`epithre-nala`

Formerly epithre-omni. The model was renamed on 2026-10-21. epithre-omni keeps working as an alias for the same model at the same price; responses echo the model name you send. New integrations should use epithre-nala. The name comes from Old Javanese, where nala means heart.

The flagship. Default choice for most chat workloads.

Capabilities

Capability	Notes
Tier	Flagship
Context window	131,072 tokens
Max output	16,384 tokens
Modalities	Text, image (vision)
Tool use	Yes; standard `tools` + `tool_choice` schema
Extended thinking	Yes; opt-in via `chat_template_kwargs={"enable_thinking": true}`
Structured output	Yes; `response_format` json_object + json_schema strict
Prompt caching	Yes; `cache_control` markers
Streaming	Yes; SSE
Indonesian fluency	Native (tuned on Indonesian corpora)

When to use

General chat with vision
Agentic tool-use chains
Anything that needs strong Indonesian register handling
Vision QA on documents, screenshots, photos
RAG synthesis after retrieval

When NOT to use

Very long context (>32K tokens) - use epithre-prme instead.
High-throughput cheap classification - use epithre-lyt (~6x cheaper).
Audio or video input - use epithre-lyt.

Pricing

Input: Rp7,000 / 1M tokens
Output: Rp25,000 / 1M tokens
Prompt cache write: 1.25x input rate (Rp8,750 / 1M)
Prompt cache read: 0.1x input rate (Rp700 / 1M)
Batch: 0.5x all rates

Performance benchmarks

Task	Score (vs.)
Indonesian general chat	Strong (native fluency)
Indonesian legal QA (with RAG)	~95% accuracy on 200-question gold set
Indonesian sentiment classification	~92% F1 (3-class)
Indonesian vision QA (invoice extraction)	~94% field-level F1
English chat	Strong on everyday and technical English

Latency

Non-streaming, 500-token reply: ~3-8s
Streaming first-token: ~0.4-1.2s
With thinking enabled: +2-10s depending on problem complexity

Caveats

tool_choice="required" is reliable as of May 2026 — prior stall on long prompts (>5K tokens) is resolved. Bare "required", named tool choice, and "auto" all work cleanly on Nala.
Structured output (json_schema) benefits from the same fix. Long-input schema-constrained generation is stable; fall back to json_object only if you hit a specific edge case.