Google's largest open-weight Gemma 3 model, running vision-plus-text workloads on a single GPU with a 128K context.
Modality:
Text
Image
Video
PDF
model ID
google/gemma-3-27b-it
Output Speed *
0.00
tok/s
Intelligence Index *
4.9
/ 100
Context Window *
131072
tokens
Input price
0.48
Anytoken
Output price
0.96
Anytoken
Gemma 3 27B: Multimodal Open Weights That Run on One GPU
Gemma 3 27B is the largest model in Google DeepMind's open-weight Gemma 3 family, built from the same research as Gemini. It accepts text and image input and returns text, supports 140+ languages, and exposes a 128K-token context. Its defining trait is capability-per-accelerator: it delivers competitive quality while remaining deployable on a single GPU, including consumer cards with quantization. It suits teams that need self-hosted, license-permissive multimodal inference for document analysis, multilingual generation, and visual question answering without depending on a proprietary API.
Start building with the Gemma 3 27B API on AnyAPI.ai.
Performance
Where Gemma 3 27B Earns Its Place: Quality Per Accelerator
Gemma 3 27B's strongest characteristic is delivering solid general and multimodal quality at a size that fits a single accelerator. It scores 67.5 on MMLU-Pro and 89.0 on MATH in Google's technical report, and reached an LMArena Elo around 1338 at release, competing with much larger open models. Because that quality runs on one GPU, teams can self-host multimodal inference without multi-node serving or per-token API dependence. The practical consequence: predictable, license-permissive deployment for document, vision, and multilingual workloads where controlling infrastructure matters more than frontier reasoning scores.
Benchmarks
Gemma 3 27B Benchmarks: Strong Math, Modest Coding and Agents
In Google's Gemma 3 technical report, the instruction-tuned 27B scores 67.5 on MMLU-Pro and 89.0 on MATH, indicating solid knowledge and symbolic reasoning for its size. Coding is more modest: it reaches 29.7 on LiveCodeBench, and independent aggregators place its coding percentile lower among peers. Agentic tool use is a clear weak point, scoring only 6.6 on τ²-bench (Retail) in later Google comparisons. FACTS Grounding of 74.9 contrasts with a low SimpleQA of 10.0, meaning it grounds well on provided documents but is weaker on unaided factual recall.
Output Speed
*
0.00
tok/s
Intelligence Index
*
4.9
/ 100
MMLU *
Broad world knowledge and problem-solving
67
%
GPQA *
PhD-level scientific reasoning across physics, biology, chemistry.
43
%
HLE *
Adherence to multi-step structured instructions.
4
%
LiveCodeBench *
Tool-calling reliability in long agentic loops.
14
%
Technical Specifications
What the model supports
Gemma 3 27B accepts text and image input and produces text output; it is not a reasoning model and has no extended-thinking mode. The 128K-token context and 140+ language coverage are its most production-relevant specifications, enabling long-document and multilingual work in a single request. Note the distinction between context window and output: it processes up to 128K tokens of input but generates far fewer per completion. Open weights under the Gemma license permit commercial use and self-hosting, which is the main reason teams choose it over closed alternatives.
Verified Specifications —
Gemma 3 27B Instruct
*
Input modalities
Text
Image
Video
PDF
output modalities
Text
Context window
131072
tokens
Maximum output tokens
16384
Reasoning
No
Knowledge cutoff
March 2025
Pricing (standard)
0.48
AnyTokens in
/
0.96
AnyTokens out
Quickstart
import requests
url = "https://api.anyapi.ai/v1/chat/completions"
payload = {
"stream": False,
"tool_choice": "auto",
"logprobs": False,
"model": "Model_Name",
"messages": [
{
"content": [
{
"type": "text",
"text": "Hello"
},
{
"image_url": {
"detail": "auto",
"url": "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"
},
"type": "image_url"
}
],
"role": "user"
}
]
}
headers = {
"Authorization": "Bearer AnyAPI_API_KEY",
"Content-Type": "application/json"
}
response = requests.post(url, json=payload, headers=headers)
print(response.json())
import requests url = "https://api.anyapi.ai/v1/chat/completions" payload = { "stream": False, "tool_choice": "auto", "logprobs": False, "model": "Model_Name", "messages": [ { "content": [ { "type": "text", "text": "Hello" }, { "image_url": { "detail": "auto", "url": "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg" }, "type": "image_url" } ], "role": "user" } ]
}
headers = { "Authorization": "Bearer AnyAPI_API_KEY", "Content-Type": "application/json"
} response = requests.post(url, json=payload, headers=headers) print(response.json())
View docs
Copy
Code is copied
const url = 'https://api.anyapi.ai/v1/chat/completions';
const options = {
method: 'POST',
headers: {Authorization: 'Bearer AnyAPI_API_KEY', 'Content-Type': 'application/json'},
body: '{"stream":false,"tool_choice":"auto","logprobs":false,"model":"Model_Name","messages":[{"content":[{"type":"text","text":"Hello"},{"image_url":{"detail":"auto","url":"https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"},"type":"image_url"}],"role":"user"}]}'
};
try {
const response = await fetch(url, options);
const data = await response.json();
console.log(data);
} catch (error) {
console.error(error);
}const url = 'https://api.anyapi.ai/v1/chat/completions';
const options = { method: 'POST', headers: {Authorization: 'Bearer AnyAPI_API_KEY', 'Content-Type': 'application/json'}, body: '{"stream":false,"tool_choice":"auto","logprobs":false,"model":"Model_Name","messages":[{"content":[{"type":"text","text":"Hello"},{"image_url":{"detail":"auto","url":"https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"},"type":"image_url"}],"role":"user"}]}'
}; try { const response = await fetch(url, options); const data = await response.json(); console.log(data);
} catch (error) { console.error(error);
}
View docs
Copy
Code is copied
curl --request POST \
--url https://api.anyapi.ai/v1/chat/completions \
--header 'Authorization: Bearer AnyAPI_API_KEY' \
--header 'Content-Type: application/json' \
--data '{
"stream": false,
"tool_choice": "auto",
"logprobs": false,
"model": "Model_Name",
"messages": [
{
"content": [
{
"type": "text",
"text": "Hello"
},
{
"image_url": {
"detail": "auto",
"url": "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"
},
"type": "image_url"
}
],
"role": "user"
}
]
}'curl --request POST \ --url https://api.anyapi.ai/v1/chat/completions \ --header 'Authorization: Bearer AnyAPI_API_KEY' \ --header 'Content-Type: application/json' \ --data '{ "stream": false, "tool_choice": "auto", "logprobs": false, "model": "Model_Name", "messages": [ { "content": [ { "type": "text", "text": "Hello" }, { "image_url": { "detail": "auto", "url": "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg" }, "type": "image_url" } ], "role": "user" } ] }'
View docs
Copy
Code is copied
View docs
Limitations & Trade-offs
1
No reasoning mode. Gemma 3 27B is confirmed as a non-reasoning model that produces direct responses without extended chain-of-thought. On tasks that reward step-by-step deliberation — hard math contests, multi-step planning, complex debugging — it trails models with explicit thinking modes. Its later successor's dramatic gains on AIME and agentic benchmarks illustrate the gap. If your workload depends on deep reasoning or self-correction, choose a dedicated reasoning model instead.
2
Weak agentic tool use. Independent Google comparisons show Gemma 3 27B scoring only 6.6 on τ²-bench (Retail), a measure of multi-step tool calling and error handling. While the model supports function calling and structured outputs on many providers, its reliability in autonomous, multi-turn agent loops is limited. For production agents that call tools, execute steps, and recover from failures, this model is a risky default and a stronger tool-use model is preferable.
3
Modest coding performance. On LiveCodeBench the 27B reaches roughly 29.7, and independent aggregators place its coding percentile low among comparable models. Real-world testing found it useful for small-scale corrections and implementation support but unreliable on large-codebase understanding and complex refactors. It can assist coding, but it is not a strong choice as the primary engine for coding agents or repository-scale generation.
4
Weak unaided factual recall. Gemma 3 27B scores only 10.0 on SimpleQA while scoring 74.9 on FACTS Grounding, meaning it answers well when grounded on supplied documents but is unreliable on open-ended factual questions from memory. Combined with an August 2024 knowledge cutoff and no built-in web search, this makes retrieval augmentation essential for factual applications rather than optional.
Best-Fit Workloads
Where this model earns its place
01
Visual question answering and document analysis
Text-plus-image input makes Gemma 3 27B suitable for VQA, image description, and document understanding, with a SigLIP-based vision encoder and MMMU around 64.9. Combined with the 128K context, it can analyze long documents that mix text and figures in a single request. It fits internal tools that extract or summarize from scanned pages, screenshots, and mixed-media inputs where self-hosting is preferred.
02
Multilingual generation and translation-style tasks
With support for over 140 languages and strong post-training on multilingual data, Gemma 3 27B handles generation, summarization, and cross-lingual tasks across a wide language range. Global-MMLU-Lite of 75.1 supports broad multilingual competence. It suits products serving international audiences that need on-prem or license-permissive multilingual inference rather than a proprietary API.
03
Retrieval-augmented generation (RAG)
A 74.9 FACTS Grounding score alongside a low SimpleQA indicates the model performs best when answers are grounded on provided context. The 128K window lets you feed substantial retrieved passages per query. This makes Gemma 3 27B a practical open-weight generator inside RAG pipelines, where retrieval compensates for the August 2024 cutoff and weak unaided recall.
04
Self-hosted, cost-controlled inference
Because Gemma 3 27B is the most capable Gemma 3 size that runs on a single GPU — including consumer cards with quantization-aware training — it fits teams that need predictable, self-managed inference. Open weights under the Gemma license permit commercial use and full deployment control, avoiding per-token API dependence for high-volume or privacy-sensitive workloads.
Pricing in anytokens via AnyAPI
Input
0.48
₳
Output
0.96
₳
Cache write
—
₳
Cache read
—
₳
Integration
01
One API integration
02
Easy model switching
03
Flexible for production
04