Qwen3.8 Max and Kimi K3: two Chinese flagships in two weeks

The short version

Two Chinese flagship models shipped in two weeks. Kimi K3 from Moonshot AI landed on 27 July with 2.8 trillion parameters and the largest open weights published to date. Qwen3.8 Max from Alibaba followed on 3 August with 2.4 trillion parameters and native multimodality. Both carry a one-million-token context. Both are now in the NetRoom catalogue.

Condensed: Qwen is nearly three times cheaper on output tokens and accepts video input. Kimi is stronger on frontend work and ships weights you can download and self-host.

Qwen3.8 Max: built for autonomy

Alibaba previewed the model at the WAIC conference in Shanghai on 19 July and made it generally available on 3 August. Inside is a sparse MoE architecture holding 2.4 trillion parameters, roughly 95 billion of which activate per request. For scale, the previous Qwen3.6 Max Preview (20 April) was a third of that size in total.

Context runs to 1,000,000 tokens with output up to 131,072. The model takes text, images and video as input.

The release is built around agentic work. In Alibaba's internal runs the model sustained autonomous software development for more than ten days straight — writing code, running its own tests, reading logs and iterating without human input. In another test it reproduced a machine learning research paper from scratch: 33 rounds of GPU training over roughly 125 hours, 7,600 lines of code, followed by 18 proposed improvements, some of which outperformed the original method.

Benchmark movement against 3.6 is substantial:

• Terminal-Bench 2.1 — 86.6 versus 51.5 for Qwen3.6 on version 2.0
• SWE-bench Pro — 67.7 versus 49.5

On Arena.AI it immediately became the top-ranked Chinese model for text tasks and second globally for vision, behind only a Claude Fable 5 variant. Alibaba said open weights would follow the release; they were not available at the time of writing.

Kimi K3: the largest open weights so far

Moonshot shipped K3 on 27 July and released the weights for unrestricted download straight away. At 2.8 trillion parameters it is the largest open-weight model published. The previous K2.6 (also 20 April) ran 1 trillion parameters with 32 billion active per token and competed differently — through an Agent Swarm system scaling to 300 specialised sub-agents and runs of up to 4,000 coordinated steps.

K3 carries the same one-million-token context. It is tuned for navigating large repositories, tool use, debugging, and iterating against logs, tests and screenshots.

Benchmark numbers:

• Frontend Code Arena — first place at 1679, ahead of Claude Fable 5 in blind developer testing
• GDPval-AA v2 — 1687, third behind Claude Fable 5 Max (1815) and GPT-5.6 Sol Max (1747.8), ahead of Claude Opus 4.8 (1600)
• AA-Briefcase — 1527, second place, beating GPT-5.6 Sol Max (1495) and trailing Fable 5 Max (1587)

Two details worth noting. Demand after launch was heavy enough that Moonshot temporarily restricted access. And at comparable quality, K3 is meaningfully easier to run than its rivals — which matters a great deal for open weights, since somebody has to actually stand the thing up.

Pricing is where they split

This is the sharpest difference between the two.

• Qwen3.8 Max — $2.00 per million input tokens, $6.00 per million output.
• Kimi K3 — around $2.90 per million input and $14.00 per million output. Direct from Moonshot, input is $3.00, dropping to roughly $0.30 on a cache hit.

That puts Kimi 2.3x above Qwen on output. For chat and short answers the gap barely registers; for agentic runs generating thousands of lines of code it becomes the dominant line item. K3 also costs about five times what K2.6 did — Moonshot has clearly moved the model into a different tier.

Caching offsets part of this. If you run long conversations with a stable system prompt and the same codebase in context, the real bill comes in well under the headline rate.

Which to reach for

Qwen3.8 Max when cost per volume matters and you need image or video understanding. Long agentic runs, batch document analysis, whole-project work.

Kimi K3 for frontend and UI work — that Arena result is not an accident — for log- and screenshot-driven debugging, or when open weights are a requirement rather than a preference: the ability to download the model and not depend on a vendor changing terms or pricing.

Both are live on NetRoom, running in the browser with per-token billing. Model pages: Qwen3.8 Max and Kimi K3.

The wider read

Two weeks, two releases, both sitting near the top of global leaderboards, both cheaper than Western flagships. A year ago the framing around Chinese models was catching up. Right now Kimi K3 holds first place in Frontend Code Arena and Qwen3.8 Max sits second worldwide on vision. Whatever gap remains is measured in leaderboard positions, not generations.