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

As of August 2026 the open-weight frontier is a three-way race. Alibaba's Qwen3.8-Max (2.4T MoE, launched Aug 3) claims the top benchmarks but independent scores are pending. Kimi K3 ($3/$15 API, open weights July 27) is the proven, deployable leader. DeepSeek V4 Pro ($0.435/$0.87 off-peak) is by far the cheapest to run. Pick K3 for a battle-tested deploy, DeepSeek V4 Pro for cost, Qwen3.8-Max if its benchmarks and weights hold up.

Is Qwen3.8-Max better than Kimi K3?

Alibaba claims Qwen3.8-Max ranks higher than Kimi K3 on several benchmarks. As of August 4, 2026 those are vendor numbers — independent SWE-bench and reasoning results aren't in yet. Kimi K3 remains the more proven choice for production because its open weights (July 27) and pricing are established.

What's the cheapest open-weight frontier model to run in 2026?

DeepSeek V4 Pro at $0.435 input / $0.87 output per million tokens (off-peak; 2x during peak windows, cache-hit ~$0.043) is dramatically cheaper than Kimi K3 ($3/$15). For a 30K-in/5K-out task, DeepSeek V4 Pro costs about $0.017 versus $0.16 for Kimi K3.

Quick Answer

Qwen3.8-Max vs Kimi K3 vs DeepSeek V4 Pro (Aug 2026)

Published: August 4, 2026

The Short Answer

Three open-frontier contenders as of August 2026:

Qwen3.8-Max (Alibaba, Aug 3) — biggest and newest, top benchmark claims, weights signaled.
Kimi K3 (Moonshot, open weights July 27) — the proven, deployable leader at $3/$15.
DeepSeek V4 Pro — the cost king at $0.435/$0.87 off-peak.

Side-by-Side

	Qwen3.8-Max	Kimi K3	DeepSeek V4 Pro
Launched	Aug 3, 2026	Open weights Jul 27, 2026	GA Jul 24, 2026
Params	2.4T (MoE)	Large MoE	MoE
Context	1M	Large	Long
API price (in/out per MTok)	TBD	$3 / $15 (flat)	$0.435 / $0.87 (off-peak, 2x peak)
Cache-hit input	TBD	—	~$0.043
Open weights	Signaled	✅ Yes	✅ Yes
30K/5K task cost	TBD	~$0.16	~$0.017

How To Choose

Want the safest production deploy today → Kimi K3. Open weights are out, pricing is flat (no peak/valley surprises), and it was the open-weight benchmark leader before Qwen3.8-Max’s claim.
Optimizing for cost → DeepSeek V4 Pro. At ~~$0.017 per typical task it’s an order of magnitude cheaper than K3. Watch the 2x peak-hour surcharge (1–4 & 6–10 UTC) and cache your prompts (~~$0.043 cache-hit).
Chasing the absolute frontier → Qwen3.8-Max, if the independent benchmarks confirm Alibaba’s claims and the weights ship on schedule. At 2.4T it’s the largest of the three.

Watch Outs

Qwen3.8-Max benchmarks are vendor-reported as of Aug 4, 2026. Don’t rebuild your stack around unverified numbers.
DeepSeek peak pricing doubles in Beijing-time windows — model your real traffic pattern before assuming the off-peak rate.
Scale ≠ deployability. All three are large MoE models; even Qwen3.8-Max’s “open weights” will demand serious GPU memory.

Verdict

Best proven open-weight model → Kimi K3
Cheapest to run → DeepSeek V4 Pro
Biggest / potential new leader → Qwen3.8-Max (pending independent benchmarks)

Sources

SiliconANGLE — Qwen3.8-Max: siliconangle.com
Bloomberg — Qwen3.8-Max vs Kimi K3 benchmarks: bloomberg.com
DeepSeek — pricing: deepseek.ai/pricing
Moonshot — Kimi: moonshot.ai