Qwen3.8 Max vs Kimi K3
Two Chinese frontier models, released about 72 hours apart in July 2026. Most comparisons you'll read are vibes; this one separates what's verified from what nobody knows yet.
If you need to choose today (2026-07-21): Kimi K3 is the safer operational choice — published $3/$15 per-M-token pricing, documented benchmarks, broad API access, and open weights scheduled for July 27. Qwen3.8 Max is the bigger open question — a claimed 2.4T-parameter flagship that currently has no public API, no pricing, and no benchmark table, but very low-cost trial access (Token Plan preview at ~10% credit rates).
The one independent head-to-head so far: Kimi K3 83, Qwen3.8 Max 80 — a single blind run, effectively a tie in signal terms.
Side-by-side: verified vs unknown
✔ = verified against official sources on 2026-07-21 · "claim" = vendor statement · "—" = not published
| Qwen3.8 Max | Kimi K3 | |
|---|---|---|
| Model ID | qwen3.8-max-preview ✔ | Kimi K3 (Moonshot) |
| Announced | 2026-07-19 ✔ | ~2026-07-16 (72h earlier) |
| Total parameters | 2.4T (claim, no model card) | 2.8T (claim) |
| Context window | 1M ✔ | 1M (published) |
| API pricing | — (PAYG not announced) | $3 in / $15 out per M tokens (published) |
| Access today | Token Plan / Qoder / QoderWork only ✔ | Hosted API + broad availability |
| Open weights | Promised, no date | Scheduled 2026-07-27 |
| Official benchmark table | — (claim: "second only to Fable 5") | Published |
| Native vision | Reported multimodal; details unpublished | Published |
| Cheapest real trial | $6/mo Token Plan Lite at ~10% preview rates ✔ | PAYG from first token |
What the one real head-to-head showed
Trilogy AI ran a matched, blind repository task (269 files) on both models:
- Scores: Kimi K3 83, Qwen3.8 Max 80.
- Efficiency: Kimi finished faster and used fewer tokens.
- Quality texture: Qwen produced cleaner system boundaries and stronger replay metadata.
One run, one task — treat it as a coin-flip-adjacent data point, not a ranking. Community one-shot demos on X (e.g., landing-page generation) have gone both ways.
Decision guide
| Your situation | Sensible pick today |
|---|---|
| Production app, need pricing you can budget | Kimi K3 (published pricing) or a generally-available Qwen model — Qwen3.8 Max literally can't be bought per-token yet |
| Interactive coding agent, cost-sensitive evaluation | Qwen3.8 Max via $6 Token Plan — the preview 10% credit rate is the cheapest frontier-model trial running |
| Self-hosting / fine-tuning plans | Kimi K3 if July 27 weights land; Qwen3.8 weights are promised with no date |
| Waiting for real data | Both models are 2 days old with preview drift risk — see our benchmark tracker; our own multi-model coding-agent evaluation is in progress |
Dates to watch
- 2026-07-27 — Kimi K3 scheduled open-weight release (this table updates that day)
- Unannounced — Qwen3.8 Max benchmark table, PAYG pricing, open weights; tracked on our pricing page
Verified 2026-07-21. Both models are in their first week — expect this page to change; every revision keeps its date stamps.
Sources: Qwen Cloud docs · Trilogy AI head-to-head · OrcaRouter comparison (K3 specs) · MarkTechPost