Tencent Hy4 Flash: 224B-Parameter Multimodal Model Priced Below Claude Haiku 5.5
LatentMoE and Engram cut KV cache to one-sixth of Hy4's; Tencent benchmarks lack independent verification

Tencent released Hy4 Flash on November 16, 2026, pricing the model at $0.070 per million input tokens and $0.281 per million output tokens — cheaper than both Claude Haiku 5.5 and Gemini 2.5 Flash, and roughly 92% and 89% less expensive than the full Hy4 it was built to complement. With a one-million-token context window and native support for text, images, audio, and video, the model is intended for the high-volume everyday tasks that make up most enterprise AI workloads.
Flash-Tier Positioning: Speed and Scale Over Peak Intelligence
Hy4 Flash is not Tencent's most capable model. It is Tencent's cheapest one to run, and that is deliberate. The company positions the Hy model family with two tiers: the full Hy4 for tasks that require high reasoning quality, and Flash for tasks where request volume and cost per unit outweigh the need for peak intelligence. Document summarization, presentation generation, spreadsheet repair, audio transcription, and visual understanding — the category Tencent groups under its WorkIndex benchmark — are cited as the intended use cases.
The model is live on Tencent Cloud's TokenHub API marketplace under the identifier hy4-flash, and also embedded in Tencent's own productivity product WorkBuddy, its assistant app Yuanbao, enterprise communication product WeCom, QQ, and the ima knowledge-management product. An OpenRouter listing is coming soon. Pricing through November 30 is discounted 50% further, with WorkBuddy access free for the same period.
How Engram Parameters and LatentMoE Reduce KV Cache
The architecture behind Hy4 Flash explains why a 224-billion-parameter model can be served cheaply enough to undercut flash-tier competitors.
The 224 billion figure covers three parameter types: standard transformer parameters, Mixture-of-Experts routing parameters, and what Tencent calls Engram parameters — 64 billion of the total. Tencent describes Engram as a memory-style component, though no architecture paper has been published. What the company does document is the deployment consequence: Engram parameters are not part of the per-token activation path. During inference, only 12 billion parameters are active on each token, so the effective compute cost per token is determined by that 12B active count rather than the 224B total.
The second mechanism is LatentMoE, Tencent's label for its mixture-of-experts design. Standard dense transformers activate all parameters for every token. MoE architectures route each token to only a subset of expert modules, increasing total model capacity without a proportional increase in per-token computation. Tencent's design combines LatentMoE routing with hybrid attention — interleaving standard multi-head attention with other mechanisms — in a way the company says cuts the KV cache to one-sixth the size of Hy4's and one-twenty-fourth the size of Hy3's. KV cache is the memory a serving system must hold per request to allow attention over previously generated tokens; a smaller cache means more concurrent sessions per GPU and lower memory bandwidth per request. That reduction is what drives the serving economics that make the pricing tier possible.
Post-training adds another efficiency layer. Hy4 Flash was trained using self-distillation from the full Hy4 — compressing its reasoning patterns into the smaller active footprint — followed by reinforcement learning that cut average output token count by approximately 50%. The model tends to give shorter, more direct answers than its predecessor, reducing both inference latency and the cost per request under output-priced billing. This explains why the per-task cost figures Tencent cites diverge substantially from the list price ratio alone.
The inference stack uses megakernel optimization, 4-bit quantization, and an internal system Tencent calls DFly. The company claims sustained decoding throughput above 150 tokens per second.
Where Flash Sits in Flash-Tier Pricing
At $0.070 input and $0.281 output per million tokens, Hy4 Flash undercuts the main competitive flash-tier models on output pricing. Claude Haiku 5.5 is listed at $0.10 input and $0.50 output for prompts under 100,000 tokens, making Hy4 Flash 30% cheaper on input and 44% cheaper on output at list prices. Gemini 2.5 Flash charges $0.30 input and $2.50 output, making Hy4 Flash substantially cheaper on both. Gemini 2.5 Flash-Lite is closer at $0.10 input and $0.40 output, with Hy4 Flash still cheaper on both dimensions. DeepSeek V4.1 Flash, the nearest DeepSeek flash-tier competitor, is priced at $0.30 input and $1.20 output at peak on its first-party API, with an off-peak rate of $0.15 input and $0.60 output.
The multimodal capability matters at this price point. DeepSeek V4.1 Flash supports text and image inputs but not audio or video processing natively, and Gemini 2.5 Flash-Lite's multimodal capabilities are more constrained than Gemini 2.5 Flash's full feature set. Hy4 Flash processes text, images, audio, and video natively within the same model and within the same one-million-token context window, which allows mixed-modality enterprise tasks — transcribing a meeting recording, then summarizing the attached slides, then drafting a follow-up document — to run against a single endpoint at a consistent pricing rate.
What Tencent Claims, and What Remains Unverified
Tencent published two categories of performance evidence with the launch. The first is WorkIndex, an internal benchmark measuring performance on office productivity tasks: slide creation, document editing, spreadsheet repair, audio and video processing, and visual understanding. On WorkIndex, Tencent reports Hy4 Flash scores slightly above Hy4 Preview — the open-weight variant released in August — while costing one-thirteenth of Hy4 Preview and one-twentieth of full Hy4 per task, according to the company's calculation.
The second is a blind satisfaction test conducted using 85 internal WorkBuddy users who evaluated 259 tasks across 27 categories. The company's research page describes the result as Hy4 Flash performing "close to" Claude Haiku 5.5; marketing materials characterize this more strongly as matching Haiku 5.5. No external organization has published independent benchmark scores for Hy4 Flash as of its release date, and an 85-person internal test represents a narrow basis for a head-to-head capability claim against Anthropic's current small-model tier.
The 18x per-task cost advantage Tencent cites over Haiku 5.5 also requires scrutiny. The list price difference between the two models is roughly 1.4x on input and 1.8x on output. The gap to 18x is driven primarily by shorter output token counts from RL training — meaning fewer tokens billed per task — rather than by pricing alone. That savings depends on the model generating shorter, complete, accurate responses, and there is currently no independent evaluation of output quality at reduced length. The figure also does not account for Haiku 5.5's higher per-token pricing for longer prompts, which Anthropic tiers above 100,000 tokens.
Open-Source Plans, Undocumented Infrastructure, and What Comes Next
Tencent released preview weights for the full Hy4 model under Apache 2.0 license in August 2026. No Flash weights are currently available, and Tencent has stated only that open-sourcing is planned without committing to a date. No model card, safety disclosure, or formal architecture paper has been published for Hy4 Flash. The DFly inference system remains undocumented beyond its name.
The November 30 promotional pricing deadline creates a compressed evaluation window for developers and enterprises comparing Hy4 Flash against existing flash-tier providers. The model's real competitive standing will become clearer as independent evaluators benchmark its multimodal capabilities, output coherence at reduced token counts, and long-context stability — the dimensions WorkIndex and an internal satisfaction survey cannot fully address. Whether the Engram parameter architecture represents a novel memory mechanism or a marketing label for a known technique remains an open question until Tencent publishes the supporting technical material.
Enterprise buyers evaluating Tencent Cloud API endpoints should note that Tencent operates under Chinese regulatory jurisdiction, including data-access obligations established by China's Cybersecurity Law and Data Security Law. Developers routing workloads through Tencent Cloud TokenHub should review the platform's data-handling terms and applicable residency requirements for their deployment context.



