Tencent's Hy4 flash packs 224B total parameters with 12B active, matches Claude Haiku 5.5 on daily tasks at 1/18th the cost
-
Tencent released Hy4 flash , a cheap MoE model with 224B total and 12B active parameters.
-
API pricing: $0.003 cached, $0.070 input, $0.281 output per 1M tokens.
-
Claims 18x task cost reduction versus flagship Hy4 while matching Claude Haiku 5.5 on daily productivity.
-
Uses Engram, a 64B-parameter memory lookup module that caches reusable computations to cut inference cost.
-
Inherits full Hy4 feature set: 1M context, vision understanding, and computer-use agent capabilities.
-
Positioned against DeepSeek-V4.1-Flash, Qwen3.8-Flash-Next, GLM-5.3-Flash, and GPT-6 Luna in the flash tier. Tencent’s Hunyuan team has released Hy4 Flash, a lower-cost sibling of its flagship Hy4 model. The company says the new model retains multimodal input, agent capabilities and a one-million-token context window while targeting high-volume workloads such as chat, retrieval-augmented generation, document processing and routine code changes.
224B total, 12B active #
Hy4 Flash uses a mixture-of-experts architecture with 224 billion total parameters, including 64 billion Engram parameters, while activating 12 billion parameters for each token. Mixture-of-experts models route tokens through selected components, allowing total model capacity to exceed the computation used for each token.
| Specification | Hy4 Flash |
|---|---|
| Total parameters | 224B |
| Engram parameters | 64B |
| Active parameters per token | 12B |
| Context window | Up to 1M tokens |
| Inputs | Text and images |
| Agent capabilities | Tool use and computer interaction |
Prices built for volume #
Tencent lists three API rates, each calculated per one million tokens:
| Token type | Price per 1M tokens |
|---|---|
| Cached input | $0.003 |
| Uncached input | $0.070 |
| Output | $0.281 |
At those rates, a request containing one million uncached input tokens and one million output tokens costs $0.351. Billing all input as cache hits reduces the same request to $0.284, with generated tokens accounting for most of the remaining cost.
The company says uncached input, output and cached input cost about one-twelfth, one-ninth and one-fifteenth of Hy4’s corresponding rates. Tencent also advertises an 18-fold reduction in per-task cost. A fixed token mix cannot produce that figure from the three component ratios alone, so the task-level estimate requires a different baseline, altered token usage or another unstated assumption.
Where Tencent says it lands #
Tencent positions Hy4 Flash for everyday productivity work. The company reports that it matched Claude Haiku 5.5 in high-reasoning mode during human blind tests covering daily tasks. Reproducing that result requires the underlying prompts, sample size, rater protocol and scoring criteria, followed by evaluation on each team’s production data.
The vendor’s cost-versus-performance chart places Hy4 Flash alongside DeepSeek-V4.1-Flash, Qwen3.8-Flash-Next, GLM-5.3-Flash and GPT-6 Luna. The same chart assigns Hy4 Flash a score of 42.1 and the flagship Hy4 a score of 51.4, indicating that Tencent still expects the larger model to handle demanding reasoning and generation workloads.
| Workload | Likely model choice |
|---|---|
| Chat, classification and email drafting | Hy4 Flash |
| RAG answers grounded in supplied documents | Hy4 Flash |
| OCR, chart reading and interface parsing | Hy4 Flash |
| Routine code edits and refactoring | Hy4 Flash |
| Deep multi-step reasoning | Hy4 |
| Long-horizon planning and complex code generation | Hy4 |
Engram and prompt caching do different jobs #
Engram is a learned memory module inside the model. According to Tencent’s description, it retrieves representations for recurring token patterns through table lookups, reducing dense computation for those patterns. Engram and mixture-of-experts routing together help the model combine 224 billion total parameters with 12 billion active parameters per token.
An API prompt-cache hit is a separate service-layer mechanism. The discounted rate applies when Tencent’s platform recognizes input that qualifies for reuse, such as repeated prompt prefixes, system instructions, tool schemas or document sections.
Reliable budget forecasts require the provider’s cache eligibility rules, retention period, prefix-matching behavior, minimum token count and any cache-write charges. The quoted launch prices establish the hit rate but do not define those operating terms.
The flagship feature set carries over #
Hy4 Flash follows the flagship Hy4, which Tencent introduced with vision understanding and computer-use capabilities. The company says Flash inherits those features along with the one-million-token context window.
- Long documents: question answering and summarization within the available context budget.
- Vision: OCR, screenshot analysis, chart interpretation and interface parsing.
- Computer use: screen interpretation and action selection for agent loops.
- Tool use: calls to application-defined functions and external services.
- Code: routine edits, refactoring and lightweight generation.
A one-million-token context window describes maximum capacity, not guaranteed accuracy across an entire request. Images, tool definitions, retrieved documents and generated output may also consume the available budget, depending on Tencent’s API accounting. Computer-use deployments additionally require an execution harness, permission controls and safeguards around external actions.
Production questions remain #
- API contract: model identifiers, endpoint availability, SDK compatibility, structured outputs and tool-call schemas.
- Performance: time to first token, generation speed, concurrency limits and sustained throughput.
- Limits: maximum output length, image constraints, rate limits and context allocation.
- Operations: regional availability, uptime commitments, retries and observability.
- Data handling: retention, training policies, residency options and compliance terms.
- Quality: internal evaluations for hallucination rates, tool accuracy, long-context retrieval and code correctness.
The economics favor repetitive workloads #
At the posted rates, a monthly workload with 100 million input tokens and 20 million output tokens would cost $12.62 with no input caching. If 80 million input tokens qualified for the cache-hit rate, the token bill would fall to $7.26.
Those estimates exclude platform fees and assume every designated token qualifies for the stated rate. Hy4 Flash’s clearest fit is high-volume work with predictable prompts and moderate reasoning demands. Internal quality tests, cache behavior and service limits will determine whether the advertised token prices translate into lower production costs.