Qwen 3.8 vs. Kimi K3: benchmarks, licenses, and what it takes to run them

Kimi K3’s weights landed on Hugging Face on July 27. Qwen3.8-Max followed about two weeks later.

What actually shipped

Moonshot announced K3 in mid-July and published the weights on July 27. It’s one model: a 2.8-trillion-parameter mixture-of-experts with 104B active parameters per token, a native 1M-token context window (featherless serves it at 256k on Developer Plan), and vision built in through the new 401M-parameter MoonViT-V2 encoder. The checkpoint ships natively in MXFP4, so the whole model fits in roughly 1.56TB of weights (more on that below).

Alibaba’s release on August 3 is a family, and the naming deserves a moment of care. The headline model, Qwen3.8-Max, is what you get through Alibaba’s API: multimodal, with a 1M-token context window. The open checkpoint, published five days later as Qwen3.8-2.4T-A95B, is the same scale but text-only: 2.4T total parameters, 95B active, 262K native context extensible to about 1M. It reasons before every answer whether you want it to or not. There’s a reasoning_effort knob (low, medium, xhigh) but no off switch. It is advisable to use it at low reasoning as the model tends to overthink.

The quieter release is the one most people will actually run: Qwen3.8-27B, a dense multimodal model under plain Apache 2.0. Coming in at only 56 gb in total, Qwen 3.8 27B is a fraction of the size of Kimi K3, but still very powerful. Available on Featherless.

The architecture

Both are hybrid-attention MoEs that keep full attention in only a quarter of their layers. K3 runs 69 layers of Kimi Delta Attention, Moonshot’s linear-attention design, against 24 layers of gated MLA. Qwen3.8 stacks 23 repeating blocks of three Gated DeltaNet layers followed by one gated-attention layer, 92 layers in all.

Full attention’s KV cache is what makes million-token contexts expensive to serve; linear attention carries a fixed-size state instead. Moonshot’s Kimi Linear paper, where KDA debuted last October, claimed up to 75% KV-cache reduction and up to 6x decoding throughput at a 1M-token context. Moonshot hasn’t republished those multipliers for K3 itself, but the direction matches what we see serving hybrid models.

Where they diverge is sparsity and precision. K3 is wider: 896 experts with 16 routed plus 2 shared per token, and it went through quantization-aware training from the supervised fine-tuning stage onward, so the 4-bit weights are the model rather than a compressed copy of it. Qwen3.8 uses 512 experts with 10 routed plus 1 shared, and ships at full precision, with a separate FP8 repo for lighter deployments.

The benchmarks that overlap

Neither card benchmarks against the other model. Both labs compare themselves to Claude and GPT instead. Ten evals appear on both cards. Two of those, FrontierSWE and Agents’ Last Exam, show comparator scores that visibly disagree from card to card, a sign the labs ran different versions or setups, so we’ve left them out. The eight below line up, and on five of them both labs quote identical scores for the third-party comparators, which means they’re at least reporting the same tests. These are still vendor numbers; apply the usual salt.

Sources: Kimi K3 model card, Qwen3.8-2.4T-A95B model card. Qwen/Qwen3.8-27B on Hugging Face, Qwen ran its coding-agent evals through the Claude Code harness with five-hour timeouts; Moonshot reports its own setup at max reasoning effort. Same tests, not always the same conditions.

Independent signals, for what they’re worth this early: Artificial Analysis currently scores K3 at 60 on its Intelligence Index, the highest of any open-weights model, with Qwen3.8 right behind at 58. The strongest closed model sits at 63.

What they cost to run

K3’s native MXFP4 weights come to about 1.4TB, and community serving estimates start around eight nodes of 8x80GB GPUs once you add KV cache and activations. Qwen3.8’s full-precision checkpoint is 4.89TB, or ~2.5TB for the FP8 build.

Through first-party APIs, Alibaba serves Qwen 3.8-Max at $2 per million input tokens and $6 per million output, with cached input at $0.25. Moonshot charges $3 and $15, cached input at $0.30. That output gap matters more than it might seem, because both are reasoning models that spend thousands of output tokens thinking before they answer. On a long agent run, Qwen3.8-Max is less than half the price.

On Featherless, Kimi K3 has been live since July 29; and runs at 256K. Qwen3.8-27B is in the catalog as well.

Prices are per million tokens.

If you need 2.4T Qwen, or K3 at its full 1M context, talk to us about a dedicated deployment.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.featherless.ai/v1",
    api_key="YOUR_API_KEY",
)

response = client.chat.completions.create(
    model="moonshotai/Kimi-K3",  # or "Qwen/Qwen3.8-27B"
    messages=[{"role": "user", "content": "Summarize this repo's architecture."}],
)

Qwen or Kimi?

Pick K3 for deep-research and browsing agents, MCP-heavy tool use, or anything that needs to see: its BrowseComp and MCPMark scores top every model on its card. Pick Qwen3.8-Max when instruction-following precision and long-document retrieval matter, or when output-token cost dominates; its IFBench and 256K-retrieval numbers are the best either card reports. Pick Qwen3.8-27B when you want the most value for money model which punches above its weight.