Tencent Hunyuan released Hy4 flash on November 16, 2026, and opened three weights on November 18 — the full model, an FP8 variant, and the base checkpoint — across GitHub, Hugging Face, ModelScope and GitCode. The model topped the OpenRouter daily ranking on its first day of availability. Tencent reports first-day token volume above 50 trillion, a record for the Hunyuan series.
The headline specification is unusual. Total parameters sit at 224B, but only about 12B activate per token. The remainder is largely an N-gram embedding table rather than dense compute. That restructuring is what makes a 12B-per-token activation budget deliverable on 8×A100 or H100 hardware, and what puts per-task cost at ¥0.26 in Tencent's internal measurements.
What Hy4 flash Is, and What It Is Not
Hy4 flash is not a scaled-down Hy4. Tencent positions it as a knowledge-distilled derivative of the Hy4 flagship, rebuilt around cost efficiency for high-frequency daily tasks rather than parameter-count reduction.
Three architectural changes separate it from the flagship, per the official GitHub README dated November 18, 2026.
| Change | Mechanism | Effect |
|---|---|---|
| Hybrid Attention | Sliding-window attention (512-token window) interleaved with IndexPool sparse attention (up to 2,048 tokens) in a 1:3 ratio | KV cache compressed to 1/6 of Hy4, 1/24 of Hy3 |
| LatentMoE | Router experts computed in a 1,536-dim low-rank space instead of the full 3,072-dim space | 16 of 512 routed experts plus 1 shared expert active per token, ≈11.3B text + 0.43B vision |
| Engram N-gram embedding | Multi-head hash tables storing bi-gram and tri-gram representations, looked up from the last two or three tokens and injected into hidden states | 63.4B embedding parameters, concentrated in layers 1/11/21/31, excluded from inference matmuls |
Two details in that table carry most of the engineering weight.
KV cache is the deployment bottleneck, and it drops by a factor of six. Hybrid Attention splits the work: sliding-window attention handles adjacent-token relations inside a 512-token window, while IndexPool sparse attention reaches into a wider span and selects up to 2,048 tokens. The two alternate at a 1:3 ratio. Attention cost scales with the sum, not the product, and the compression applies to every request regardless of batch size.
Engram parameters cost memory but not compute. The N-gram component accounts for 63.4B parameters, the single largest contributor to the 224B total. It does not participate in matrix multiplication during inference. Lookup and injection are cheap, which keeps activated parameters low even as stored parameters grow. The design trades weight footprint for compute, and on modern accelerators that ratio is favorable.
The combined effect shows up on WorkIndex, Tencent's internal evaluation covering eight office productivity task categories — slide generation, document editing, audio and video processing, spreadsheet repair. Hy4 flash scores 42.1 against Hy4 preview's 39.7, with Hy4 at 51.4 as the upper reference. Per-task cost is 1/13 of Hy4 preview and 1/20 of Hy4.
Benchmark Data Behind the OpenRouter Placement
Ranking placement on OpenRouter reflects what developers actually chose to call, not what a leaderboard model scored. The figures below come from Tencent's disclosure and independent evaluation sets, not a single ranking.
User blind test. 259 tasks across 27 task categories, run with 85 internal WorkBuddy users.
- Satisfaction: 2.79 out of 4
- Median end-to-end completion time: 6.7 minutes
- Cost per task: ¥0.26
- Source: Tencent Hunyuan official GitHub, 2026-11-18
WorkIndex internal evaluation, eight office productivity tasks.
- Hy4 flash: 42.1
- Hy4 preview: 39.7
- Hy4 (upper reference): 51.4
Public benchmarks, few-shot, base-model protocol.
- MMLU (5-shot): 88.7 — Hy4 Base 91.6, DeepSeek-V4-Pro-Base 91.1
- MMLU-Pro (5-shot): 71.1 — Hy4 Base 73.9
- C-Eval (5-shot): 92.0 — Hy4 Base 95.0
- U-MATH (5-shot): 69.8 — Hy4 Base 71.6
One caveat belongs with these numbers. They are few-shot results on the base model, not on the instruction-tuned release. Tencent's README states that the base model was exposed to large volumes of synthetic instruction data, so it does not always follow the expected answer format in few-shot settings. Scores may therefore understate the tuned model's real behavior. For everyday use the relevant references are WorkIndex and the blind test, both of which measure the instruction-tuned Hy4 flash.
Inference supports two reasoning modes: reasoning_effort: "high", the default with thinking enabled, and reasoning_effort: "no_think", which emits output directly without generating reasoning tokens and suits latency-sensitive paths.
Local Inference With Transformers
Transformers support is not yet merged into the main branch. Install from the specified PR:
A low-latency text example in no_think mode:
Recommended sampling parameters: temperature=0.9, top_p=1.0. Swap reasoning_effort to "high" when a reasoning chain is wanted.
Image input follows the same message structure, with an image content block carrying a URL alongside a text question. The vision path adds 0.43B activated parameters, negligible against the text budget.
Production Deployment: vLLM and SGLang
Both runtimes ship official prebuilt images with DFLY2 speculative decoding and FlashMLA sparse attention integrated. FP8 quantization (Hy4-flash-FP8) reduces memory footprint and targets 8×A100 or H100 servers. Local testing can start from the GGUF build (AngelSlim/Hy4-flash-GGUF).
vLLM
Three flags matter. --speculative-config '{"method":"dfly"}' enables DFLY2 speculative decoding using the built-in 0.85B draft model, which adds no extra VRAM. --attention-backend FLASHMLA_SPARSE activates the sparse attention kernel. --tensor-parallel-size 8 shards across eight cards; the FP8 build runs on 8×A100 80GB.
SGLang
--cuda-graph-backend-prefill disabled is required for DFLY2; omitting it can cause failures in the prefill stage.
Calling the Service
Both runtimes expose an OpenAI-compatible endpoint, so no client rewrite is needed:
The extra_body field is where the reasoning mode is set on the OpenAI-compatible path. It is easy to miss and produces noticeably slower responses when omitted, since high is the default.
Three Deployment Routes
| Item | Local (GGUF) | Production API (vLLM) | Production API (SGLang) |
|---|---|---|---|
| Weights | AngelSlim/Hy4-flash-GGUF | tencent/Hy4-flash-FP8 | tencent/Hy4-flash-FP8 |
| Image | llama.cpp / Ollama | vllm/vllm-openai:hy4-flash | lmsysorg/sglang:hy4-flash |
| Hardware | ≥32GB unified memory (Q4) | 8×A100/H100 80GB | 8×A100/H100 80GB |
| DFLY2 | Partial GGUF support | method: dfly, 5 tokens | DFLY, block size 6 |
| Sparse attention | Kernel selected post-quantization | FLASHMLA_SPARSE | --trust-remote-code |
| Interface | OpenAI-compatible | OpenAI-compatible | OpenAI-compatible |
| Use case | Validation / personal dev | High-concurrency service | High-concurrency service |
Scenario one — local validation. GGUF plus llama.cpp or Ollama, no multi-GPU requirement. The Q4_K_M build runs within roughly 32GB unified memory on Apple Silicon. This path validates behavior, not throughput.
Scenario two — small-batch API. Single node, 8×A100/H100, FP8 with either official image. With DFLY2 and FLASHMLA_SPARSE enabled, measured decode throughput exceeds 150 tokens per second.
Scenario three — large-scale clusters. Follow the vLLM Recipe or the SGLang Cookbook, both of which document multi-node configuration and PD (prefill-decode) disaggregation. DFLY2's KV Share and Index Share mechanisms mean only the last layer's hidden state crosses the network during PD separation, so bandwidth pressure stays low.
For teams already running several models in one serving stack, the endpoint layer is worth settling before workload grows. Routing every model behind one gateway means per-task cost accounting stays comparable across backends, and a model swap does not require touching client code. 4sapi serves as that single access layer for mixed deployments, which is the difference between measuring cost per task and guessing at it.
Common Questions
Is the Tencent Hy License an open-source license, and can it be used commercially? The Tencent Hy License is a custom agreement, not Apache, MIT or GPL. Commercial terms are governed by the LICENSE file in the GitHub repository. Review it against your own business scenario before deployment.
Hy4 flash or Hy4 — which fits which task? Per official guidance: deep search, document processing, slide generation, simple frontend display, computer use and other high-frequency daily work go to Hy4 flash. Long-horizon software engineering and other complex tasks go to Hy4, which has a higher capability ceiling. Internally, Hy4 flash's 42.1 on WorkIndex against Hy4 preview's 39.7 indicates the flash variant received targeted tuning for routine workloads.
How should `no_think` and `high` be chosen? Latency-sensitive paths — chat assistants, real-time completion — use reasoning_effort: "no_think", which produces no reasoning tokens and responds faster. Accuracy-critical paths — complex analysis, multi-step reasoning — use "high", where the model thinks before answering. Both modes switch on the same weights; no model swap is required.
Conclusion
Hy4 flash's first-day placement on OpenRouter and its 50-trillion-token debut trace back to three cost-oriented architectural decisions rather than to raw parameter count: KV cache compressed to one-sixth, LatentMoE routing in low-rank space, and an Engram N-gram table that raises parameter count without raising compute. Together they bring cost per task to ¥0.26.
On the deployment side, the official images provide the vLLM and SGLang paths, and the two docker run commands above are the core of it. Local validation via GGUF is simpler still.
One licensing note stands: the Tencent Hy License is not a standard open-source agreement, and commercial use requires verifying its terms first. The data in this article is based on Tencent Hunyuan's official disclosure and GitHub README as of November 19, 2026; framework versions and image tags will change with future releases.




