LLM Models
Shisa LLM ships in three tiers. All three use the same OpenAI-compatible chat completions endpoint — you select a model with the model field in your request.
| Model ID | Tier | Parameters | Best for |
|---|---|---|---|
shisa-ai/shisa-v2.1-unphi4-14b | Flash | 14B | Cost-efficient, high-throughput everyday tasks |
shisa-ai/shisa-v2.1-llama3.3-70b | Pro (recommended) | 70B | The best balance of quality and speed for bilingual JA/EN work |
shisa-ai/shisa-v2-405b | Flagship | 405B | Maximum quality; access by request |
Shisa V2.1 Flash
shisa-ai/shisa-v2.1-unphi4-14b
The efficient tier. A 14B model built on a Phi-4 base, tuned for Japanese. Use it when latency and cost matter more than squeezing out the last few points of quality — chat widgets, classification, high-volume batch jobs.
Shisa V2.1 Pro
shisa-ai/shisa-v2.1-llama3.3-70b
The recommended default. A 70B model on a Llama 3.3 base. It posts a Japanese MT-Bench of 9.26 and a 73.1% Japanese average, making it the right choice for most production bilingual workloads. Read the Shisa V2.1 write-up for full evaluation details.
Shisa V2 405B
shisa-ai/shisa-v2-405b
The flagship. A 405B model on a Llama 3.1 base that reaches an 80.5% Japanese average — the highest quality Shisa offers.
The 405B model is available to whitelisted accounts. Contact us to request access. Read the Shisa V2 405B write-up for details.
Context windows
Each tier has its own context window: Shisa V2.1 Flash supports 16K tokens, Shisa V2.1 Pro supports 65K tokens, and Shisa V2 405B supports 131K tokens. Confirm the exact limits for your account in the Shisa platform, as served limits can change.
Next steps
- Quickstart — make a request against any model.
- Chat completions reference — the full request/response schema.
- Pricing — how each tier is billed.