Language model

Himalaya Gemma 4 E2B

A bilingual Nepali–English Gemma 4 fine-tune from HimalayaAI.

Status: Available
Parameters: 5.1B
Effective: 2.3B
Context: 4K
Best for: Bilingual Nepali–English assistants, content, and language-aware applications

Model IDs

Model ID	Precision	Use
`himalaya-gemma-4-q4`	Q4 K M	Smallest footprint
`himalaya-gemma-4-q8`	Q8	Balanced default
`himalaya-gemma-4-bf16`	BF16	Highest-precision variant
`himalaya-gemma-4-q4-private`	Q4 K M	Private · no internal retention
`himalaya-gemma-4-q8-private`	Q8	Private · no internal retention
`himalaya-gemma-4-bf16-private`	BF16	Private · no internal retention

Use one of the model IDs above with Chat Completions or Responses at https://tarka.rest/v1.

Pricing

himalaya-gemma-4-q4: Free

himalaya-gemma-4-q8: Free

himalaya-gemma-4-bf16: Free

himalaya-gemma-4-q4-private: $0.06 input · $0.02 cached input · $0.16 output per million tokens

himalaya-gemma-4-q8-private: $0.07 input · $0.02 cached input · $0.18 output per million tokens

himalaya-gemma-4-bf16-private: $0.08 input · $0.02 cached input · $0.20 output per million tokens

USD · catalog 2026-11-04.1. Release verified 2026-11-04. EUR uses the catalog conversion. Live catalog.

Overview

Himalaya Gemma 4 E2B is HimalayaAI’s full-parameter fine-tune of Gemma 4 E2B for Nepali and English instruction following. Its training mix combines Nepali instruction data with OpenHermes, making it the larger bilingual option in Tarka’s currently available language-model catalog.

Available aliases

Choose the precision for your workload without changing the API shape:

himalaya-gemma-4-q4 — the smallest serving footprint
himalaya-gemma-4-q8 — the balanced default
himalaya-gemma-4-bf16 — the highest-precision variant

All three use the deployed 4K context window and support streaming Chat Completions and Responses through https://tarka.rest/v1.

When to pick it

Nepali–English assistants and customer-support workflows
Content generation that moves between Nepali and English
Applications that need more capacity than HimalayaGPT 0.5B while retaining a Nepal-focused training mix

When to pick something else

The smallest possible Nepali route → HimalayaGPT 0.5B
Context beyond the deployed 4K route
Tool-calling workflows; the current aliases do not advertise tool support

The model and weights are by HimalayaAI. See the Hugging Face model card for upstream training, datasets, formats, and Apache 2.0 licensing details.

Data retention

The original free aliases retain inputs and outputs in Tarka’s private internal archive for evaluation and debugging until Tarka administrators delete them. They are not used for model training.

Choose an alias ending in -private to avoid internal retention. Private aliases use the inference API-key scope and are currently priced at zero. Organization capture, when enabled, writes only to your organization’s own bucket.