Gemma 4 26B A4B

google/gemma-4-26b-a4b-it

The multimodal workhorse and the cheapest thing here to serve. Sparse — roughly 3.8B active of 26.5B — with hybrid attention: 25 of 30 layers use a 1024-token sliding window and only 5 are global. That is why a 262k context costs 5.1GB of KV rather than the 12GB the Qwen 30B needs.

Get an API key

Endpoints

Each row is one provider serving these weights at one precision. Prices are per million tokens.

Provider	Quantisation	Context	Max out	Input	Output	Cached in
Infersia	AWQ int4	256K	32K	$0.06	$0.30	$0.01

Prefix caching is enabled and the discount is passed through. If your agent replays the same system prompt each iteration, the repeated portion bills at the cached rate.

Agent config

Paste this into your coding agent and it can configure itself. The timeout is derived from this variant's measured throughput, not a guess — a long generation that outlives a client's default ceiling is the most common way a working request looks broken.

# Infersia — google/gemma-4-26b-a4b-it
# Paste this to your agent. Values are measured, not aspirational.

provider:
  type: openai-compatible
  base_url: https://api.infersia.com/v1
  api_key: ${INFERSIA_API_KEY}   # from https://infersia.com/dashboard/keys
  model: google/gemma-4-26b-a4b-it

limits:
  max_input_tokens: 262144      # hard limit — over this returns HTTP 413
  max_output_tokens: 32768
  timeout_seconds: 480                   # covers the full 32768 output budget at 145 tok/s, doubled for the tail
  # allow ~14s per 1,000 output tokens if you cap max_tokens lower

features:
  streaming: true
  tools: true
  vision: true
  reasoning: true   # OFF by default; send chat_template_kwargs.enable_thinking=true to opt in
  prompt_caching: true   # automatic; reuse a stable prefix and it bills at the cache rate

pricing_usd_per_million_tokens:
  input: 0.06
  output: 0.3
  cached_input: 0.015

What AWQ int4 means

Activation-aware 4-bit weight quantisation. Small, measurable quality cost.

The quantisation is returned on every request in the x-infersia-quantization response header — so you can assert on it in your own tests rather than trusting this page.

Call it

curlbash

curl https://api.infersia.com/v1/chat/completions \
  -H "Authorization: Bearer $INFERSIA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/gemma-4-26b-a4b-it",
    "messages": [{"role": "user", "content": "Hello"}],
    "stream": true
  }'

Pin a provider with google/gemma-4-26b-a4b-it@infersia, or take the cheapest with the bare id. Add :free for the rate-limited free tier where one exists.