Artificial Analysis, the independent index, doesn’t list Qwen 3.8 at all. Its current top looks like this: Claude Fable 5 at 60, GPT-5.6 at 59, Kimi K3 at 57, Claude Opus 4.8 at 56. A genuine number-two would sit around 59, above GPT-5.6. Maybe Qwen 3.8 lands there when someone measures it. But we’ve seen this exact move one generation ago: Qwen 3.7 Max was billed as China’s number one, and when Artificial Analysis actually scored it, it came in at 56.6 — fifth. Still strong. Just not what the announcement said.

I want to be precise about the criticism here, because it’s easy to overshoot. Qwen builds good models. 3.7 Max at fifth in the world is a real achievement. The problem isn’t the models. The problem is a self-reported ranking presented as fact before anyone independent has checked it, attached to weights nobody can download. Claim and delivery, both running ahead of reality.

What Qwen’s own history says about your 3.8 weights

Now the useful part, the one you can’t get from the horse-race coverage. If you want to know when — or whether — you’ll get runnable 3.8 weights, the best predictor is Qwen’s own track record. I pulled the actual Hugging Face commit dates for the last two generations.

Qwen 3.5 launched February 16, 2026, with the big open 397B-A17B on day one. The runnable weights followed fast: the 35B-A3B MoE and the 27B dense both hit Hugging Face on February 24, and the 9B — the one that fits an 8GB card — on February 27. Call it eight to eleven days from flagship to something you can actually load.

Qwen 3.6 was even tighter. The 35B-A3B open MoE landed April 15–16, the 27B dense April 21–22, and the closed Max-Preview showed up around April 20 — meaning the open weights arrived alongside the paid tier, the MoE actually a few days ahead of it. No wait at all.

So for two straight generations, Qwen was one of the good actors: runnable open weights, consumer-sized, within a week or two of the headline, sometimes before it. That’s the pattern that makes 3.8’s silence loud.

Because then came Qwen 3.7, and the pattern broke. Qwen3.7-Max shipped in May as a proprietary, API-only model. No open weights ever followed — not a 27B, not a 9B, nothing. I checked again this week: there is still no official open Qwen 3.7 of any size. The open line went 3.6, skipped 3.7 entirely, and now arrives at 3.8 as a 2.4T model with no small variant announced.

So here’s the honest forecast, and it’s a fork, not a date. If 3.8 follows the 3.5/3.6 cadence, a runnable 27B or 35B-A3B-class open variant appears within roughly one to two weeks and this whole worry evaporates. If it follows the 3.7 precedent, Max tier only with the runnable open tier quietly skipped, you get a 2.4T datacenter drop labeled “open” and nothing for your GPU.

Update, July 27: day eight, and the 3.5/3.6 window is closing. Both of those generations had shipped consumer-sized weights by now — 3.5 took eight to eleven days, 3.6 didn’t make anyone wait at all. Nothing has appeared. That isn’t proof of the 3.7 outcome yet, and I’m not going to call it one, but the branch that would have made this a non-story has mostly run out of room. Alibaba has never pre-committed a date for the small weights, so anyone telling you a specific day is guessing. Watch the Qwen Hugging Face org, not the tweets. Weights or a named small variant is the only signal that counts.

Update, August 4: a named small variant now exists. Alibaba announced Qwen3.8-27B for open weights on August 3, alongside Max. That is the signal this paragraph was waiting on. It has not shipped, and no specs or license are published yet.

Did they say why?

No. I looked — the announcement thread, the Qwen blog, statements from the team — and Alibaba has given no reason for going Max-first or for the open-weight delay. The tweet just says “soon.” I’m not going to invent a motive, and you should be skeptical of anyone who does. We know the pattern (Max-first, twice now). We don’t know why, and there’s a real difference between reporting the first and guessing the second. For the longer view on where Qwen’s open posture may be heading, I dug into that separately in Is Qwen Going Closed?.

There is one piece of documented external context worth knowing — and only as context. On July 7, Reuters reported that China’s Ministry of Commerce held talks with Alibaba, ByteDance, and Z.ai about restricting overseas access to the country’s most advanced AI models, talks that per the reporting explicitly covered open-weight releases and even unreleased future models. Alibaba was a named participant. That’s real, recent, and relevant to open-weight timing across every Chinese lab. It is not a stated cause of anything here: the talks are at the consultation stage, no rules exist, no timeline was given, and Alibaba has not connected them to Qwen’s release plans. File it as background, not as the reason for 3.8. The honest position is that we know the pattern and we don’t know the motive.

Who actually ships weights you can run

The trillion-parameter labs aren’t the whole field. Some labs still ship consumer-sized open weights the day they announce, and it’s worth naming them, because it’s a real difference in behavior, not vibes.

Google’s Gemma 4 is the clean example. It launched April 2, 2026, Apache 2.0, with open weights on Hugging Face the same day — and in sizes that fit one card. The 31B dense runs in about 17–18GB at Q4 on a 24GB GPU. The 26B-A4B MoE (roughly 4B active per token) is lighter still and comfortable on 16GB. No trillion-parameter flex, no “coming soon.” Announce and ship, runnable, day one.

The contrast with the other big open releases is instructive. Meta’s Llama 4 was open at launch in April — but Scout (109B total) and Maverick (~400B) are datacenter MoE; neither runs on a single 24GB card in any honest sense. Mistral’s newest, Small 4, is Apache 2.0 and open, but it’s a 119B MoE that only “loads” on 24GB via heavy RAM offload at single-digit tokens per second. The genuinely runnable Mistral is still last generation’s Small 3.2 24B dense. So “open at announcement” and “runnable on your hardware” have quietly split into two different things, and only a couple of labs — Google most cleanly, and Qwen itself back in the 3.5/3.6 days — deliver both at once.

That’s the real fault line in mid-2026 open AI. Not China versus America. Not Qwen versus Kimi. It’s whether a lab ships weights sized for the machine you own, or weights sized for the machine they own.

Kimi K3 is the cleanest illustration yet, which is why it deserves better than being filed under broken promises. Moonshot did everything right on the openness axis: shipped on the announced date, permissive MIT-derived terms, and released 4-bit weights that are the ones the model was trained into rather than a degraded copy made afterward. Every criticism you could aim at Qwen 3.8’s silence, K3 answers. And the result on your desk is identical to a model that was never released at all, because openness and runnability turn out to be independent axes. A lab can score full marks on the first and still hand you something you have no way to load.

Run this today

While the trillion-parameter headlines play out, here’s what actually fits your GPU right now. None of it is new this week, which is exactly the point — the runnable tier didn’t move, so you don’t have to wait for anything.

The best open model you can run on a single 24GB card is still Qwen 3.6-27B dense. On the independent Artificial Analysis Intelligence Index it scores 37 — number one among open models in its class — with Gemma 4 31B behind it at 29. And it’s genuinely fast: on my own RTX 3090 the 27B runs around 38 tok/s at Q4 on llama.cpp, and about 2.5x that with DFlash speculative decoding on batch coding and math. If you’d rather have the MoE, our 35B-A3B guide clocks Qwen 3.6-35B-A3B at 101 tok/s at UD-Q4_K_XL (22.4GB) on a 3090, because only 3B of its parameters fire per token. Either one is the “run this instead of waiting” answer.

Stepping down the tiers:

16GB: Gemma 4 26B-A4B (the MoE, ~4B active) is the natural fit, or Qwen 3.6-35B-A3B at UD-Q3_K_M (16.6GB) if you want the Qwen family.
12GB: the smaller Qwen 3.5 dense models — the 9B especially — or Gemma 4’s compact variants. Quality per gigabyte here is the best it’s ever been.
8GB: Qwen 3.5-9B at Q4 lands around 5–6GB and leaves room for context. For the full tier-by-tier breakdown, including the uncensored/abliterated options, the best uncensored local LLMs guide and the VRAM requirements guide go deeper than I can here.

If you’re weighing whether a bigger model is worth the squeeze, the running 70B models locally guide has the honest math on where quantization stops paying off.

The bottom line

Two trillion-parameter models went “open” in ten days and neither one is for you, for two completely different reasons. Kimi K3 shipped on time under a license that genuinely lets you do what you want, and needs 64 accelerators to hold. Qwen 3.8’s weights are still unshipped, promised “next week” as of August 3, at 2.4T, with a 27B now named but unspecified and a top-of-the-charts claim no independent index has confirmed. K3 is a physics problem. Qwen 3.8 is a delivery problem. Neither is a scandal, and only one of them is anybody’s fault; this is mostly just a category of release that stopped being aimed at consumer hardware.

The good news is the tier that is aimed at you never slowed down. Qwen 3.6 still tops the independent open-model charts on a 24GB card and runs at 38 tok/s on a used 3090. Gemma 4 ships runnable and Apache-licensed the day it’s announced. The headlines will keep chasing the trillion-parameter number, and you can keep ignoring them, because the model that fits your GPU is already downloaded, already fast, and already better than anything you could have run a year ago.

Watch the Qwen Hugging Face org for a small 3.8 variant. Day eight has passed with nothing, which is later than either of the last two generations took, so I’d temper your expectations. And watch what Moonshot does next, because K3 proved the openness half of this is solvable by a lab that wants to solve it. The half nobody has solved is making the result fit in 24GB.

Either way, you didn’t lose anything this month. You were never the audience for the 2.4T drop, and you weren’t the audience for the 2.8T one either — even after it arrived, on time, with the license it promised.