Research training capability, distinct from the historical 16K API serving contract.
attention
Three sliding-window attention layers per full-attention layer; cross-layer KV reuse between consecutive full-attention layersSource for attention (opens in a new tab)
Internal attention/cache design, not a billable prompt-cache API.
Prices retain their published currencies and billing units below.
Published billing units · original currency · observed 14 Oct 2026
Category
Usage
Rate
Unit
Historical text inference
Input1
CN¥0.99
1M tokens
Historical text inference
Output1
CN¥0.99
1M tokens
Rate conditions
Price in the December 5, 2024 archive, charged on input plus output tokens; conversation history is billable input. Not the final pre-shutdown price and not a current offer. No currency conversion.
Retired with the public China platform on September 3, 2026. Requests could be served by a Yi model or a named third-party model; this is not a pinned Yi checkpoint endpoint.
Regions
Host-specific historical service; no current availability or residency guarantee inferred.
Router protocol, not independent evidence for every backend.
Prices retain their published currencies and billing units below.
Published billing units · original currency · observed 14 Oct 2026
Category
Usage
Rate
Unit
Historical text inference
Input1
CN¥0.99
1M tokens
Historical text inference
Output1
CN¥0.99
1M tokens
Rate conditions
Later documentation price for a mixed-model routing service, billed on input plus output tokens. Public service has stopped. Not a pinned Yi checkpoint price; no current inference offer.
Retention, training use, regional processing and enterprise terms have not been summarized for this endpoint.
Caching
Caching can reuse earlier input processing. Availability, lifetime and billing depend on the endpoint. A cache hit still generates a new response.
01 / INITIAL INPUTProcess & retain
The system processes eligible input and retains reusable state.
02 / LATER REQUESTCheck available state
Reuse depends on matching input, session state and the cache policy.
03 / CACHE HITReuse & generate
Reusing state can reduce repeated work. Any charges follow the billing rules below.
01.AI China API (historical)
yi-lightning
Not verified
No model-specific API prompt-cache contract verified.
How it works
Not collected
No model-specific API prompt-cache contract verified.
Minimum reusable input
Not collected
No model-specific API prompt-cache contract verified.
Lifetime & refresh
Not collected
No model-specific API prompt-cache contract verified.
How to enable it
Not collected
No model-specific API prompt-cache contract verified.
Usage fields
Not collected
No model-specific API prompt-cache contract verified.
Billing rules
Not collected
No model-specific API prompt-cache contract verified.
Limits & data boundaries
Not collected
No model-specific API prompt-cache contract verified.
No verified cache rate for this endpoint. An empty rate is not a zero charge.
01.AI China mixed-model router (retired)
yi-lightning
Not verified
No model-specific API prompt-cache contract verified.
How it works
Not collected
No model-specific API prompt-cache contract verified.
Minimum reusable input
Not collected
No model-specific API prompt-cache contract verified.
Lifetime & refresh
Not collected
No model-specific API prompt-cache contract verified.
How to enable it
Not collected
No model-specific API prompt-cache contract verified.
Usage fields
Not collected
No model-specific API prompt-cache contract verified.
Billing rules
Not collected
No model-specific API prompt-cache contract verified.
Limits & data boundaries
Not collected
No model-specific API prompt-cache contract verified.
No verified cache rate for this endpoint. An empty rate is not a zero charge.
What the measurements say
GPQA 0-shot
01.AI · provider-reported
50.9%
Provider-reported results in Table 2 of arXiv 2412.01253v1. The report notes that locally deployed results may exceed API results because serving trades speed against accuracy. This does not evaluate later mixed-model routers.
Provider-reported results in Table 2 of arXiv 2412.01253v1. The report notes that locally deployed results may exceed API results because serving trades speed against accuracy. This does not evaluate later mixed-model routers.
Provider-reported results in Table 2 of arXiv 2412.01253v1. The report notes that locally deployed results may exceed API results because serving trades speed against accuracy. This does not evaluate later mixed-model routers.
Provider-reported results in Table 2 of arXiv 2412.01253v1. The report notes that locally deployed results may exceed API results because serving trades speed against accuracy. This does not evaluate later mixed-model routers.
Provider-reported results in Table 2 of arXiv 2412.01253v1. Judged with GPT-4o-0513. The report notes that locally deployed results may exceed API results because serving trades speed against accuracy. This does not evaluate later mixed-model routers.
Provider-reported results in Table 2 of arXiv 2412.01253v1. Judged with GPT-4o-0513. The report notes that locally deployed results may exceed API results because serving trades speed against accuracy. This does not evaluate later mixed-model routers.
Provider-reported results in Table 2 of arXiv 2412.01253v1. Judged with GPT-4o-0513. The report notes that locally deployed results may exceed API results because serving trades speed against accuracy. This does not evaluate later mixed-model routers.
Provider-reported results in Table 2 of arXiv 2412.01253v1. Judged with GPT-4o-0513. The report notes that locally deployed results may exceed API results because serving trades speed against accuracy. This does not evaluate later mixed-model routers.
Latency and throughput depend on the host, request size, reasoning settings and load. There is no measured speed or global usage figure in this record.
The record behind the model
Why it is included
Inferred for this named proprietary API/research variant from the developer’s release description and public weight inventory. Exact weights for this variant were not found. Public Yi, Yi-1.5, Yi-Coder and Yi-VL checkpoints remain excluded; API naming alone does not prove an independently trained base.
