Shannon 2 Lite

Most production workloads don't need a model's absolute ceiling on every call — they need consistent quality at a price and latency that scale. Shannon 2 Lite is built for exactly that: the full frontier-distilled Shannon 2 behaviour, quantized to FP8 so you can put it in front of high-traffic products and dozens-of-turns agents without the bill of a full-precision frontier model.

01The foundation: Kimi K2.7

Shannon 2 Lite is based on Kimi K2.7, Moonshot AI's open-weights flagship (released June 12, 2026): a sparse Mixture-of-Experts model where only a small fraction of a trillion parameters activate per token, delivering frontier-class quality at a serving cost far below a dense model of the same size.

Because the weights are open, we host and quantize the model ourselves rather than renting it — which is what makes Lite's FP8 economics possible.

02FP8 quantization — the heart of Lite

Shannon 2 Lite is quantized to FP8: 8-bit floating point for weights and activations. Versus full precision, FP8 roughly halves memory bandwidth and materially increases tokens per second, while modern per-tensor scaling keeps quality loss on instruction-following tasks small. The practical result:

• Lower cost per token — the biggest lever for high-volume products.
• Lower latency — faster first token and higher sustained throughput.
• Smaller footprint — fits far fewer accelerators per replica.
• Same behaviour — identical 256K context and the same distilled instruction-following as Pro.

03Frontier distillation

Lite and Pro share one post-training pass: 30,000 curated, frontier-grade reasoning and instruction examples. The aim is to sharpen how the model answers — cleaner instruction-following, more consistent formatting, better tool-call discipline, and fewer needless refusals on legitimate professional work — not to change what it knows. Applied identically to both builds so they stay behaviourally aligned.

04Cost & performance, honestly

Lite's headline is economics. On list API prices, the K2.7 foundation undercuts today's leading closed models by roughly 6x on output tokens:

Metric	Shannon 2 Lite	Claude Opus 4.8	GPT-5.5
Input / 1M tokens	$0.95	$5.00	$5.00
Output / 1M tokens	$4.00	$25.00	$30.00
Open weights	Yes	No	No
Context window	256K	1M	~1M

On capability, the honest reference point is MCPMark Verified (real-world agentic software tasks) — the only public benchmark on which the K2.7 foundation, Claude Opus 4.8, and GPT-5.5 all report numbers on the same test:

The foundation beats Claude Opus 4.8 on agentic tasks and trails GPT-5.5 — at a fraction of either one's price. For high-volume work, that price-to-capability ratio is the whole point of Lite.

MCPMark Verified & list API prices, June 2026. K2.7 figures are Moonshot-reported; independent third-party benchmarks are pending. GPT-5.5 and Claude Opus 4.8 are shown for reference.

05Minimal censorship, maximum responsibility

Shannon 2 Lite is tuned for minimal censorship: on legitimate security, red-team, and research tasks it stays direct instead of refusing by reflex. It is a professional tool — access is gated to verified professionals, usage is continuously audited, and the model is operated under our Responsible Use Policy.

06Where Lite shines

• High-volume assistants — FP8 economics make it the default for user-facing, high-traffic features.
• Agent loops — cheap enough to run for dozens of turns; 256K context for long trajectories.
• Recon & triage — fast, low-cost first-pass analysis in security workflows.
• Retrieval & classification — high throughput for pipelines and batch jobs.

07Frequently asked questions

---

Sources: Moonshot AI (Kimi K2.7) · K2.7 vs GPT-5.5 vs Claude Opus 4.8 comparison · Independent K2.7 pricing analysis. K2.7 benchmarks are Moonshot-reported and provisional pending independent verification.