BroadNet Gemma 4 VL 26B API
Updated
BroadNet Gemma 4 VL 26B, based on Google's Gemma 4, runs on BroadNet Inference as broadnet/gemma-4-26b-a4b through the OpenAI-compatible https://inference.broadnet.ai/v1. It costs $0.10 input and $0.22 output per 1M tokens, with a 32,768-token context. It can use the starting credit, and inference content is not retained.
At a glance
- Text
- Vision (image input)
- Starting credit eligible: no top-up needed
- Gemma hourly key: one hour, one stream, a separate product from the Meteor plan
- Text and image input
- BroadNet runs it in production with a hot backup
- BroadNet reports it matched Claude Sonnet on 5 of 6 steps of its chatbot pipeline in a blind evaluation

| Model ID | broadnet/gemma-4-26b-a4b |
|---|---|
| Endpoint | /v1/chat/completions |
| Supported inputs | text, image (vision: understanding, not generation) |
| Context | 32,768 tokens, prompt and answer together |
| Input | $0.10 per 1M tokens |
| Cached input | Not published |
| Output | $0.22 per 1M tokens |
| Pay-as-you-go limits | 30 requests a minute, 4 in flight per key |
| Access | Starting credit eligible |
| Hourly key | $0.50 for one hour, one stream (a separate product from the Meteor plan) |
Prices on this page are from BroadNet's price feed snapshot dated , not checked live here. See official pricing for current rates.
official pricingReport a wrong price or number (email to info@broadnet.ai, with this page's link filled in).
What does BroadNet publish about BroadNet Gemma 4 VL 26B?
From the BroadNet Gemma 4 VL 26B page on broadnet.ai, checked 15 November 2026. Measurements are BroadNet's, on its hardware, unless a public leaderboard is named; we have not reproduced them.
| Architecture | Mixture of experts: 26B parameters in total, about 4B active per token |
|---|---|
| Serving | 4-bit AWQ weights, multi-token-prediction speculative decoding (k=2), fp8 KV cache, BroadNet conversational chat template |
| Modalities | Text and image in, text out |
| Context on the API | 32K tokens; BroadNet suggests one page per request for long documents |
| Hardware | NVIDIA Blackwell-class GPUs with a hot backup; failover is automatic |
| Licence | Apache-2.0 (Google) |
BroadNet's measurements
- Internal quality battery: 90.86 of 100, the best local model BroadNet says it has measured.
- Blind three-judge evaluation on 240 production conversations across 11 tenants: matched Claude Sonnet on 5 of 6 steps of BroadNet's chatbot pipeline, and answered 1.2 to 6.5 times faster per step.
- About 136 tokens per second on a single stream with speculative decoding.
- The 26B mixture of experts beat the dense Gemma 4 31B on 5 of 6 pipeline steps with about one eighth of the active parameters.
- The same weights served by a hosted provider scored 0.13 lower on average than BroadNet's own serving.
Base weights: google/gemma-4-26B-A4B-it on Hugging Face.
How BroadNet chooses and serves models: its GPU hardware and serving approach and the models it evaluated and ruled out.
Published benchmark results
BroadNet did not run these benchmarks; results are as published by the Google model card and Epoch AI and may use different settings. BroadRouter is BroadNet's site. They refer to the published model google/gemma-4-26B-A4B-it; BroadNet serves its own 4-bit AWQ build of it, so results on BroadNet can differ.
| Benchmark | Result | Setting or note | Source |
|---|---|---|---|
| MMLU Pro | 82.6% | Instruction-tuned model | Google model card, checked 2026-11-16 |
| GPQA Diamond | 82.3% | Instruction-tuned model | Google model card, checked 2026-11-16 |
| LiveCodeBench v6 | 77.1% | Instruction-tuned model | Google model card, checked 2026-11-16 |
| AIME 2026 no tools | 88.3% | Instruction-tuned model | Google model card, checked 2026-11-16 |
| MMMU Pro | 73.8% | Instruction-tuned model | Google model card, checked 2026-11-16 |
| MATH-Vision | 82.4% | Instruction-tuned model | Google model card, checked 2026-11-16 |
| GPQA Diamond | 0.732 | Minimal reasoning setting; Epoch run gemma-4-26b-a4b-it_minimal, mean score on a 0 to 1 scale | Epoch AI, checked 2026-11-16 |
| OTIS Mock AIME 2024-2025 | 0.822 | Minimal reasoning setting; Epoch run gemma-4-26b-a4b-it_minimal, mean score on a 0 to 1 scale | Epoch AI, checked 2026-11-16 |
Epoch AI rows: Epoch AI, Capabilities & benchmarking, licensed CC BY 4.0, shown as the mean score on a 0 to 1 scale, rounded to three decimals (hover a value for the exact figure in Epoch's data file).
Side by side: BroadNet Gemma 4 VL 26B vs DeepSeek V4.1 Flash, BroadNet Gemma 4 VL 26B vs GLM 5.3 and BroadNet Gemma 4 VL 26B vs Qwen3.8 27B, with prices, context and workload costs. Gemma 4 26B A4B is also sold by 11 other providers, from $0.07 input and $0.34 output per 1M tokens on their own pricing pages: Gemma 4 26B A4B API pricing by provider. The same text is a different number of tokens on each model: see how BroadNet Gemma 4 VL 26B's tokenizer compares on Arabic, English and code in the tokenizer efficiency dataset.
Call it in Python
Install the OpenAI SDK with pip install openai, export BROADNET_API_KEY, then run:
from openai import OpenAI
import base64, os
client = OpenAI(base_url="https://inference.broadnet.ai/v1", api_key=os.environ["BROADNET_API_KEY"])
image = base64.b64encode(open("invoice.jpg", "rb").read()).decode()
response = client.chat.completions.create(
model="broadnet/gemma-4-26b-a4b",
max_tokens=2000,
messages=[{"role": "user", "content": [
{"type": "text", "text": "What is the total on this invoice?"},
{"type": "image_url", "image_url": {"url": "data:image/jpeg;base64," + image}},
]}],
)
print(response.choices[0].message.content)Or with curl
curl https://inference.broadnet.ai/v1/chat/completions \
-H "Authorization: Bearer $BROADNET_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"broadnet/gemma-4-26b-a4b","messages":[{"role":"user","content":"Reply with one short sentence."}]}'Guides for BroadNet Gemma 4 VL 26B
- Use the OpenAI Python SDK with BroadNet Inference
- Use the OpenAI Node.js SDK with BroadNet Inference
- Try BroadNet Inference in your browser, no account
- First request to BroadNet Inference with curl
- Stream chat completions and handle dropped connections
- Use Cline with BroadNet Inference (OpenAI Compatible provider)
- Connect Open WebUI to BroadNet Inference
- Add BroadNet Inference to LibreChat as a custom endpoint
- Use BroadNet Inference in n8n (OpenAI credential, Base URL)
- Use the Vercel AI SDK with BroadNet Inference
- Route LiteLLM to BroadNet Inference (SDK and proxy)
- Read Arabic invoices and receipts with a vision model
- Use Jan with BroadNet Inference as a custom OpenAI-compatible provider
- Connect SillyTavern to BroadNet Inference (Custom OpenAI-compatible)
- Call BroadNet Inference from Go with the openai-go SDK
- Use Haystack pipelines and agents with BroadNet Inference
- Use DSPy with BroadNet Inference (set model_type="chat")
- Evaluate BroadNet models with promptfoo
- Connect Flowise to BroadNet Inference (OpenAI Custom Model)
- Connect Langflow to BroadNet Inference
Use BroadNet Gemma 4 VL 26B in the console
The BroadNet console on inference.broadnet.ai manages accounts and keys. It does not receive your choice from this page, so:
- Sign in or create an account.
- Create a pay-as-you-go key; the starting credit covers this model.
- Send this model ID as
modelto/v1/chat/completions:broadnet/gemma-4-26b-a4b
Open the BroadNet consoleEstimate the costRead the API reference
Compare BroadNet Gemma 4 VL 26B
Questions
What is the model ID for BroadNet Gemma 4 VL 26B?
Pass broadnet/gemma-4-26b-a4b as model to /v1/chat/completions. The base URL is https://inference.broadnet.ai/v1.
How much does BroadNet Gemma 4 VL 26B cost?
$0.10 input and $0.22 output per 1M tokens, from your account balance. It is also the model on the Gemma hourly key, a separate flat-price product.
What is the context window of BroadNet Gemma 4 VL 26B?
32,768 tokens per request, prompt and answer together.
Can I use BroadNet Gemma 4 VL 26B without a top-up?
Yes. It can use the starting credit for new accounts.
Where does BroadNet Gemma 4 VL 26B run?
BroadNet's model page says: NVIDIA Blackwell-class GPUs with a hot backup; failover is automatic. Its privacy statement adds that processing locations for data-residency needs are agreed with BroadNet before you send data.
How does BroadNet Gemma 4 VL 26B score on benchmarks?
As published, the Google model card reports 82.6% on MMLU Pro and 82.3% on GPQA Diamond; Epoch AI measured 0.732 on GPQA Diamond in its own run, as a mean score on a 0 to 1 scale (checked 2026-11-16). BroadNet did not run these benchmarks, and sources may use different settings, so the figures are not directly comparable. Every published result we found, with its setting and source.
Is my data stored?
Inference content is not retained: prompts, images and answers are processed for the request only. Necessary account, billing and security metadata, such as timings, token counts and cost, is retained as described in BroadNet's privacy statement. Read how.