Which open-weight AI model is best in August 2026?

As of August 2026, Qwen3.8-Max leads on hosted cost-per-capability ($2/$6 per MTok, weights releasing the week of Aug 10) and edges ahead on coding. Kimi K3 ($3/$15, weights out July 27) has the longest open track record. DeepSeek V4 is the cheapest to run and the most permissively licensed. Pick Qwen3.8-Max for coding value, Kimi K3 for a proven agentic model, DeepSeek V4 for lowest cost and cleanest license.

How much do Qwen3.8-Max, Kimi K3, and DeepSeek V4 cost via API?

Qwen3.8-Max is $2 input / $6 output per million tokens (GA Aug 3, 2026). Kimi K3 is $3 / $15 flat. DeepSeek V4 Flash is $0.14 / $0.28 (off-peak, 2x during peak) and DeepSeek V4 Pro is roughly $0.44 / $0.87 off-peak — the cheapest of the three by a wide margin.

Are Qwen3.8-Max weights actually open?

Qwen3.8-Max weights are being published the week of August 10, 2026 on Hugging Face and ModelScope — but under a custom license with conditions for some large model-as-a-service and consumer deployments, not Apache-2.0. Kimi K3 and DeepSeek V4 offer more permissive terms, so verify the license against your use case.

Quick Answer

Qwen3.8-Max vs Kimi K3 vs DeepSeek V4: Open Frontier

Published: August 11, 2026

The Short Answer

As of August 2026, the open-weight frontier is a three-way race: Qwen3.8-Max (best coding value, weights week of Aug 10, $2/$6), Kimi K3 (proven agentic model, weights out July 27, $3/$15), and DeepSeek V4 (cheapest and most permissive). Pick by coding value, track record, or cost + license.

Quick Comparison

	Qwen3.8-Max	Kimi K3	DeepSeek V4
Vendor	Alibaba	Moonshot	DeepSeek
Weights	Week of Aug 10, 2026	Jul 27, 2026	Released
API price (in/out per MTok)	$2 / $6	$3 / $15	Flash $0.14 / $0.28; Pro ~$0.44 / $0.87
License	Custom (conditions)	More permissive	Most permissive
Best for	Coding value	Proven agentic	Cheapest + cleanest license

Qwen3.8-Max — Best Coding Value

Qwen3.8-Max is the first Max-class Qwen released with open weights (week of Aug 10, 2026). Artificial Analysis scored its max config at 57, it tracks Kimi K3 and DeepSeek V4 Flash across most categories, and it’s stronger on coding and software tasks. Hosted at $2/$6 — the best cost-per-capability of the three. Caveat: custom license.

Kimi K3 — Proven Agentic Model

Kimi K3 (weights July 27, 2026) is the established open agentic pick, at $3/$15 flat (no peak/valley surge). Longer track record in production agent stacks and a cleaner license than Qwen3.8-Max. Costs more per token than the other two.

DeepSeek V4 — Cheapest + Cleanest License

DeepSeek V4 is the budget king: Flash at $0.14/$0.28 off-peak, Pro at ~$0.44/$0.87 off-peak — far cheaper than either rival, with the most permissive license. Peak pricing (1–4 & 6–10 UTC) doubles rates, so schedule batch work off-peak.

Which Should You Pick?

Best coding + hosted value → Qwen3.8-Max.
Proven agentic, simple flat pricing → Kimi K3.
Cheapest to run + safest license → DeepSeek V4.

The Reality Check

All three trail the closed frontier (Claude Opus 5, GPT-5.6 Sol) on the hardest agentic tasks — but they’ve closed the gap enough that self-hosting is a real cost play. Choose on license fit first, then price, then coding quality.

Sources

OrcaRouter — Qwen 3.8 vs Kimi K3 (GA Aug 3, $2/$6 vs $3/$15): orcarouter.ai
Qubrid AI — Kimi K3 vs Qwen3.8-Max full comparison: qubrid.com
DeepSeek — API pricing: api-docs.deepseek.com