Chinese Open-Source AI Models: Kimi K3, Qwen 3.8, GLM 5.2

> **TL;DR:** Four Chinese labs — Moonshot AI, Alibaba, Z.ai and DeepSeek — have each released a frontier-class model with open weights and no license fee, including Kimi K3, billed as the largest open-source model shipped to date. DeepSeek's V4 Pro Max was reportedly trained for under $6 million and is served at $0.40 per million input tokens and $0.87 per million output tokens, roughly 10–50x below comparable GPT and Claude tiers. Nearly 60% of AI usage inside US companies is now attributed to Chinese models, with Pinterest, Airbnb and Coinbase named among users.

Key Takeaways

- Kimi K3 launched as the largest open-source model released so far and reportedly beats GPT-5.6 and Fable 5 on some benchmarks — demand forced Moonshot to pause new sign-ups. - Qwen 3.8 (Alibaba) and GLM 5.2 (Z.ai) landed in the same tier, both free and open-weight, as part of an ecosystem-building play rather than a revenue play. - DeepSeek V4 Pro Max was reportedly trained for under $6 million using 256 hyper-specialised experts with only ~8 activated per query. - Multi-head latent attention compresses the KV cache by a claimed 90%+, making long conversations dramatically cheaper to serve. - US export controls on H100s, TSMC and ASML pushed China toward domestic silicon (Huawei Ascend, SMIC) and state-run shared compute rather than out of the race.

Four Chinese labs shipped frontier-class models with downloadable weights in the space of a single news cycle, and none of them charges a licence fee. Moonshot AI's **Kimi K3** arrived as the largest open-source model released to date. Alibaba's **Qwen 3.8** and Z.ai's **GLM 5.2** landed in the same performance tier, free. **DeepSeek V4 Pro Max** was reportedly trained for under $6 million — against the billions spent by Anthropic and its peers — and is priced to make the comparison hurt. Taken together, this is the clearest signal yet that open weights have stopped being the budget option. We're tracking the fallout across our [LLM Launches & Updates](https://speka.info/llm-updates/) coverage.

The four launches, and what each one actually changes

Kimi K3: the largest open-source model yet

Moonshot AI positions Kimi K3 as the biggest open-source model released so far, and reports it outperforming GPT-5.6 and Fable 5 on some benchmarks. Benchmark wins are the least interesting part of that sentence. The operational detail matters more: demand after launch was heavy enough that Moonshot had to **pause new sign-ups**. Capacity constraints are a strange kind of good news for an open-weight release — the weights are downloadable regardless, so the queue is a measure of how many people wanted the hosted convenience on day one.

Qwen 3.8: Alibaba's free frontier tier

Alibaba's Qwen 3.8 is positioned as comparably powerful to Kimi K3 and released free and open-source. Qwen has been the most consistently shipped open-weight family out of China, and 3.8 continues the pattern of releasing at the top of the lineup rather than holding the best checkpoint back for a paid API.

GLM 5.2: Z.ai plays the same hand

Z.ai's GLM 5.2 rounds out the trio, cited alongside Kimi K3 and Qwen 3.8 as a free open-source model now competitive with GPT and Claude. The stated logic is ecosystem-building: give the model away, become the default substrate developers build tooling and fine-tunes on, monetise later or elsewhere. When three labs run the same play in the same quarter, it stops looking like a strategy and starts looking like a market condition.

DeepSeek V4 Pro Max: the $6 million claim

DeepSeek claims V4 Pro Max was trained for just under **$6 million**, with performance presented as comparable to Opus 4.6 Max, GPT 5.4 and Gemini 3.1 Pro. Treat the figure with the usual caution — training-cost claims rarely specify whether they include failed runs, data pipeline work, salaries or amortised hardware, and DeepSeek has not published a full accounting. But even a generous restatement leaves a gap of two orders of magnitude against frontier US training budgets, and the architecture below explains where a lot of that gap comes from.


Two architectural bets doing the heavy lifting

256 experts, roughly eight awake at a time

Mixture-of-experts routing is standard at the frontier now, but DeepSeek pushed the granularity much further than most. Where OpenAI and others split a model into dozens of expert clusters, DeepSeek sliced its model into **256 hyper-specialised experts and routes each query to about eight of them**. The remaining ~248 stay dormant for that token. The effect is that total parameter count — which determines what the model knows — decouples almost entirely from the compute burned per response. You pay for a small model at inference time and get a large model's knowledge base.

Finer-grained routing is harder to train well; it makes load-balancing across experts and routing collapse genuine engineering risks. That DeepSeek got it working at 256-way granularity is the substantive claim here, more than the dollar figure attached to it.

Multi-head latent attention and the KV-cache problem

The second bet targets memory rather than compute. In a standard transformer, the key-value cache — the model's working memory of the conversation so far — grows with every token and has to be held in expensive GPU memory. Long context is therefore not just slow, it is structurally costly.

DeepSeek's **multi-head latent attention (MLA)** compresses that cache into a compact learned summary instead of storing every detail, with a reported reduction in short-term memory requirements of **more than 90%**. If that holds in production, it changes the unit economics of exactly the workloads enterprises care about: long documents, long agent traces, long multi-turn sessions.

The price list is the competitive weapon

Here is where the architecture cashes out. DeepSeek's top model is quoted at **$0.40 per million input tokens and $0.87 per million output tokens**. The comparable GPT tier is cited at $5 input and $30 output; the comparable Claude tier at $10 input and $50 output.

Run the arithmetic and the spread is roughly 10x to 50x depending on which model and which token direction you compare. For a chat product with modest volume, that difference is a rounding error. For a company running document pipelines, code review bots or anything agentic — where a single task can burn millions of tokens — it is the difference between a viable margin and no product.

Adoption has already moved

Nearly **60% of AI usage inside US companies is now attributed to Chinese models**, with Pinterest, Airbnb and Coinbase named among the users. Low cost and open weights are cited as the two drivers, and they compound: open weights mean you can self-host, avoid per-token billing entirely, keep data in your own environment, and fine-tune without asking permission. That combination answers procurement, finance and legal objections in one move.

The export controls that produced this

None of this happened in a vacuum. In **October 2022** the US barred Nvidia from exporting the H100 to China, blocked sales of the advanced machines needed to manufacture such chips, and added Huawei and dozens of other Chinese firms to a banned entity list. Nvidia responded with the **H20**, a deliberately weakened part built for the China market, and has continued lobbying to unlock advanced sales while Beijing pushes domestic production instead.

The controls reach further than American exports. The **Foreign Direct Product Rule**, a 1959 mechanism, lets Washington block the sale of any product made using US technology even when it is manufactured abroad. Because TSMC in Taiwan and ASML in the Netherlands depend on US technology, software and IP, that rule effectively gives the US a say over who buys the world's most advanced chips and lithography machines.

The workarounds started crude. Smuggling routes ran through Malaysia, Japan, Hong Kong and Singapore, including human couriers — one student reportedly paid $100 per chip for six carried out of Singapore. By all accounts that barely dented China's compute needs. What did move the needle was domestic capability: Huawei shipped its own **Ascend** chips in the 2023 Mate 60 Pro, restoring 5G in a phone many observers expected to be stuck on 4G, and **SMIC** found ways to produce advanced chips without leading-edge lithography — less efficient and more expensive than TSMC, but functional enough to reduce dependence on foreign suppliers.

The structural response is the **National Integrated Computing Power Network**, targeted for 2030, which treats compute as public infrastructure: large data centres sited where land and renewable energy are cheap, linked to cities by high-speed fibre, with startups, universities and researchers renting GPU capacity on demand rather than building their own facilities. Scarce hardware plus shared allocation is a strong incentive to optimise — which is roughly a description of how MLA and 256-way expert routing came to exist.

What this means if you're building

The practical takeaway is that model choice is now a portfolio decision rather than a vendor commitment. Route cheap, high-volume work to open weights; reserve premium closed models for the tasks where the quality delta is measurable and worth 30x. That only works if your tooling is model-agnostic — which is why the open agent-tooling ecosystem matters more than ever, from the [Claude Skills repos dominating GitHub trending](https://speka.info/blog/claude-skills-repos-take-over-github-trending) to open coding agents like those we covered in [OpenCode, Firecrawl and more](https://speka.info/blog/trending-github-repos-opencode-firecrawl-5-more).

The other half is evaluation discipline. Benchmark claims from any lab — Chinese or American — are marketing until you've tested them on your own workload. Build a small internal eval set before you migrate anything; provider-published recipe collections like [Anthropic's Claude Cookbook](https://speka.info/blog/claude-cookbook-anthropics-official-recipe-hub-goes-viral) are a reasonable starting point for patterns even if you end up running a different model behind them.

The frontier no longer has a price floor. Plan accordingly.

Frequently Asked Questions

Is Kimi K3 really the largest open-source AI model?

Moonshot AI's Kimi K3 is described as the largest open-source model released to date, and is reported to beat GPT-5.6 and Fable 5 on some benchmarks. Demand at launch was high enough that Moonshot paused new sign-ups.

How much did DeepSeek V4 Pro Max cost to train?

DeepSeek claims V4 Pro Max was trained for just under $6 million, compared with the billions spent by large US labs. The figure is the company's own claim and has not been independently audited, so treat it as a reported number rather than a verified one.

How much cheaper is the DeepSeek API than GPT or Claude?

DeepSeek's top model is quoted at $0.40 per million input tokens and $0.87 per million output tokens, against cited rates of $5 input / $30 output for the comparable GPT tier and $10 input / $50 output for Claude. That works out to roughly a 10–50x gap depending on the comparison.

What is multi-head latent attention (MLA)?

MLA is a DeepSeek-developed technique that compresses the key-value cache — the model's short-term memory of a conversation — into a compact summary rather than storing every detail. It reportedly cuts short-term memory requirements by more than 90%, making long conversations far cheaper to serve.

Why do 256 experts with only 8 activated make a model cheaper?

In a mixture-of-experts model, only the routed experts run for a given query. By slicing the model into 256 hyper-specialised experts and activating about eight per query, DeepSeek keeps a large model's knowledge while paying a small model's compute cost per response.

Are US companies actually using Chinese AI models?

Nearly 60% of AI usage inside US companies is now attributed to Chinese models, with Pinterest, Airbnb and Coinbase named among users. Low cost and open weights — which allow self-hosting and fine-tuning — are cited as the main drivers.

Didn't US chip export controls stop this?

The October 2022 rules blocked Nvidia H100 sales to China and, via the Foreign Direct Product Rule, gave Washington leverage over TSMC and ASML. China responded with domestic silicon from Huawei and SMIC, a state-backed shared compute network targeted for 2030, and heavy investment in efficiency techniques.