Qwen3.8 vs Kimi K3: Two Open-Source 2T+ Models Land Within 8 Days of Each Other

Between July 12 and July 19, 2026, two Chinese labs released open-weight language models that cross the 2-trillion-parameter threshold for the first time: Moonshot AI's Kimi K3 (2.8T total parameters, 1M-token context, native vision) and Alibaba's Qwen3.8-Max-Preview (2.4T total parameters, positioned as the open-weights answer to GPT-5.5 / Claude Opus 4.7 / Gemini 3.1 Pro). This article is an API-side, not a benchmark-side, comparison: what you actually get when you wire each model into a real production endpoint, what it costs per million tokens, what tooling it supports, and which one you should pick for which use case.

Both models are reachable today through managed APIs that speak the OpenAI and Anthropic chat-completion protocols, so no PyTorch cluster is required to evaluate them. The pricing tables, context windows, and tool-calling features below were captured directly from each vendor's official documentation on July 20, 2026: help.aliyun.com/zh/model-studio/text-generation-model for Qwen3.8-Max-Preview, and platform.kimi.com/docs for Kimi K3.

Qwen3.8-Max-Preview at a Glance

Qwen3.8-Max-Preview is the top-tier Qwen 3.8 endpoint on Alibaba Cloud Bailian (Model Studio). Alibaba's own positioning table places it against GPT-5.5, Claude Opus 4.7, and Gemini 3.1 Pro on capability, and it sits one tier above the more generally available qwen3.7-max model. As of July 19, 2026, the model card on Bailian reads "Token Plan only" — you cannot buy it on pure pay-as-you-go; you must subscribe to a Token Plan that bundles Qwen3.8-Max-Preview with a quota of tokens per month.

Key technical points from the official documentation:

• Parameter count: 2.4T total parameters (Mixture-of-Experts, per the @Alibaba_Qwen launch post on July 19, 2026).
• Context window: long-context support comparable to the 1M-token class, with the 3.8-Max-Preview specifically called out for repository-scale code tasks.
• Protocols: OpenAI-compatible, Anthropic-compatible, and the native DashScope protocol — all three are exposed through the same base_url.
• Regions: Beijing (cn-beijing), Singapore, Tokyo, Frankfurt, Virginia — five regional endpoints at launch.
• Tooling: full function calling / tool_choice support, structured output (JSON Schema), and visual reasoning if you route Qwen-VL-class endpoints for image input.
• Access gate: Token Plan subscription required (Bailian will redirect you to the subscription page from the model card).

Because the model is gated behind a Token Plan, the pricing structure is different from Kimi K3's pay-as-you-go rates. You buy a token quota and the model spend draws against that quota; overshoot the quota and you either upgrade the plan or pause until the next billing cycle. The 3.8 preview tier is positioned for the strongest reasoning workload, not for cheap bulk generation — Alibaba's own "balanced" pick for coding agents is qwen3.7-plus.

Kimi K3 at a Glance

Kimi K3 is Moonshot AI's flagship, and it is reachable through a standard pay-as-you-go API at platform.kimi.com. The model documentation lists the same headline features that made Kimi K2.6/K2.7 popular with Chinese developer teams, scaled up: 2.8T total parameters, 1,048,576-token context window, native image and video understanding, automatic context caching, JSON mode, structured output, and tool-choice controls.

Pricing captured July 20, 2026 from the official Kimi docs (CNY per million tokens):

• Cache-hit input: ¥2 / M tokens.
• Cache-miss input: ¥20 / M tokens.
• Output: ¥100 / M tokens.
• Automatic caching: the platform automatically caches repeated prefixes and bills them at the cache-hit rate, which is the single biggest cost lever for long-running agents.

Kimi K3 also ships with tool_choice, dynamic tool loading, JSON Schema structured output, and reasoning effort controls. A new-user ¥15 coupon does not apply to K3 (it is excluded from the coupon's eligible-model list), so budgeting from the first production test matters. Two cheaper siblings — Kimi K2.7 Code and Kimi K2.6 — share the same 256K context window and the same automatic caching, but cost roughly a tenth of K3 on output. Most teams that do not need the 1M context window end up routing 90% of traffic to K2.7 or K2.6 anyway.

Qwen3.8-Max-Preview vs Kimi K3: Verified Pricing & Specs

Side-by-side, the two flagship models line up like this:

A pragmatic translation: Kimi K3 is the easier model to budget and to integrate (OpenAI SDK, automatic caching, a single pay-as-you-go rate card). Qwen3.8-Max-Preview is the more internationally-distributed option (5 regions, dual OpenAI/Anthropic compatibility) and the only one of the two that has any official English-language documentation surface that you can hand to a global team.

Calling Qwen3.8-Max-Preview via the OpenAI-Compatible API

The Bailian model card exposes a one-click "copy SDK config" snippet. Once you have provisioned a workspace on the Beijing region and subscribed to a Token Plan, the endpoint is:

curl https://<WorkspaceId>.cn-beijing.maas.aliyuncs.com/compatible-mode/v1/chat/completions   -H "Authorization: Bearer ***   -H "Content-Type: application/json"   -d '{
    "model": "qwen3.8-max-preview",
    "messages": [
      {"role": "system", "content": "You are a precise coding assistant."},
      {"role": "user", "content": "Refactor this function to use async/await."}
    ],
    "max_tokens": 4096,
    "temperature": 0.2
  }'

Python clients work the same way — point the openai SDK at the base_url shown above and pass qwen3.8-max-preview as the model id. The Anthropic-compatible base URL is also available on the same workspace, which means Claude Code and any other Anthropic-protocol tooling can drop Qwen3.8-Max-Preview in as a model backend without a wrapper.

Calling Kimi K3 via the OpenAI-Compatible API

Kimi's platform exposes an OpenAI-compatible base URL. The current endpoint is:

curl https://api.moonshot.cn/v1/chat/completions   -H "Authorization: Bearer YOUR_K...EY"   -H "Content-Type: application/json"   -d '{
    "model": "kimi-k3",
    "messages": [
      {"role": "system", "content": "You are a precise coding assistant."},
      {"role": "user", "content": "Refactor this function to use async/await."}
    ],
    "max_tokens": 4096,
    "temperature": 0.2
  }'

For long-context workloads, Kimi's automatic context caching is the key cost lever: any prompt prefix you reuse will be billed at ¥2/M (cache hit) instead of ¥20/M (cache miss). For a 200K-token conversation that repeats an 180K-token system message on every call, the cache-hit rate cuts effective input cost roughly tenfold.

When to Pick Qwen3.8-Max-Preview

Qwen3.8-Max-Preview is the right call in three specific situations:

• You need a non-China region for compliance or latency reasons. The five-region rollout (Singapore, Tokyo, Frankfurt, Virginia, plus Beijing) means you can keep data residency outside of mainland China — important if you are shipping a product to European or US customers and cannot route through a Chinese-hosted endpoint.
• Your stack already speaks Anthropic's Messages API. Claude Code, Cursor's Claude-mode agent, and a growing number of agentic frameworks call the Anthropic protocol natively. The Bailian Anthropic-compatible base URL lets you keep that stack and swap in Qwen3.8-Max-Preview without writing a translation layer.
• You want to evaluate the open-weights weights directly. Alibaba publishes the Qwen3.8 weights on Hugging Face and ModelScope under a permissive license, so if the managed API is the wrong cost shape for your workload you can self-host on your own GPU fleet.

The tradeoff is that the 3.8-Max-Preview tier is gated behind a Token Plan. If your usage is highly variable (bursty, spiky, or unpredictable), committing to a plan quota is operationally annoying. For steady, predictable workloads that you can size upfront, the plan pricing is competitive with the closed-model tier.

When to Pick Kimi K3

Kimi K3 is the right call in three different situations:

• You need native vision in the same model id. Kimi K3 accepts images and videos directly on the chat-completions endpoint. With Qwen3.8-Max-Preview you have to switch to a separate Qwen-VL model id to handle images, which means routing logic in your client.
• Your agent runs long-running loops with cached system prompts. Automatic context caching makes Kimi K3 the cheapest 1M-context API in 2026 for workloads that repeat large prefixes (code repos, RAG context, system instructions). The ¥2/M cache-hit rate is roughly an order of magnitude cheaper than the cache-miss rate, and there is no manual cache-management code to write.
• You want one model that handles tool calls and structured output without a separate provider switch. K3's tool_choice, JSON mode, and JSON Schema support are all on the same endpoint with the same SDK.

The tradeoff is regional: Kimi K3 is currently a mainland-China service. If you need a US/EU data residency story, Kimi K3 is not the model to pick.

Cost Modeling: A 200K-Agent Workload

Concretely, what does each model cost for a long-running agent that processes ~200K tokens per turn, repeats a 150K-token system prompt, and produces ~5K tokens of output? Assumptions: 1,000 turns/day, automatic cache hits on Kimi K3, Token Plan quota on Qwen3.8-Max-Preview.

• Kimi K3 (cache hit): (150K × ¥2 + 50K × ¥20 + 5K × ¥100) / 1M × 1000 = ¥(0.30 + 1.00 + 0.50) × 1000 = ¥1,800/day ≈ ¥54,000/month.
• Kimi K3 (cache miss): (200K × ¥20 + 5K × ¥100) / 1M × 1000 = ¥(4.00 + 0.50) × 1000 = ¥4,500/day ≈ ¥135,000/month.
• Qwen3.8-Max-Preview: cost is plan-dependent and not directly comparable to the Kimi per-token rates. A Token Plan at the top tier gives a multi-million-token monthly quota that, in practice, maps to ¥80,000-¥150,000/month for the same workload — within the same order of magnitude as Kimi K3 cache-hit, but with no automatic caching discount.

If your agent is truly long-running and the cache-hit rate is high, Kimi K3 wins on raw cost. If you can only hit 30-40% cache hits (short conversations, varied system prompts), the gap narrows and the Qwen3.8-Max-Preview Token Plan starts to look more attractive, especially with the international regions factored in.

Why Both Matter: The 2T+ Open-Weights Inflection Point

The bigger story is structural. Until July 2026, the 2T-parameter threshold was a closed-weight club — GPT-5.5, Claude Opus 4.7, Gemini 3.1 Pro, Grok 4. Within an eight-day window, two Chinese labs put open-weight 2T+ models into production API endpoints. The weights are downloadable, the code is open, and the managed endpoints follow OpenAI and Anthropic protocols that any developer already knows.

For buyers, this changes the negotiation. When you can self-host Qwen3.8-Max or Kimi K3 on your own GPU fleet at $1-3/hour per H100 node, the managed-API price is the ceiling, not the floor. For open-weights advocates, the release confirms the trend line that the previous Kimi K2 generation already started: the open-weight camp is now within a generation of the closed-weight flagship, and the closed-weight pricing premium will have to keep shrinking to stay defensible.

Verdict

Both Qwen3.8-Max-Preview and Kimi K3 are usable, real, 2T+ flagship-class APIs as of July 21, 2026 — and both are reachable with off-the-shelf OpenAI or Anthropic SDKs. The choice is not "which is better" but "which fits your stack":

• Pick Qwen3.8-Max-Preview for international regions, Anthropic-protocol tooling, and self-hosting flexibility.
• Pick Kimi K3 for native vision, automatic caching, mainland-China latency, and a single OpenAI-compatible endpoint that does everything.

If you are evaluating AI agent infrastructure in 2026 and you are not testing both endpoints side by side, you are leaving the open-weights pricing pressure on the table. Start with a 100-turn benchmark on each, route traffic to whichever fits your cost-and-region profile, and keep the other endpoint warm as a fallback. The closed-weight 2T+ models are no longer the only game in town.

Sources

• Alibaba Cloud, Text Generation Models — Model Studio, 2026-07-19: help.aliyun.com/zh/model-studio/text-generation-model
• Alibaba Cloud, Model List, 2026-07-19: help.aliyun.com/zh/model-studio/models
• Alibaba Cloud, Token Plan overview, 2026-07-19: bailian.console.aliyun.com — Token Plan
• Moonshot AI, Kimi Platform Documentation, 2026-07-20: platform.kimi.com/docs
• Moonshot AI, Kimi Pricing, 2026-07-20: platform.kimi.com
• @Alibaba_Qwen, Qwen3.8 launch announcement, 2026-07-19: x.com/Alibaba_Qwen/status/2078754377473601787
• @AYi_AInotes, Kimi K3 coding benchmark, 2026-07-19: x.com/AYi_AInotes/status/2077981025905316253