Primary use stated or inferred cautiously from official documentation.
Tencent Hy4 Flash Base
Pretrained checkpoint in the Hy4 Flash family, published in the official repository on November 17. It retains the visual encoder and MoE architecture but is distinct from the instruction-tuned model and its FP8 format.
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
Limits and integration
| API ID | tencent/Hy4-flash-base |
|---|---|
| Model type | Multimodal · Text and reasoning |
| Access model | Downloadable weights |
| License | Tencent Hy License |
| Deployment | Self-hosted |
| Release | 17 NOV 2026 |
| Knowledge cutoff | Not published |
| Entrance | Text · Image |
| Exit | Text |
| Context window | 1,048,576 tokens in the weights configuration |
| maximum output | Text; generation maximum not specified for the base checkpoint |
| Reasoning | Not published |
| Published tools | |
| Structured Outings | Not published |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face |
Research and adaptation of a base model with text and image input. The release includes weights and configuration. Pretraining measurements use few-shot protocols and must remain separate from assistant evaluations.
The assistant’s instructions, tools and results are not inherited. The Tencent Hy license has its own conditions and restrictions; it is not Apache-2.0. Deployment needs substantial infrastructure and a compatible runtime.
What it can do
Orientation
Research and adaptation of a base model with text and image input. The release includes weights and configuration. Pretraining measurements use few-shot protocols and must remain separate from assistant evaluations.
Context
1,048,576 tokens in the weights configuration · Text; generation maximum not specified for the base checkpoint
Tools and integration
Not published
Access
Hugging Face · downloadable checkpoint
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | No published official API rate | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Self-hosted infrastructure cost | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
No official dedicated API price has been verified. Downloading weights does not eliminate compute, memory or operating costs. TokenHub prices for hy4-flash apply to the assistant service, not this base checkpoint.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
How to read the results
The pretraining table reports 71.1% on MMLU-Pro with five examples and macro-averaged accuracy. The figure is for Hy4 Flash Base, not the Instruct model or a test reproduced by Inferama.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| MMLU-Pro (Hy4 Base 5-shot) | 71.1% | Macro accuracy · 5-shot | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
Chronology
tencent/Hy4-flash-base
Pretrained checkpoint published; revision 95b9f823 verified on November 18.