Skip to content
Model infoModelsSnapshot: 15 Oct 2026

Ricoh / Ricoh LLM

Japanese-GPT-6B

Original Japanese autoregressive research model trained from scratch on 170B tokens, presented in March 2023.

Language models (LLMs)
historicalClosed weightsText generation
Official documentation (opens in a new tab)Reviewed 14 Oct 2026
9 of 20 specification fields collected
Context window2K2,048 tokens
Maximum outputNot collectedAwaiting source review
Input billingNot collectedRicoh research (historical)
Output billingNot collectedNo inferred price

Specifications

Provider-documented facts. “Not collected” means we have not verified that field yet, not that the feature is unsupported.

Context / input limit
2,048 tokensSource for Context / input limit (opens in a new tab)

Published max_context_width, shared sequence capacity; separate serving input/output limits are not established.

Maximum input tokens
Not collected
Maximum output
Not collected
Input modalities
TextSource for Input modalities (opens in a new tab)
Output modalities
TextSource for Output modalities (opens in a new tab)
Reasoning
Not collected
Tool calling
Not collected
Structured output
Not collected
Streaming
Not collected
Batch processing
Not collected
Fine-tuning
Not collected
Knowledge cutoff
Not collected
Release date
2023-03Source for Release date (opens in a new tab)

Conference month; no exact public deployment date.

Parameters
6BSource for Parameters (opens in a new tab)
Architecture
GPT-3-style decoder, 28 layers, hidden width 4096, 16 attention headsSource for Architecture (opens in a new tab)
Languages
JapaneseSource for Languages (opens in a new tab)

Documented focus; not an exhaustive supported-language list.

Context caching
Not collected

No model-specific prompt-cache contract, lifetime or cached-input price verified. Quantization and image-token compression are not prompt caching.

Published price
Not collected

No numeric model-specific production rate verified. Private deployment does not establish free inference.

Tokenizer
SentencePiece unigram with byte fallback; 50,272 vocabulary entriesSource for Tokenizer (opens in a new tab)
Published training data
170B Japanese tokens from Wikipedia, CC100, OSCAR and mC4; one epochSource for Published training data (opens in a new tab)

Access & pricing

Prices belong to an endpoint and billing tier. A provider listing does not guarantee access for every account or region.

Ricoh research (historical)

Provider source (opens in a new tab)
API model ID
Not collected

Published model name is not a verified API request selector.

Access
Published research model; no live production API or released checkpoint verified.
Regions
Customer-selected private environment; region and residency depend on the deployment agreement.

Endpoint pricing has not been collected. Check the provider for applicable rates.

Data & enterprise terms

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.

Ricoh research (historical)

Not verified
How it works
Not collected
Minimum reusable input
Not collected
Lifetime & refresh
Not collected
How to enable it
Not collected
Usage fields
Not collected
Billing rules
Not collected
Limits & data boundaries
Not collected

No verified cache rate for this endpoint. An empty rate is not a zero charge.

Caching reviewed 14 Oct 2026

What the measurements say

JCommonsenseQA

Ricoh · vendor-reported
37.4%

Original March 2023 paper, table 5; JGLUE v1.1.0, greedy decoding, no instruction tuning. Three-shot exact-match accuracy on a held-out split of 1,119 questions.

Evaluation source (opens in a new tab)

JSQuAD exact match

Ricoh · vendor-reported
57%

Original March 2023 paper, table 5; JGLUE v1.1.0, greedy decoding, no instruction tuning. Two-shot evaluation; six of 4,777 held-out cases excluded for exceeding context.

Evaluation source (opens in a new tab)

JSQuAD character F1

Ricoh · vendor-reported
72.6%

Original March 2023 paper, table 5; JGLUE v1.1.0, greedy decoding, no instruction tuning. Two-shot evaluation with character-level F1; same filtered split.

Evaluation source (opens in a new tab)

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

Provider identifies this separately trained model for research or private deployment. No matching public derivative was verified in the two provider-linked weight inventories. Closed-weight classification is inferred for this exact revision, not its public upstream base; private distribution terms remain to be confirmed.

Classification evidence (opens in a new tab)

Sources reviewed