Tencent released the junior model of its Hy4 family — Hy4 flash — on November 16. Instead of scaling up dense compute blocks, the developers opted to separate working memory and active weights, directly cutting server resource costs when processing incoming data. This approach solves a practical problem: how to make million-token contexts affordable for everyday services without multiplying compute bills. The introduction of an external lookup table alongside redesigned attention mechanisms drastically slashes hardware overhead during inference while maintaining an acceptable baseline for standard transactional operations, which is critical for the unit economics of high-frequency request services.

Tables Instead of Neurons

The idea of a lookup table structure itself was proposed by the DeepSeek laboratory. In this scheme, addresses are known in advance, and for each token, the model reads several rows from the table.

The Engram architecture offloads 64 billion parameters into a lookup table with multi-row reads per token, enabling queries to standard RAM rather than expensive VRAM.

Beyond tabular memory, the architecture relies on hybrid attention and LatentMoE. Because of them, the KV cache of Hy4 flash is six times smaller than that of Hy4 and 24 times smaller than that of Hy3. This is crucial for long contexts, as the cache is precisely what eats up memory at a million tokens, while Tencent estimates decoding speeds above 150 tokens per second.

Test Results and System Limitations

It is generally assumed that such aggressive pruning of active computations inevitably destroys overall generation quality in practical scenarios. In reality, internal blind test data demonstrated system stability: 85 users of the WorkBuddy corporate application tested 259 tasks across 27 distinct categories evaluated on a scale of up to 4 points, providing an objective view of the model's applicability in real business processes.

In the WorkBuddy blind test across 259 tasks, Hy4 flash delivered quality on par with Claude Haiku 5.5 while cutting the execution cost of every task by 18 times.

Furthermore, the system nearly matched Anthropic's Claude Haiku 5.5 and outperformed Zhipu AI's GLM-5.3-Flash, while running tasks at 18 times lower cost than Anthropic's solution. However, this cost reduction has hard limits. On long programming tasks, the gap with the older model is noticeable: in Terminal-Bench 4.0, Hy4 flash scored 19.2 points, compared to 44.4 for the full Hy4.

The Hy4 flash model is already available in the Selectel AI router with an OpenAI-compatible API and per-token billing. Because the creators have not yet open-sourced the weights, the remote API remains the sole integration path. For standard document processing, search, and interface operations, the architecture sets a new price benchmark, but critical development tasks still require full-scale systems because architectural compromises inevitably damage specialized model capabilities.

For comparison

Cutting the server expenses of processing million-token contexts no longer requires buying ever-larger compute clusters, as shown by Hy4 flash offloading memory into lookup tables. Architectural savings on memory cache solve the unit economics of high-frequency workflows while leaving advanced engineering tasks to heavy, full-weight systems.

Large Language ModelsModel InferenceCost ReductionInference Cost