Kimi K3: 2.8T Open-Weight Model That Rivals Frontier

Moonshot AI just released what might be the largest open-weight model ever: Kimi K3, with 2.8 trillion parameters. The model is live on their platform and claims frontier-level performance, ranking second only to Claude Fable 5 and GPT-5.6 Sol on internal evaluations. Full weights are promised in the coming days. This is a serious entry into the open AI arms race.

Kimi K3: Performance and Pricing Details

Kimi K3 is Moonshot's latest model, trained at massive scale. According to the official announcement, it scores 1687 on the GDPval-AA v2 leaderboard, which evaluates AI on real-world tasks across 44 occupations. The model supports a 1-million-token context window and forces "max" reasoning effort—there's no way to dial it down. As one HN commenter noted:

reasoning_effort currently supports only max; K3 always has thinking mode enabled.

Pricing is $3 per million input tokens and $15 per million output tokens, with a cached input rate of $0.3. That's essentially identical to Anthropic Sonnet's pricing (until Sonnet 5's current discount) and close to Terra's $2.5 input rate. This is expensive for a Chinese open-weight model, but if performance holds, it may be justified.

Full model weights will be released in the coming days, along with a technical report detailing architecture, training, and evaluation.

Community Reaction to Kimi K3 Open-Weight Model

The HN thread has 150+ points and 100+ comments, mixing excitement and skepticism. Many are impressed by the model's size—2.8T parameters dwarfs DeepSeek-V4-Pro's 1.6T and Mistral Large 3's 675B. Others question the high pricing and fixed reasoning effort. One commenter shared a real-world cost:

"95 input, 16,658 output = 25 cents! That's the most expensive Pelican I've rendered through a Chinese model so far."

That reference to using Kimi K3 via OpenRouter highlights that while performance may be near frontier, cost is real. The community is watching closely because open-weight releases from Chinese labs have historically driven down prices and increased competition.

Kimi K3: Strengths and Caveats

Kimi K3 is a bold move. Moonshot is essentially saying, "We can hang with Anthropic and OpenAI, and we're open-sourcing it." That's a powerful narrative, but details matter.

First, the 2.8T parameter count is massive, but as with all dense MoE models, effective capacity depends on active parameters. DeepSeek-V4-Pro has 1.6T total but only 49B active; Kimi K3's active parameter count isn't disclosed. That number matters for inference cost and speed.

Second, the pricing is aggressive—not cheap, but competitive with frontier. For builders, this means another top-tier option, but the lack of flexibility (fixed reasoning effort, no temperature adjustments) limits use cases. You can't use it for creative writing or lightweight tasks without paying the full reasoning cost.

Third, the open-weight promise is huge. If weights are truly open and performance holds, it could fuel fine-tuned variants and self-hosted deployments, pressuring existing providers to lower prices or improve transparency.

But treat the performance claims with healthy skepticism. The benchmark scores come from Moonshot's own testing. Wait for third-party evaluations and the technical report before committing.

Kimi K3: Implications for AI Builders

If you're building applications requiring top-tier reasoning—code generation, legal analysis, scientific research—Kimi K3 is worth a trial. The 1M context window is a key advantage over GPT-4o (128K) or Claude Sonnet (200K). For large-document processing, K3 might be a better fit.

• Pricing comparison: For a typical batch of 1000 queries (500 input, 500 output tokens each), Kimi K3 and Claude Sonnet both cost $9.00. DeepSeek-V4 costs ~$0.5 for the same volume. K3 is premium-priced.

• Self-hosting: Open weights could save on inference at scale, but you need serious GPU capacity to run a 2.8T model. Even with quantization and MoE tricks, infrastructure demands are high.

• API access: For now, use the API through providers like OpenRouter, which already supports K3. Monitor the pricing page for cache discounts and updates.

# Pricing per 1M tokens
k3_input = 3.0 / 1_000_000
k3_output = 15.0 / 1_000_000

# Cost for 1000 queries
input_tokens = 500 * 1000  # 500K
output_tokens = 500 * 1000  # 500K
cost_input = input_tokens * k3_input
cost_output = output_tokens * k3_output
total_cost = cost_input + cost_output

print(f"Kimi K3 total: ${total_cost:.2f}")
# Output: Kimi K3 total: $9.00

Kimi K3: Verdict for Your Use Case

If you're an AI researcher or builder working on complex reasoning tasks, yes—Kimi K3 is a new contender that could improve results or reduce costs if self-hosted. If you're a casual user or building simple chatbots, skip it; cheaper models like GPT-4o mini or DeepSeek's V-series are more cost-effective.

The real story is the accelerating pace of open-weight releases, which will eventually democratize access to frontier-level AI. For now, K3 is promising but unproven—watch the technical report and community benchmarks before committing.