Qwen 3.8 Max vs Kimi K3：本格的なコーディング作業に対応できるのはどちらか？

Qwen 3.8 Max vs Kimi K3 at a Glance

Alibaba’s own Token Plan documentation says the Qwen Preview may be continually updated and later removed or replaced. Moonshot’s Kimi K3 announcement and API guide disclose substantially more about the model and its current interface.

Why This Is Not an Equal Comparison Yet

Qwen 3.8 Max Is Still a Moving Preview

Alibaba describes Qwen 3.8 Max as a 2.4T model, but the hosted model available today is specifically qwen3.8-max-preview. That suffix matters. Alibaba states that its capabilities may be upgraded throughout the Preview and that the endpoint may later be removed or replaced by a production version.

This makes every current benchmark a dated snapshot. A useful Qwen result should record the test date, endpoint, reasoning level, agent shell, tool permissions, and raw output. Without those details, two people may believe they tested the same model while actually receiving different Preview behavior.

The access terms also limit what developers can conclude from the low launch price. Alibaba’s Personal Token Plan currently offers Qwen 3.8 Max through a Credits-based subscription, but its official terms restrict the key to supported interactive coding and agent tools. Automation scripts, custom application backends, and non-interactive batch calls are prohibited on that plan.

So Qwen is callable. It is not yet a normal pay-as-you-go backend for a customer-facing app.

Kimi K3 Has Completed Its Open-Weight Release

Kimi K3 is now available in two forms: a conventional hosted API and a downloadable checkpoint. Moonshot AI has published the full weights in the official `moonshotai/Kimi-K3` repository, together with a model card, deployment guidance, a custom license, and the K3 technical report.

The release adds details that were unavailable when this comparison was first published. K3 has 104B activated parameters per token, uses MXFP4 weights with MXFP8 activations, and is currently documented for vLLM, SGLang, and TokenSpeed. The repository is about 1.56 TB across 96 safetensors shards.
That is a real open-weight release. It is not the same as an unrestricted MIT release or an easy local install. The Kimi K3 License adds conditions for large Model-as-a-Service businesses and very large commercial products, while Moonshot recommends supernode configurations with 64 or more accelerators for deployment.
In plain English: Kimi has removed the ownership uncertainty. It has not removed the infrastructure cost.

Coding Comparison: What the 269-File Test Actually Shows

The best current head-to-head evidence is a matched StackPerf architecture test. Both models received the same read-only snapshot of two unfamiliar software projects containing 269 files. They used OpenCode 1.17.13, identical task instructions, the same permissions, a 60-minute limit, and a 65,536-token completion ceiling. Editing, web access, plugins, and subagents were disabled.

The final reports were anonymized before scoring, then checked separately for unsupported claims.

Where Kimi K3 Did Better

Kimi’s report modeled the editing lifecycle more completely. It tracked revisions, superseded takes, retry attempts, fallback history, artifact hashes, invalidation, editorial selection, and trimming. That made its proposed contract better suited to a system in which one scene may be generated, rejected, revised, and generated again without losing its history.

Kimi also finished sooner and used fewer tokens on the tested route. That result is encouraging for long coding tasks, but it includes the serving system around the model. Kimi ran through a Kimi subscription endpoint, while Qwen ran through Alibaba’s international Token Plan endpoint. Provider capacity, caching, and launch-period traffic all affected the observed result.

My reading is that Kimi won this task because it understood changing state more completely, not because it wrote prettier code or proved itself better at every software task.

Where Qwen 3.8 Max Did Better

Qwen drew a cleaner boundary between the two systems and created a stronger replay record. Its proposed metadata included the model, seed, prompt, references, provider request ID, media location, duration, and cost. Those fields matter when a team needs to reproduce an output or investigate why the same workflow changed after a provider update.

Its tool behavior was also disciplined: 44 tool calls, all successful. Qwen produced the longer report with fewer requests and received a higher tool-use score in the blind review.

The cost of that cleaner architecture was weaker lifecycle modeling. Its contract did not fully cover scene order, revisions, supersession, retry history, or take invalidation. Clean boundaries are valuable. They are not enough when the product’s state keeps changing.

What This Test Cannot Prove

One architecture-analysis run cannot tell us which model produces better frontend code, fixes more real bugs, passes more tests, or needs fewer human corrections over a month of development. It also cannot isolate model intelligence from endpoint performance.

The most important warning is hidden in the factual review: both models cited real repository locations, yet both produced seven groups of claims that went beyond what the code proved. A report can contain hundreds of valid citations and still reach an unsafe conclusion.

That is the real takeaway. Use either model as an engineering assistant. Do not treat either one as its own reviewer.

How Developers Should Test Qwen 3.8 Max Against Kimi K3

Keep the agent shell and permissions identical. Set a fixed time limit. Record the full model ID and date, especially for Qwen’s moving Preview. Then capture task completion, wall-clock time, input and output tokens, cache hits, failed tool calls, retries, human interventions, and final tests passed.

Do not score an answer because it “looks thorough.” Check whether the patch works and whether the model’s claims survive review.

For high-value architecture decisions, there is another useful pattern: run both models independently, hide their identities during review, and compare their disagreements. The 269-file test produced its strongest design only after combining Qwen’s system boundaries with Kimi’s lifecycle model. Sometimes the right answer to Qwen versus Kimi is both, followed by verification.

Qwen 3.8 Max vs Kimi K3 Pricing and Cost

Kimi K3 Has Predictable Token Pricing

Kimi K3 has an ordinary token-based price. The current GPT Proto Kimi K3 API page lists $2.70 per 1M input tokens and $13.50 per 1M output tokens, 10% below Moonshot’s current $3/$15 list rate.

That makes a basic cost estimate possible before a test begins:

estimated cost = (input tokens ÷ 1,000,000 × $2.70) + (output tokens ÷ 1,000,000 × $13.50)

A run with 200,000 input tokens and 20,000 output tokens would cost about $0.81 at the current GPT Proto rate: $0.54 for input and $0.27 for output. Retries, additional turns, tool results, and retained reasoning history would increase the total.

Predictable does not mean automatically cheap. Kimi always reasons, and reasoning content counts toward token use. Long agent histories can therefore become expensive even when the visible final answer is short.

Qwen 3.8 Max Is Cheaper to Experiment With, but Harder to Budget

Qwen 3.8 Max Preview currently uses subscriptions and Credits rather than a disclosed per-million-token price. Alibaba’s Personal Token Plan lists promotional monthly prices of CNY 39, CNY 139, and CNY 499 for its three tiers. During the Preview, Qwen calls may consume as little as 10% of the standard Credits rate, with an additional time-limited night discount documented for calls between 22:00 and 08:00 UTC+8.

That is attractive for interactive testing. It is not a clean cost comparison.

Credits can vary with token use, caching, reasoning, and tools, so a CNY subscription cannot honestly be converted into a stable $X per 1M tokens figure from the information currently published. The Personal Plan restriction also means that its promotional economics do not represent the cost of running Qwen behind your own application.

Qwen wins on low-cost Preview access. Kimi wins on production cost visibility.

Open Weights and Self-Hosting

Kimi K3 now has a concrete advantage over Qwen 3.8 Max: its exact public checkpoint can be downloaded today. Moonshot AI has released the full 2.8T-parameter weights, model card, license, and technical report. Alibaba has announced open weights for Qwen 3.8 Max, but the exact hosted Max checkpoint, release date, license, and deployment requirements remain unconfirmed.
The word “open” still needs qualification. Kimi K3 is open-weight under the custom Kimi K3 License, not MIT. The license broadly permits use, modification, deployment, fine-tuning, and redistribution, but adds a separate-agreement requirement for Model-as-a-Service businesses above the stated revenue threshold and an attribution requirement for very large commercial products.
Hardware is the other boundary. The official K3 repository is about 1.56 TB, and Moonshot recommends supernode deployments with at least 64 accelerators. That makes self-hosting plausible for well-funded infrastructure teams, not a substitute for an API key on a developer workstation.
Choose the hosted Kimi K3 API when you want predictable token billing and no cluster operations. Choose the weights when data control, custom inference, fine-tuning, or deployment ownership is worth operating the hardware and reviewing the license. For Qwen 3.8 Max, wait until Alibaba publishes the actual checkpoint before making the same calculation.

Multimodal Inputs, Reasoning, and Agent Behavior

Qwen’s documented vision support makes it relevant for screenshot-based debugging. Kimi goes further by accepting video, which is useful when the input is a screen recording, animation reference, or product demo that would be difficult to describe frame by frame.

Kimi’s richer documented interface also adds integration work. Its Preserved Thinking behavior means a multi-turn application should return the complete assistant message, including reasoning content and tool calls, rather than keeping only the visible answer. That history occupies context and is billed. The 1M-token window is large, but it is not free storage.

Both models default to their highest reasoning setting. For evaluation, keep those settings consistent. For production, test lower effort on simpler tasks. A model that solves 99% of requests at lower effort may be cheaper and faster than one left at maximum reasoning for every autocomplete, classification, or short transformation.

Which Model Should You Choose?

For an application going live this week, I would choose Kimi K3. It has enough published information to estimate cost, define integration behavior, and repeat a test against a stable model name.

For an internal coding experiment, I would not ignore Qwen. Its Preview handled a long repository analysis with zero failed tool calls and showed better architectural boundaries than Kimi in the matched test. That is a serious capability signal. It is not yet a production contract.

How to Try Kimi K3 Through GPT Proto

Qwen 3.8 Max is not yet available on GPT Proto, so the current integration example uses Kimi K3 only. You can call it through GPT Proto’s OpenAI-compatible Chat Completions endpoint with the kimi-k3 model string:

curl https://gptproto.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: $GPTPROTO_API_KEY" \
  -d '{
    "model": "kimi-k3",
    "messages": [
      {
        "role": "user",
        "content": "Review this migration plan. Identify unsupported assumptions, missing rollback steps, and the tests required before production."
      }
    ]
  }'

The endpoint and authorization structure follow the GPT Proto API quickstart. For multi-turn Kimi workflows, preserve the complete assistant message returned by the API, including reasoning and tool-call fields, rather than storing only the final visible text.

You can try the Kimi K3 API, browse 200+ AI models, or use the GPT Proto unified AI API to compare Kimi with other text, image, video, and audio models. When Qwen 3.8 Max is added, the useful test will be the same task, prompt, tool permissions, and scoring method on both model endpoints.

Final Verdict

Qwen 3.8 Max and Kimi K3 appear close enough in coding capability that one small benchmark should not decide the next year of your architecture. Qwen’s cleaner system boundaries and flawless tool execution deserve attention. Kimi’s stronger lifecycle reasoning and three-point lead deserve attention too.

The product decision is now clearer. Kimi K3 combines a standard API, public token pricing, a documented 1M context window, multimodal input, and a released open-weight checkpoint. Qwen 3.8 Max remains a changeable Preview with Credits-based access, no ordinary production price, and no public weights.
Kimi K3 is the better model to deploy today, whether you begin with its hosted API or have the infrastructure to manage the checkpoint yourself. Qwen 3.8 Max may still become the stronger model, but developers should wait for its production endpoint, ordinary pricing, model card, license, and actual weight release before making that bet.