Kimi K3.1 Pro vs Qwen3.8-Max: Same $2 Input, $4 Apart on Output

The two models shipped 81 days apart, each as its maker's flagship: Kimi K3.1 Pro is Moonshot's post-trained successor to Kimi K3, and Qwen3.8-Max was Alibaba's strongest Qwen model in August. On spec sheets both are 2-trillion-class MoE models with 1M context, and both charge $2 per million input tokens via API; the dividing lines are openness and output price — Kimi's flagship itself ships open weights, Qwen's flagship API is closed and hosted, and its output costs $4 less.

The timing matters: Alibaba put the Qwen4-Max Preview live on Nov 11 for a window ending 24:00 on Nov 25, with no parameter count, benchmarks or pay-as-you-go pricing published, and the model ID goes offline afterwards; Qwen3.8-Max remains the flagship Alibaba sells by the token today, while Kimi K3.1 Pro's full weights have been public since Nov 4.

1. Publisher and dates: Moonshot AI; Kimi K3.1 Pro released Oct 23, 2026; full weights published on Hugging Face Nov 4, 2026
2. Publisher and dates: Alibaba; Qwen3.8-Max generally available Aug 3, 2026; upgraded snapshot qwen3.8-max-0902 on Sep 2, with qwen3.8-max routing to it since Sep 5
3. Method: specs and prices from each vendor's official pages (checked Nov 12, 2026); benchmarks split into vendor-reported figures and Artificial Analysis independent measurements, all on Index v5.0.1; vendor tables use different harnesses, so their scores are not a ranking under one set of conditions
4. Context: Alibaba put the Qwen4-Max Preview live on Nov 11 for a window ending 24:00 on Nov 25, with no parameter count, benchmarks or pay-as-you-go pricing published

On specs both are 2-trillion-class MoE models with 1M context, with Kimi carrying 0.4 trillion more total parameters; on price the input rates match while output and cache rates diverge; on openness the dividing line is that Kimi's flagship itself ships open weights, while Qwen's flagship API is closed and hosted, with a separate text-only open checkpoint.

1. Scale: Kimi has 2.8 trillion total parameters with 104 billion active per token, 16 of 896 experts routed per token plus 2 shared experts; Qwen has 2.4 trillion total with 95 billion active per token (per its release post)
2. Context and output caps: Kimi has a 1,048,576-token context window, with no published per-response output cap; Qwen has a 1,000,000-token context window (max input 991,808), a 131,072-token output cap and a 262,144-token chain-of-thought cap (Model Studio documentation)
3. Input modalities: Kimi takes text and image input and returns text; Qwen takes text, image and video input and returns text; both support function calling and structured outputs, and Qwen additionally supports web search and context caching (Model Studio documentation)
4. API pricing (per 1M tokens): Kimi at $2.00 input, $0.20 cached input and $10.00 output, with a Highspeed tier at $4.00 input and $20.00 output; Qwen's international region at $2.00 input, $6.00 output, $0.25 implicit cached input and $0.17 explicit cache read, and its Beijing region at $1.65 input and $4.951 output
5. Output price gap: Qwen's output costs $4.00 less than Kimi's, 40% lower; input prices match, and Kimi's cached input is $0.05 cheaper
6. Open weights: Kimi's full weights were published on Hugging Face on Nov 4 under the Kimi K3.1-Pro Licence, with commercial-use restrictions including a separate agreement for model-as-a-service revenue above $20 million a year; Qwen's hosted model is closed, and its same-generation open weights are the text-only Qwen3.8-2.4T-A95B published on Hugging Face in August (262,144-token native context) under the Qwen3.8-Max Licence
7. Access: Kimi is served via the Kimi API and 22 providers, with vLLM, SGLang and TokenSpeed recommended for self-hosting; Qwen is served via Alibaba Cloud Model Studio, with 2 providers listed by Artificial Analysis

Read the scores in two groups: vendor tables are each lab's own runs on its own harness, so only the same benchmark at the same version sits side by side; the Artificial Analysis Intelligence Index v5.0.1 and Terminal-Bench 4.0 are the same institution measuring both models on the same protocol, while older index readings use a different protocol and do not compare with the new one.

1. Artificial Analysis Intelligence Index v5.0.1 (checked Nov 12, 2026): Kimi 46, Qwen3.8-Max (0902) 34, a 12-point gap
2. Terminal-Bench 4.0 (measured independently by Artificial Analysis, leaderboard of Nov 11, 2026): Kimi 67.2%, ranked 1st overall; Qwen3.8-Max 38.9%, ranked 45th; a 28.3-point gap
3. DeepSWE v1.1 (vendor-reported): Kimi 75.5%, Qwen3.8-Max-0902 69.3% (Alibaba's Sep 2 snapshot table), a 6.2-point gap
4. MMMU-Pro (vendor-reported): Kimi 84.1%, Qwen3.8-Max-0902 82.7% (Alibaba's Sep 2 snapshot table), a 1.4-point gap
5. GPQA Diamond (vendor-reported): Kimi 94.2%, Qwen3.8-Max 92.6% (Aug 3 release table, not updated in the 0902 snapshot), a 1.6-point gap
6. SWE-bench Pro (vendor-reported): Qwen3.8-Max 67.7%; Kimi has not published a score for this benchmark
7. Terminal-Bench Science 0.1 (vendor-reported): Kimi 38.1%; Qwen has not published a score for this benchmark; on Kimi's own table the figure trails Claude Opus 5.5 at 63.3% and GPT-6 Astra at 68.1%
8. Speed (measured by Artificial Analysis, checked Nov 12, 2026): time per Intelligence Index task is 37.5 minutes for Kimi and 52.8 minutes for Qwen; time to first answer token is 169.70s for Kimi and 118.77s for Qwen

Match the scenario to the numbers on function and price: self-hosted weights, terminal coding and the unpublished output cap sit on Kimi's side of the table, while video input, output price and time to first token sit on Qwen's; where neither side covers the need, the unpublished items in the table call for re-testing on the target task first.

1. Running flagship weights on your own hardware: Kimi's full weights are public (Nov 4), with vLLM, SGLang and TokenSpeed recommended for self-hosting; Qwen's hosted model is closed, and the downloadable same-generation weights are the text-only Qwen3.8-2.4T-A95B
2. Video input: Qwen supports text, image and video input; Kimi supports text and image input, not video
3. Pay-as-you-go calls with a high share of output tokens: Qwen's output costs $6.00 per million tokens against Kimi's $10.00, with input at $2.00 for both; cached input is $0.20 per million tokens for Kimi and $0.25 implicit cache for Qwen
4. Terminal coding: on Artificial Analysis' independent Terminal-Bench 4.0 run, Kimi scores 67.2% and Qwen3.8-Max 38.9%
5. Interactive use sensitive to first-token latency: Artificial Analysis measures time to first answer token at 118.77s for Qwen and 169.70s for Kimi; time per Intelligence Index task is 37.5 minutes for Kimi and 52.8 minutes for Qwen
6. Long single responses: Qwen's per-response output cap is 131,072 tokens (Model Studio documentation); Kimi has not published a per-response output cap