Yes, if…
You run a lot of short-to-medium requests, keep reusing the same system prompt, or already pay a frontier model for work a 12B-active model handles fine. The cache tier makes the first and second case especially cheap.
Hy4 Flash is Tencent's cheap-and-fast member of the Hy4 family: a 224B mixture-of-experts model with 12B active parameters, a 1M-token context and multimodal input. On TokenRa it costs $0.056 per 1M input tokens and runs on an OpenAI-compatible endpoint, so an existing OpenAI SDK integration only needs a new base URL, key and model ID.
Hy4 Flash is priced per token with no monthly minimum. TokenRa bills every tier below Tencent Cloud's direct list price, so the same model costs less through one key.
| Tier | TokenRa | Tencent Cloud list | You save |
|---|---|---|---|
| Input | $0.056 / 1M | $0.070 / 1M | 20% |
| Output (completion) | $0.225 / 1M | $0.281 / 1M | 20% |
| Cached input read | $0.0024 / 1M | $0.003 / 1M | 20% |
Prices are per 1 million tokens in USD. Tencent Cloud figures are the vendor's published list price at the time of writing; TokenRa figures are what your key is billed. Check the live model list for the current rate before committing to a budget.
A typical RAG or document-QA job sends 10M input tokens and returns 2M output tokens in a month:
Long system prompts are where the cache tier pays off. Cached reads are billed at 4.3% of the input rate ($0.0024 ÷ $0.056), so a 50K-token instruction block sent 1,000 times costs about $2.80 uncached and roughly $0.12 when it hits the cache — a saving of more than 95% on that block alone.
Keep your static system prompt and tool definitions at the start of every request, unchanged, so they qualify for cache reads. Put the volatile part — the user's document, the retrieved chunks — after it.
Tencent shipped Hy4 Flash on 16 November 2026 as the lightweight option next to the 770B-class Hy4 flagship. The architecture is sparse: 224B total parameters, of which 12B are active per token. That is why it is fast and cheap without collapsing to a small model's reasoning quality.
| Spec | Hy4 Flash |
|---|---|
| Total parameters | 224B (MoE) |
| Active per token | 12B |
| Context window | Up to 1M input tokens |
| Max output | About 64K tokens per request |
| Input modalities | Text, images |
| Tool calling | Function tools, JSON and JSON-schema output |
| Released | 16 November 2026 |
| Licence | Open release announced, weights date not yet published |
Hy4 is the flagship (770B-class, open weights) and costs several times more per token. Hy4 Flash is the volume workhorse. If you arrived looking for the flagship's pricing, the Hy4 API page has it; Hy4 Preview docs cover the earlier preview build.
The model ID is hy4-flash. Because channel naming occasionally differs from the marketing name, confirm it once against the model list your key can see:
curl https://tokenra.io/v1/models \
-H "Authorization: Bearer $NEW_API_KEY"
Then call it exactly as you would call GPT or Claude — same request shape, same SDK:
curl https://tokenra.io/v1/chat/completions \
-H "Authorization: Bearer $NEW_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "hy4-flash",
"messages": [
{
"role": "user",
"content": "Explain quantum entanglement in one paragraph."
}
],
"temperature": 0.7
}'
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["NEW_API_KEY"],
base_url="https://tokenra.io/v1",
)
resp = client.chat.completions.create(
model="hy4-flash",
messages=[
{
"role": "user",
"content": "Explain quantum entanglement in one paragraph.",
},
],
temperature=0.7,
)
print(resp.choices[0].message.content)
print(resp.usage) # prompt_tokens / completion_tokens / total_tokens
Hy4 Flash takes image input the same way the OpenAI vision models do — pass a URL or a base64 data URI as an image_url content part:
resp = client.chat.completions.create(
model="hy4-flash",
messages=[{
"role": "user",
"content": [
{"type": "text", "text": "What is wrong with this dashboard chart?"},
{"type": "image_url", "image_url": {"url": "https://example.com/chart.png"}},
],
}],
)
Function calling works on the standard tools array, and response_format accepts {"type": "json_object"} or a full json_schema when you need guaranteed shape. Anything written for gpt-4o-mini or deepseek-chat in your codebase should run on Hy4 Flash after swapping the model string.
The exact model ID returned by /v1/models (this page uses hy4-flash), and the maximum output tokens your plan allows — public listings say about 64K, but the console is the authority for your account.
These three get compared constantly because they occupy the same price band. The honest answer is that they are built for different jobs, not that one wins.
| Model | Size | Context | Pick it when |
|---|---|---|---|
| Hy4 Flash | 224B MoE, 12B active | 1M | High-volume text and image work where cost per token dominates: extraction, summarisation, bulk translation, routine code. |
| Hy4 | 770B-class flagship | 1M | Long-horizon agentic coding and hard reasoning, where a wrong answer costs more than the token bill. |
| DeepSeek V4.1 Flash | ~304B MoE | 1.31M | You want a slightly larger active footprint and the longest context, and your stack already leans DeepSeek. |
A practical rule: prototype on Hy4 Flash, and route requests to Hy4 only when the task fails a check you care about — a unit test, a schema validation, a human review. In most pipelines that keeps 80–90% of traffic on the cheap tier without any visible quality drop.
You run a lot of short-to-medium requests, keep reusing the same system prompt, or already pay a frontier model for work a 12B-active model handles fine. The cache tier makes the first and second case especially cheap.
Your workload depends on native reasoning benchmarks at the frontier, or you need self-hosted weights today — the open release is announced but not dated.
$0.056 per 1M input tokens, $0.225 per 1M output tokens and $0.0024 per 1M cached input reads. That is 20% below Tencent Cloud's direct list of $0.070, $0.281 and $0.003.
hy4-flash, as used throughout this page. Confirm the exact string with a GET /v1/models call on your key before hard-coding it into production.
Up to 1M input tokens. Output is capped at roughly 64K tokens per request on public listings — check the console for the limit on your key.
Tencent announced an open release, but no confirmed weights date had been published as of 17 November 2026. Until then, use the hosted API.
Both. Images go in as image_url content parts; tools go in the standard tools array, with JSON and JSON-schema output available through response_format.
Yes. Point base_url at https://tokenra.io/v1, use your TokenRa key, and set the model to hy4-flash. Nothing else in the request changes.