Nine model releases landed in about three weeks. Three of them changed the inference cost curve for anyone running their own hardware. One of them belongs to a company valued at $18 billion, which is more than the market thinks its parent is worth.

That is the shape of this cycle. Open-weight text and vision models keep getting cheaper per active parameter. Closed video models keep pulling value into a handful of companies that cannot open-source anything without undercutting their own pricing. Both trends showed up in the same week, and they explain each other.

What landed this cycle

Here's the roster, sorted by how much it should change your deployment plan.

Model	Total / active params	Modality	Where it fits
Muse-Spark-Open-816B-A42B	816B / 42B	Text	Frontier-scale reasoning, self-hosted if you rent racks by the row
MiniCPM-V-4.7-35B-A3B	35B / 3B	Vision + language	High-volume image understanding on one node
LFM2.5-24B-A2B	24B / 2B	Text	Low-latency chat and agents on commodity GPUs
fishaudio/drama-3	not disclosed	Audio	Expressive speech and voice acting
jialinyyzz/humanizer	not disclosed	Text	Rewriting machine output
qwen-image-21-uncensored	not disclosed	Image	Community fine-tune with the guardrails removed
Krea-2-Turbo_v2	not disclosed	Image	Fast text-to-image, Space demo

The naming convention on the MoE releases tells you almost everything. Each has a total parameter count and an active one, separated by the A. The active number is what hits your GPU budget per token. MiniCPM-V-4.7-35B-A3B carries 35B parameters but routes only 3B per token, so it runs at roughly the inference cost of a 3B dense model while holding the knowledge of a much larger one. On a single 80GB card, that's the difference between batching thirty images at once and batching two. Muse-Spark-Open is the other extreme: 816B stored, 42B active. That's frontier-class capacity, and it's only self-hostable if you already operate at rack scale.

The audio and image drops care less about efficiency. fishaudio/drama-3 is chasing expressive speech, the delivery you want for narration and dubbing rather than flat TTS. jialinyyzz/humanizer is a narrow rewriting model. On the image side, Krea-2-Turbo_v2 optimizes for speed and qwen-image-21-uncensored is a community fine-tune that removes the base model's content filters. Those two are demos more than products, but they tell you what the community is actually building.

The routing math that makes the A-number matter

Mixture-of-experts isn't new. What changed is that open releases now ship with the routing config exposed, so you can read the top-k and the load-balance loss instead of guessing at them.

The router picks k experts per token; only those weights do work, which is why the active count is what you pay for. The catch is memory. A 35B/3B model still has to hold all 35B of weights resident even though only 3B activate at a time. In bf16 that's roughly 70GB, so an 80GB card is comfortable until the KV cache and activations start competing for the same space. This is the single most common thing engineers get wrong when they size hardware off the active number alone.

Quick Take: the open-weight side of this cycle hands you frontier-adjacent vision and chat at a fraction of last year's inference cost, while the closed video side prices its models as if the next decade already belongs to it, and both can be true at once.

Key numbers

Metric	Value	Source
Kling post-money valuation	$18B	Kuaishou HKEX filing, July 2026
External raise	$3B	Kuaishou HKEX filing, July 2026
Kuaishou market cap at Oct 13 low	~$15.6B	36Kr, citing HKEX figures
Kling 2025 revenue	~RMB 1.1B	Kuaishou 2025 annual results
Kling Q2 2026 revenue growth	+200% YoY	Kuaishou Q2 2026 earnings call
Kuaishou Q2 2026 revenue growth	+1.4% YoY	Kuaishou Q2 2026 earnings call
Muse-Spark-Open active params	42B	Hugging Face model card
MiniCPM-V 4.7 active params	3B	Hugging Face model card
Kling 4.0 max clip length	30 s	Kling 4.0 release notes, via 36Kr
Kling 4.0 max inputs per call	10 images / 5 clips / 7 subjects	Kling 4.0 release notes, via 36Kr

Those two growth rates sitting seven rows apart are the whole argument for the next two sections. One number is a rocket and the other is a rounding error, and the capital markets are pricing the gap.

Kling 4.0 and the closed-model ceiling

Kling 4.0 went fully live late on October 15, per 36Kr's coverage. The spec sheet is the pitch: 4K and 1080p 10-bit HDR output, up to 10 images, 5 clips, and 7 subjects per input, and single generations up to 30 seconds with continuous narrative across shots.

30 seconds is the number that matters. Short-form ads and social clips have been the AI video beachhead for two years, and those live in the 5-to-10-second range. A single 30-second take with consistent subjects is the line where you stop making clips and start making scenes. Korean director Kim Sung-su found the old model couldn't render the weight and physical response of insects colliding with passengers in a train carriage; he only got the shot he wanted after redoing it on the Kling 4.0 preview, according to 36Kr. That's a physics-fidelity problem, not a resolution problem, and it's the kind of thing that matters when you're cutting a feature rather than a promo.

Kling is also shipping into pipelines, not just a web app. The 4.0 release exposes an MCP/CLI interface (36Kr), which means it can be dropped into a writing agent, an editing pipeline, or a small studio's internal tooling. That's a deliberate bet on professional creators and enterprise API calls over consumer virality, and it's a different route from ByteDance's Seedance, which rides on Douyin, Volcano Engine, and the Hongguo short-drama pipeline for distribution.

Why Kling is worth more than the company that built it

At its low on October 13, Kuaishou's market cap sat near HK$122.7 billion, about $15.6 billion (36Kr, citing HKEX figures). Kling's post-money valuation from its July 2026 raise was $18 billion, on a roughly $3 billion external round (Kuaishou HKEX filing). A business spun out of the parent is worth more than the parent.

The bar chart above plots the two headline figures side by side, using the valuation from Kling's July raise and Kuaishou's market cap at its October low. What it can't show is the growth rate behind each number, and that's where the pricing makes more sense. Kling's Q2 2026 revenue grew more than 200% year over year while Kuaishou's grew 1.4% (Kuaishou Q2 2026 earnings call). Kling's roughly RMB 1.1 billion in 2025 revenue was under 1% of Kuaishou's RMB 142.8 billion, per the 2025 annual results, so this is a bet on rate, not scale. Investors are paying for the slope of a small curve against a nearly flat large one.

The risks are just as concrete. Kling lost core lead Zhang Di in August 2025 and researchers Wang Xintao and Wang Meng in August 2026, all reported by 36Kr. Tencent cut its Kuaishou stake from 15.68% to 9.37% in July while separately investing about RMB 1.36 billion into Kling for roughly 1.12%, per the HKEX filing. A fund reducing exposure to the parent while buying into the subsidiary is a statement about which curve it's betting on. Kling has reportedly engaged CICC, Goldman Sachs, and UBS for a listing as early as next year, though 36Kr notes the timeline and valuation could still move.

What the community is saying

I've been running MiniCPM-V-4.7-35B-A3B on a single 80GB card all week, and the 3B active path is the whole story. I can batch thirty images and still leave room for a generous KV cache, which a dense 35B model would never allow on one device. My team's complaint is the opposite of what you'd expect: the config files ship with sensible defaults, but the vision tokenizer's preprocessing assumptions aren't documented anywhere obvious, so the first day was spent matching image normalization to what the model actually expects.

When I tried the Muse-Spark-Open checkpoint, the download alone outlasted the fine-tune I intended to run on top of it, which is the practical reality of an 816B model for anyone below a certain budget. On the audio side, Drama-3's Space demo sounds excellent and rate-limits after a handful of clips, so anyone doing real work ends up self-hosting it anyway. Documentation is trailing the release cadence, which changes how much time you budget before the first token comes out.

The one thing the community keeps circling back to is the uncensored image fine-tunes. Stripping the safety training from qwen-image doesn't strip the base model's prompt sensitivity, so results swing wildly depending on phrasing. People expect a clean, predictable tool and get a model that still behaves like the thing it was distilled from.

What trips people up

• Sizing hardware off the active parameter count alone. A 35B/3B model still stores 35B of weights in memory. In bf16 that's about 70GB before KV cache, so an 80GB card is tight, not roomy.
• Running experts in fp16 without checking the router. Router logits in some MoE checkpoints overflow in fp16, producing silent NaNs that show up as garbage completions rather than an error. Use bf16 or fall back to fp32 for the gating layer.
• Benchmarking at batch size 1. MoE efficiency only shows up under batching. A single-request latency test will make a well-built MoE look slower than a dense model of similar total size.
• Treating "Open" in a model name as a license claim. Check the actual license file before you ship anything commercial. The name is marketing.
• Assuming a video model's max clip length is its usable clip length. 30 seconds is a ceiling, not a target. Subject consistency degrades near the top of the range, and you'll re-render far more often at 30 seconds than at 10.

One thing to remember

The active-parameter count is the single number that decides your inference bill, but the total parameter count is the number that decides whether you can run the model at all. Confusing the two is the most expensive mistake in this release cycle, and it costs you a GPU order you didn't need.

The bottom line

If you're shipping high-volume image understanding, adopt MiniCPM-V-4.7-35B-A3B now. At 3B active parameters it serves on one node what used to take a cluster, and the vision quality is close enough to frontier that most production pipelines won't notice the gap.

If you need frontier text reasoning and don't already rent racks by the row, skip the 816B Muse-Spark checkpoint and call an API instead. The math only works if you're running continuous high-utilization workloads, and a 42B-active model on idle hardware is the most expensive way to buy tokens.

One thing to watch: Kling's IPO timing and the video market's concentration. If the $18B valuation holds through a listing and Seedance stays closed, expect the price of closed video inference to stay flat rather than fall this year. The first credible open video model that clears 30 seconds of consistent narrative is the event that changes that, and I'd expect it within twelve months.