Remember when “cheap AI” meant hoping your credit card survived a conversation with a frontier model? Those days are officially over. Tencent just dropped Hy4 Flash, and its pricing sheet reads like a dare to every AI company currently charging you per token like they’re minting gold.
Here’s the headline: a 224-billion-parameter model with only 12 billion active parameters per token, priced at $0.070 per million input tokens and $0.281 per million output tokens. For context, that’s roughly 92% cheaper on input and 89% cheaper on output than its bigger sibling, the full Hy4. Cached input? A laughable $0.003 per million tokens.
The “revolutionary” move here isn’t just the price tag, it’s that Tencent claims this thing outperforms its own 770B-parameter Hy4 Preview on everyday office tasks. Let’s dig into how they pulled that off and what it means for anyone building on top of these models.
The Architecture: Smart Sparsity Meets Clever Memory
Tencent’s approach combines three architectural techniques that deserve a closer look: hybrid attention, LatentMoE, and something called Engram parameters.
Of those 224 billion total parameters, 64 billion are Engram parameters, a conditional memory technique that stores frequently occurring text patterns in a fast-lookup table. Instead of recomputing common patterns, the model retrieves them. It’s a practical hack that saves compute without sacrificing much quality.
The MoE design means only 12 billion parameters activate per token. That’s the number that determines your inference cost, not the headline 224B. For comparison, the full Hy4 activates 49 billion parameters per token. Flash is using roughly a quarter of the compute per token while claiming comparable performance on everyday tasks.
The KV cache story is arguably more important than the parameter count. Tencent reports the KV cache size is just 1/24 of Hy3 and 1/6 of Hy4. For a 1M-token context window, that’s the difference between a model that fits on one machine and one that requires a small cluster. This is the kind of optimization that quietly decides whether a model is economically viable at scale.
The Blind Test: Claims vs. Reality
Hy4 Flash scored nearly identical to Claude Haiku 5.5 on satisfaction while costing roughly 18x less per task. But let’s be honest about what this does, and doesn’t, tell us.
These are internal evaluations on Tencent’s own product surface with Tencent’s own users. They’re useful signals about relative cost, not independent capability rankings. The model trails DeepSeek’s Flash variant on satisfaction, which is worth noting.
More importantly, the model falls flat on harder benchmarks. On Terminal-Bench 4.0, Hy4 Flash scores 19.2 against 44.4 for the full Hy4. On ProgramBench, it’s 16.0 versus 30.0. The gap is real and significant.
The Hard Numbers: Public Benchmark Performance
Looking at the LLMLearner benchmark data, Hy4 Flash holds its own in specific niches:
| Benchmark | Hy4 Flash | Hy4 | DeepSeek V4.1 Flash |
|---|---|---|---|
| HLE | 55.2 | 59.1 | 58.8 |
| MMMU-Pro | 75.8 | 78.6 | 82.5 |
| DeepSWE | 64.6 | 74.9 | 77.0 |
| APEX-Agents | 63.1 | 75.2 | 63.4 |
| AutomationBench | 52.2 | , | , |
| OSWorld 2.1 | 68.1 | 78.9 | 69.3 |
Notice the pattern: Hy4 Flash is competitive on agentic and office-automation benchmarks (APEX-Agents, AutomationBench) but falls behind on complex coding and reasoning. Its AutomationBench score of 52.2 places it 8th out of 62 models, that’s genuinely impressive for a budget model.
Tencent’s positioning is refreshingly honest: use Hy4 Flash for everyday tasks, use the full Hy4 for long-horizon software engineering. The company isn’t claiming Flash is a frontier model, it’s claiming Flash is good enough for 80% of what people actually use AI for.
What the “Flash Model Race” Actually Means
Over the past two weeks, Chinese labs have moved in lockstep. Alibaba shipped Qwen4-Max with 2.4T parameters. DeepSeek released its V4.1-Flash API suite. Baidu pushed ERNIE 5.5 Flash Preview. And Google launched Gemini 4 Hydrogen at $1.50 per million tokens, which suddenly looks expensive by comparison.
The pattern is clear: after the flagship race, the second race is the efficiency race. Every major lab is shipping “flash” variants that trade a slice of capability for a massive cut in cost and latency.
The real question isn’t whether Hy4 Flash beats the flagship. It doesn’t, and Tencent isn’t claiming it does. The question is whether your workload, drafting, summarizing, editing spreadsheets, triaging long documents, is close enough in quality at one-twentieth the cost.
The WorkIndex: Benchmarks That Actually Mean Something
One of the most interesting details here is Tencent’s decision to build WorkIndex, an internal benchmark grounded in real WorkBuddy tasks. It covers eight evaluation sets: slide creation, slide editing, human-AI slide co-writing, Markdown human-AI co-writing, document editing, spreadsheet repair, audio/video processing, and visual understanding.
This is notable because public benchmark scores often fail to predict real-world utility. A model that crushes MMLU might produce verbose, unfocused output on actual office tasks. Tencent is measuring deliverable quality, single-turn output length, and task turnaround time, the things that actually matter for productivity.
On WorkIndex, Hy4 Flash scores 42.1 versus 39.7 for the 770B Hy4 Preview. That’s a smaller model outperforming its bigger sibling on the tasks that count for office work.
The Cost Breakdown: Understanding the Real Savings
Tencent’s pricing table tells a compelling story:
| API Pricing (per million tokens) | Cached Input | Input | Output |
|---|---|---|---|
| Hy4 Flash | $0.003 | $0.070 | $0.281 |
| Hy4 | $0.042 | $0.834 | $2.501 |
The output token price is the one that matters most for agentic workloads, that’s where your costs balloon when models write excessively. Tencent says reinforcement learning cut average end-to-end output tokens by 50%. Combined with the lower per-token rate, that’s how they arrive at a 20x per-task cost reduction.
On the internal WorkIndex benchmark, average cost per task is:
– 1/13 of Hy4 Preview
– 1/20 of full Hy4
– 1/8 of DeepSeek-V4.1-Flash-1017 (Max tier, peak-hour pricing)
There’s a caveat for developers: these task-level savings depend on the model outputting shorter, more efficient responses. If your application needs verbose output, your actual cost ratio will differ from Tencent’s claims.
Real-World Integration Results
Tencent deployed Hy4 Flash across seven product surfaces before announcing it publicly. The results from these deployments offer the most honest picture of the model’s capabilities:
WeCom’s Dayuan assistant saw output speed increase 88% over Hy4 while cutting output tokens in thinking mode by roughly 60%.
Yuanbao’s agent workflows showed a 13.6% performance boost over Hy3 and a 7.56% gain in human side-by-side evaluations.
Honor of Kings’ game R&D bug analysis matched DeepSeek-V4.1-Flash-1017 in root-cause localization while cutting model requests by 22-39% and token consumption by 40.5-61.2%.
WeChat Pay’s real-device agent evaluations matched or exceeded DeepSeek across all four measured workflows while cutting execution steps by ~6% and single-turn output length by ~29%.
These are production deployments with measurable improvements, not just benchmark chest-thumping.
The Open-Weight Elephant in the Room
Tencent plans to release Hy4 Flash as an open-weight model, but hasn’t provided a date or license. For now, it’s API-only, which means:
- No local verification of Tencent’s claims
- No ability to run it on your own hardware
- No transparency about the Engram architecture’s details
The community speculation about local performance is interesting. Some estimates suggest a 4-bit version could fit in roughly 130GB and run on hardware like two NVIDIA DGX Spark servers, but these are educated guesses. There’s no llama.cpp support yet, and the architecture’s hybrid attention may complicate porting.
Some Reddit commentators have noted the Engram ratio, 64B out of 224B total parameters, is precisely 28.5%, matching what the original DeepSeek Engram paper suggested as optimal. Whether that’s intentional or coincidental, it suggests Tencent is following a well-researched playbook.
What This Means for Your AI Budget
For builders and enterprises, Hy4 Flash represents something genuinely new: an economically viable model for high-volume workloads.
If you’re currently using a frontier model for basic document processing, research, or office automation, running those tasks on Hy4 Flash could cut your inference bill by an order of magnitude. The quality tradeoff on routine tasks is minimal, Hy4 Flash matches Claude Haiku 5.5 on user satisfaction while costing roughly 18x less per task.
For teams building agentic systems, the math is even more compelling. Agents make multiple model calls per task, so per-task costs compound. A 20x reduction in per-task cost could be the difference between an agent that’s economically viable and one that burns through your budget.
The model supports the OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages formats, making it straightforward to swap into existing integrations. The API identifier is simply hy4-flash on Tencent Cloud TokenHub.
Tencent is offering a 50% discount on API pricing through November 30, and WorkBuddy access is free during the same period.
Hy4 Flash isn’t a frontier model, it’s a frontier economic model. It’s Tencent’s bet that most daily AI work deserves a cheaper runtime than the flagship race has been demanding. And when you look at the production deployment results from WeCom, Yuanbao, and WeChat Pay, that bet looks justified.
The tradeoff is straightforward: you sacrifice capability on complex coding and hard reasoning, but you gain dramatically lower costs for the everyday tasks that occupy most of your usage. For teams running high-volume workloads at scale, that trade is worth serious consideration.
If your workload is mostly routine, drafting, summarizing, editing, triaging long documents, Hy4 Flash deserves a place in your evaluation. And if you’re already running agentic systems, the per-task cost reduction might be the optimization you’ve been looking for.
Tencent’s previous breakthrough with the Hy4 model showed what’s possible with open weights. Hy4 Flash extends that story with a new twist: intelligence isn’t just about size, it’s about deploying the right size for the right task at the right price.
The inference cost wars aren’t over, they’re just getting interesting. And for the first time, budget-conscious builders are holding the winning hand.




