DEV Community

Cover image for Qwen 3.8 vs Kimi K3: China's Two Open-Weight Giants, Compared
Hassann
Hassann

Posted on • Originally published at apidog.com

Qwen 3.8 vs Kimi K3: China's Two Open-Weight Giants, Compared

July 2026 became a two-horse month for open-weight frontier models, and both contenders came from China. Moonshot AI released Kimi K3 on July 16, while Alibaba previewed Qwen 3.8-Max on July 19 and made it generally available in early August. Both are trillion-parameter mixture-of-experts (MoE) models, both work with Claude Code-style agent harnesses, and both target lower API costs than US frontier labs. The practical differences are in what you can download today, which inputs each model accepts, and what high-volume usage costs.

Try Apidog today

For the full Alibaba specification breakdown, see the Qwen 3.8-Max explainer.

Benchmark caveat: No independent, same-harness head-to-head benchmark of Qwen 3.8-Max and Kimi K3 was available as of August 3, 2026. Benchmark results below come from vendor-published tables and should be treated as indicative rather than conclusive.

The timeline: what shipped, and when

Release timing matters because “open weights” means different things for these models today.

Kimi K3

Kimi K3 launched on July 16, 2026. Moonshot promised open weights at launch and published them 11 days later, on July 27, as a 594 GB MXFP4 release on Hugging Face.

You can currently:

  • Download the weights
  • Apply additional quantization
  • Run the model on your own infrastructure
  • Evaluate a pinned model version without relying on a hosted endpoint

Qwen 3.8-Max

Qwen 3.8-Max was previewed on July 19 and reached general availability on Alibaba Cloud Model Studio in early August, according to the official Qwen release post.

Alibaba stated that weights would be released to Hugging Face and ModelScope “next week,” placing the expected date around August 10. As of August 3, 2026, those weights are not downloadable.

Implementation takeaway: If you need self-hosting this week, Kimi K3 is the only practical option. If hosted API access is acceptable, evaluate both.

Parameters: two trillion-scale MoE models

Both models use sparse MoE architectures. The total parameter count describes model size, while active parameters per token better reflect serving compute.

Spec Qwen 3.8-Max Kimi K3
Total parameters 2.4T 2.8T
Active parameters per token 95B 104B
Activation ratio ~4% ~3.7%
Architecture base Qwen 3.5 foundation, MoE MoE
Open-weight format Promised (~Aug 10) MXFP4, 594 GB on Hugging Face

Qwen 3.8-Max has 400B fewer total parameters and activates 9B fewer parameters per token than Kimi K3. That does not automatically make it more capable, but it can reduce serving cost when all other factors are equal.

For self-hosting, K3’s 594 GB MXFP4 download is the more useful number. This is multi-node deployment territory before accounting for long-context KV cache requirements. Qwen 3.8-Max will likely require similar infrastructure once its weights are available.

For most teams:

  • Use hosted endpoints for production evaluation.
  • Self-host only when compliance, research, model pinning, or custom inference infrastructure justify the operational cost.

Openness today: available weights vs a release promise

Kimi K3 is open in practice today. Its weights are public, so you can inspect the repository, review license terms, generate community quantizations, fine-tune the model, and host it independently.

Qwen 3.8-Max would be the first Qwen Max-tier model released with open weights. Previous Qwen Max models were API-only. However, until the Hugging Face or ModelScope repository is available, Qwen 3.8-Max remains an API model with an announced future weight release.

Decision rule:

  • Need downloadable weights now: choose Kimi K3.
  • Can wait until around August 10: re-check Qwen’s release status and compare license terms, quantization support, and community tooling.

Modality: Qwen supports images; K3 is text-only

This is the largest functional difference.

Qwen 3.8-Max supports text and image inputs, as shown in its official model configuration:

{
  "input_modalities": ["text", "image"]
}
Enter fullscreen mode Exit fullscreen mode

Alibaba’s launch materials also demonstrate 200-plus-page PDF understanding and 100-hour video comprehension through memory graphs. Treat those demonstrations as vendor examples, not as documented API input types. Image input, however, is available through the API.

Kimi K3 is a text model. If your workflow includes screenshots, scanned PDFs, UI interpretation, or computer-use tasks, you need to add a separate vision model and build the routing, data transfer, and result-merging logic yourself.

Choose based on input type

Workload Better fit
Text-only coding agents Either model
Repository analysis Either model
Screenshot debugging Qwen 3.8-Max
OCR and scanned documents Qwen 3.8-Max
UI automation or computer-use agents Qwen 3.8-Max
Self-hosted text-only inference Kimi K3

Context, output limits, and API compatibility

Qwen 3.8-Max provides:

  • Context window: 1,000,000 tokens
  • Maximum output: 65,536 tokens
  • Reasoning control: reasoning_effort
  • Reasoning levels: xhigh by default, plus medium and low
  • Thinking output: preserved by default
  • Billing: thinking and non-thinking output are billed at the same rate

Qwen is available through Alibaba Cloud Model Studio with regional base URLs for:

  • Beijing
  • Singapore
  • US-Virginia

It exposes both OpenAI-compatible and Anthropic-compatible API shapes. The Anthropic-compatible endpoint makes it straightforward to use with Claude Code-style tooling.

For example, configure a compatible harness with environment variables similar to:

export ANTHROPIC_BASE_URL="https://your-dashscope-anthropic-endpoint"
export ANTHROPIC_MODEL="qwen3.8-max"
export ANTHROPIC_API_KEY="your-api-key"
Enter fullscreen mode Exit fullscreen mode

Check the Model Studio models page for current regional availability and endpoint details.

If you test multiple regions, store each base URL in a separate environment. Tools such as Apidog let you switch environment variables without editing the request itself.

Kimi K3 also supports the Anthropic protocol and works with Claude Code-style harnesses. For current K3 endpoint configuration and limits, refer to the Kimi K3 explainer.

Pricing: Qwen is cheaper on list price

Published prices are per million tokens.

Rate Qwen 3.8-Max Kimi K3
Input $2.00 $3.00
Output $6.00 $15.00
Cache-hit input $0.20 $0.30
Tiering Flat across the full 1M context Per Moonshot’s published schedule

Qwen 3.8-Max charges $2 per million input tokens and $6 per million output tokens across its full 1M-token context window, according to the official Model Studio pricing page.

For Qwen:

  • Explicit cache creation costs 125% of the input rate.
  • Cache hits cost 10% of the input rate.
  • A 1M-token free quota is available in the Singapore region for 90 days.

Kimi K3 pricing is:

  • $3 per million input tokens
  • $15 per million output tokens
  • $0.30 per million cached input tokens

Estimate cost from your token mix

Use this basic formula before choosing a default model:

estimated_cost =
  (input_tokens / 1_000_000 * input_rate) +
  (output_tokens / 1_000_000 * output_rate) +
  (cached_input_tokens / 1_000_000 * cache_hit_rate)
Enter fullscreen mode Exit fullscreen mode

For output-heavy agent workloads, Qwen’s pricing advantage is substantial:

Qwen output: $6 / million tokens
Kimi K3 output: $15 / million tokens
Enter fullscreen mode Exit fullscreen mode

That makes Kimi K3 output 2.5x more expensive on list price.

However, do not estimate Qwen cost from visible final output alone. Its default reasoning_effort=xhigh generates thinking tokens, and those tokens are billed as output. Lower the reasoning effort during evaluation if your workload does not need maximum reasoning depth.

For worked examples, see the Qwen 3.8 pricing breakdown.

Benchmarks: compare methodology before scores

The key fact is simple: there is no independent benchmark table evaluating Qwen 3.8-Max and Kimi K3 under the same harness, prompt set, sampling configuration, and scaffolding.

What is available

  • Alibaba published Qwen 3.8-Max results against Claude Opus 4.8, Fable 5, GPT-5.6 Sol, and Qwen 3.7-Max.
  • Moonshot published a Kimi K3 launch benchmark table against a similar frontier-model set.
  • Both sets of results are vendor-run.

What is not available

  • Independent Qwen 3.8-Max vs Kimi K3 testing
  • Artificial Analysis results for Qwen 3.8-Max at the time of writing
  • Cross-vendor benchmark runs where both models used the same harness and setup

Alibaba’s published results, run largely through the Claude Code harness according to Alibaba’s own disclosure, include:

Benchmark Qwen 3.8-Max Claude Opus 4.8 Fable 5 GPT-5.6 Sol
Terminal Bench 2.1 86.6 84.6 84.6 88.8
SWE-bench Pro 67.7 69.2 80.0 64.6
PaperBench 93.0 80.3 88.8 90.5
GPQA Diamond 92.6 92.0 92.6 94.1
HLE 43.6 45.7 53.3 47.2

Alibaba’s own table includes losses as well as wins. Qwen 3.8-Max trails Fable 5 on SWE-bench Pro and trails every listed model on HLE. Its stronger reported areas include PaperBench, instruction following on IFBench, and multimodal evaluations.

Alibaba also reports results on in-house benchmarks such as QwenSWEBench and CoWorkBench. Those results cannot be directly compared with external vendor tables.

Moonshot’s K3 launch table showed K3 outperforming Claude Opus 4.8 across Moonshot’s selected benchmark rows, while trailing Fable 5 and GPT-5.6 Sol overall. See the Kimi K3 vs Claude Opus 4.8 comparison for those claims and caveats.

Practical conclusion: Treat both vendor benchmark tables as signals for which workloads to test, not as a purchasing decision. For more detail on Alibaba’s reported results, see the Qwen 3.8 benchmarks analysis.

Run your own head-to-head in Apidog

The most useful evaluation is a repeatable test against tasks from your own product or repository.

Both models provide OpenAI-compatible chat-completions endpoints, so you can use one request payload and switch only the target environment.

1. Create two environments

Create one environment for Qwen and one for Kimi K3.

Environment: qwen-3.8-max
BASE_URL=https://your-dashscope-endpoint
MODEL=qwen3.8-max
API_KEY=your-qwen-key
Enter fullscreen mode Exit fullscreen mode 
Environment: kimi-k3
BASE_URL=https://your-moonshot-endpoint
MODEL=your-k3-model-id
API_KEY=your-kimi-key
Enter fullscreen mode Exit fullscreen mode

2. Save one shared request

Use a production-like prompt rather than a benchmark riddle.

POST {{BASE_URL}}/v1/chat/completions
Authorization: Bearer {{API_KEY}}
Content-Type: application/json
Enter fullscreen mode Exit fullscreen mode 
{
  "model": "{{MODEL}}",
  "messages": [
    {
      "role": "system",
      "content": "You are a senior backend engineer. Return actionable implementation steps."
    },
    {
      "role": "user",
      "content": "Review this service design and identify likely failure modes: ..."
    }
  ],
  "stream": true
}
Enter fullscreen mode Exit fullscreen mode

3. Test the same workload on both models

For each prompt, record:

  • End-to-end latency
  • Time to first token
  • Output token count
  • Reasoning token behavior, if exposed
  • Schema validity
  • Task completion rate
  • Tool-call correctness
  • Actual invoice cost

4. Add repeatable assertions

Add checks for conditions that matter in your application:

- Response status is 200
- Response includes required JSON keys
- Latency stays below your threshold
- Output includes required domain terms
- Tool-call arguments match the expected schema
Enter fullscreen mode Exit fullscreen mode

Apidog can keep the request shared while you swap environments, inspect full response bodies, compare latency, and debug SSE streams. If your application renders thinking content, streaming inspection is especially useful.

Download Apidog and run 10 to 20 representative prompts through both endpoints before committing to either model.

The scorecard

Axis Winner today Caveat
Total/active parameters Qwen 3.8-Max Leaner at 2.4T/95B vs 2.8T/104B; mainly relevant to serving cost
Open weights today Kimi K3 Live on Hugging Face as a 594 GB MXFP4 release
Modality Qwen 3.8-Max Supports text and image input
Context window Qwen 3.8-Max 1M flat tier and 65,536-token maximum output
Price Qwen 3.8-Max $2/$6 vs $3/$15, but xhigh reasoning increases billed output
Benchmarks No call Both tables are vendor-run
Harness ecosystem Tie Both support Claude Code-style Anthropic-protocol integrations
Track record on open-weight delivery Kimi K3 Weights shipped 11 days after launch; Qwen’s release was still pending

Which model should you use?

Pick Kimi K3 if

  • You need downloadable weights immediately.
  • Your compliance requirements require a model you can pin, audit, and self-host.
  • Your workload is text-only.
  • Your input volume is high enough that cache-hit pricing dominates your costs.

Pick Qwen 3.8-Max if

  • Your workload includes images, screenshots, or scanned documents.
  • You run output-heavy agents for code generation or multi-step task execution.
  • You need a 1M-token context window without tiered pricing.
  • You want lower list prices for both input and output tokens.

Wait if open weights are the deciding factor

If Qwen’s weights arrive around August 10 as promised, the openness comparison changes. At that point, compare:

  1. License terms
  2. Available quantizations
  3. Inference stack support
  4. Hardware requirements
  5. Your own task-level evaluation results

The actionable conclusion is not to trust either vendor’s benchmark table alone. Both models are credible, lower-cost, harness-compatible frontier options. Route the same real prompts through both endpoints, measure quality, latency, and cost, then choose based on your production workload.

FAQ

Is Kimi K3 more open than Qwen 3.8-Max?

As of August 3, 2026, yes. Kimi K3 weights have been available on Hugging Face since July 27 as a 594 GB MXFP4 release. Qwen 3.8-Max weights were promised for around August 10 but were not downloadable yet.

If Alibaba releases the weights, both models become open-weight options. The next comparison points will be license terms, quantization quality, and community support. Until then, only K3 can be self-hosted. See the local K3 guide for deployment considerations.

Which model is better at coding?

There is no honest definitive answer yet.

Both vendors publish strong coding and agentic benchmark results, but they use their own evaluation setups. No independent same-harness comparison exists. Alibaba’s own results show Qwen 3.8-Max losing to Fable 5 on SWE-bench Pro, which reinforces why a single vendor table should not determine your choice.

Run both models against tasks from your own repository:

  • Bug fixes
  • Test generation
  • Refactoring
  • Code review
  • Tool invocation
  • Multi-file implementation tasks

Can I use both models in Claude Code?

Yes. Both models expose Anthropic-compatible endpoints.

For Qwen 3.8-Max, configure ANTHROPIC_BASE_URL with the DashScope Anthropic endpoint and set:

export ANTHROPIC_MODEL="qwen3.8-max"
Enter fullscreen mode Exit fullscreen mode

Kimi K3 supports the same general configuration pattern through Moonshot’s endpoint. This shared protocol compatibility makes side-by-side testing inexpensive to set up.

Which model is cheaper for a typical agent workload?

On list price, Qwen 3.8-Max is cheaper:

Qwen 3.8-Max: $2 input / $6 output per million tokens
Kimi K3:       $3 input / $15 output per million tokens
Enter fullscreen mode Exit fullscreen mode

Agent loops often generate many output tokens, which makes Qwen’s $6 output rate attractive compared with K3’s $15 rate.

Two qualifiers matter:

  1. K3’s $0.30 cache-hit rate can keep input-heavy, heavily cached workloads competitive.
  2. Qwen’s default xhigh reasoning setting bills thinking tokens as output, so real costs can exceed simple visible-output estimates.

Measure your actual input, cached-input, output, and reasoning token mix before making a cost decision.

Top comments (0)