Tencent has released the weights of Hy4 flash, the cheap, fast sibling of its Hy4 flagship. The files went up on Hugging Face on 17 November 2026, and Tencent Hy announced it on X on 18 November, two days after the hosted API launched.
It is a 220B mixture-of-experts model with about 12B active parameters, a 1M-token context window and native image input. Tencent says it reaches near Claude Haiku 5.5 satisfaction on everyday office work at a fraction of the cost.
This guide covers every download (BF16, FP8, base and GGUF), the real hardware you need, the exact vLLM, SGLang and Transformers commands, the llama.cpp situation, all published benchmarks (vendor and independent), the licence fine print and how it compares with DeepSeek V4.1 Flash and Haiku 5.5.
- What: Tencent Hy4 flash, an open-weights 220B MoE (11.3B text + 0.43B vision active per token), text + image in, text out, 1M context.
- Downloads:
tencent/Hy4-flash(BF16, about 440 GB),tencent/Hy4-flash-FP8(about 287 GB),tencent/Hy4-flash-base, plusAngelSlim/Hy4-flash-GGUF(135-168 GB). - Hardware: the official recipe uses 8 GPUs with FP8. The smallest GGUF still needs more than 128 GB of memory, so a 256 GB Mac Studio or two DGX Spark boxes is the realistic “local” setup.
- llama.cpp: not merged yet (PR #32517 open). Stock llama.cpp, Ollama and LM Studio cannot load the GGUFs today.
- API: $0.07 in / $0.281 out per 1M tokens on OpenRouter (
tencent/hy4-flash), cache reads $0.003. - Scores: Vals Index 50.8% (independent), below DeepSeek V4.1 Flash 61.3% and Haiku 5.5 58.7%, but far cheaper per test run.
- Licence: Tencent Hy License 1.0 (Apache-based, not Apache). Free for most uses; API providers with $50M+ revenue or 100M+ MAU need a separate deal.
What Is Hy4 flash?
Hy4 flash is the smaller, cheaper member of Tencent’s Hy4 family (the models formerly branded Hunyuan). Tencent describes it as self-distilled from the full Hy4 and tuned for “everyday productivity”: office documents, research, tool calls and light coding.
We covered the full Hy4 launch earlier in Tencent Hy4 Launched: Vision, Computer Use and Open Weights Coming. Flash keeps the vision input and the 1M context, but cuts the active compute to about 12B parameters per token.
Here is how the release unfolded:
| Date (2026) | What happened |
|---|---|
| 16 Nov | Hy4 flash offered as a hosted API (Tencent Cloud, later listed on OpenRouter) |
| 17 Nov | Weights appear on Hugging Face, not gated (BF16, FP8, Base) |
| 18 Nov | Tencent Hy posts the official open-weight announcement on X; AngelSlim GGUF builds and llama.cpp PR #32517 follow |
| 18-20 Nov | First independent numbers from Vals AI and OpenRouter speed data |
Note: Tencent’s X post says “224B”, but the model card and the safetensors header both give about 220B (220,251,201,958 parameters). The 220B figure matches the files you download.
Specs and Architecture
| Item | Hy4 flash |
|---|---|
| Maker | Tencent Hy team |
| Total parameters | about 220B (155.5B backbone + 63.5B Engram tables + 0.43B vision + 0.85B drafter) |
| Active per token | 11.3B text + 0.43B vision (about 12B) |
| Layers | 42 (1 dense + 41 MoE), hidden size 3,072 |
| Experts | 512 routed + 1 shared per MoE layer, 16 routed active (“LatentMoE”) |
| Attention | MLA, hybrid: 12 sparse layers (up to 2,048 selected tokens) + 30 sliding-window layers (512-token window) |
| Engram memory | Hashed bigram/trigram lookup tables, about 63B parameters |
| Vision encoder | Adapted SigLIP2-So400m (27 layers) |
| Built-in drafter | DFly2, 5 layers, 0.85B, for speculative decoding |
| Context | 1,048,576 tokens input; OpenRouter caps output at 65,536 |
| Modalities | Text and image in, text out (no video) |
| Reasoning modes | high (default, thinking on) or no_think |
| Licence | Tencent Hy License 1.0 |
Two design choices matter for anyone hosting it:
- Small KV cache. Tencent says the KV cache is 1/6 of Hy4 and 1/24 of Hy3, because only 12 of the 42 layers attend over the full context. One independent estimate puts it near 7 GB at 1M tokens with an 8-bit latent cache. Tencent has not published a GB figure.
- Engram tables. About 63B parameters are lookup tables, not compute. In the GGUF builds they sit in a separate ~63 GB file, which is why even the 2-bit build is 135 GB.
If Engram sounds familiar, DeepSeek uses a similar idea. Our DeepSeek V4.1 Flash local guide explains how those tables can be offloaded to SSD.
Every Download and Its Size
| Repo | What it is | Size | Use it for |
|---|---|---|---|
tencent/Hy4-flash |
Instruct, BF16 | ~440.5 GB (148 shards) | Full-quality serving, fine-tuning |
tencent/Hy4-flash-FP8 |
Instruct, FP8 | ~287.1 GB | The official vLLM/SGLang recipe |
tencent/Hy4-flash-base |
Pre-trained base, not chat-tuned | ~439 GB | Research and custom fine-tunes |
AngelSlim/Hy4-flash-GGUF Q4_K_M |
4-bit GGUF (recommended default) | 156.8 GiB (~168 GB) | Mac Studio / DGX Spark pairs |
| AngelSlim SPARK-4Bit | 4-bit tuned for Spark | 151.3 GiB (~162 GB) | NVIDIA DGX/RTX Spark |
| AngelSlim SPARK-4Bit-Fast | Faster Spark variant | 148.6 GiB (~160 GB) | NVIDIA DGX/RTX Spark |
| AngelSlim Q3_K | 3-bit GGUF | 144.7 GiB (~155 GB) | Tighter memory budgets |
| AngelSlim IQ2-AH | 2-bit GGUF | 125.7 GiB (~135 GB) | Smallest build, lowest quality |
The weights are also mirrored on ModelScope (Tencent-Hunyuan/Hy4-flash) and GitCode, which helps if Hugging Face is slow in your region. Tencent says the GGUF builds are “for end-to-end testing purposes”, so treat them as early.
What Hardware Can Run It?
Be realistic here: Hy4 flash is “flash” for a data centre, not for a gaming PC. Only 12B parameters are active, so it is fast once loaded, but all 220B still have to sit in memory.
| Setup | Fits? | Notes |
|---|---|---|
| 8x 80 GB GPUs (H100/H200/A100 80G) | Yes, FP8 or BF16 | Official recipe uses TP 8 with FP8 (~36 GB weights per GPU); BF16 is ~55 GB per GPU |
| Mac Studio 256 GB or 512 GB | Yes, GGUF | Q4_K_M fits; needs a patched llama.cpp build |
| 2x DGX Spark (2x 128 GB) | Yes, GGUF | AngelSlim ships Spark-specific 4-bit builds |
| Mac 128 GB / single DGX Spark | No (without SSD tricks) | Even IQ2-AH (~135 GB) is over 128 GB |
| Gaming PC, 2x RTX 5080 + 128 GB RAM | Not practical | ~160 GB total memory, but no stock llama.cpp support and heavy offload |
| Single consumer GPU (16-32 GB) | No | Use the API |
AngelSlim published its own decode speeds (not our tests, and not independently checked):
| Device (AngelSlim tests) | Build | 4K ctx | 16K ctx | 32K ctx |
|---|---|---|---|---|
| M5 Max 128 GB | Q4_K_M | 48.1 tok/s | 45.2 tok/s | 44.2 tok/s |
| M5 Max 128 GB | Q3_K | 50.3 tok/s | 49.5 tok/s | 47.4 tok/s |
| RTX Spark 128 GB | SPARK-4Bit-Fast | 32.0 tok/s | 33.1 tok/s | 33.4 tok/s |
| RTX Spark 128 GB | SPARK-4Bit | 25.8 tok/s | 28.2 tok/s | 30.3 tok/s |
Warning: Those 128 GB devices are below the file sizes, so the runs probably read the Engram file from SSD. AngelSlim has not confirmed how. Expect slower and less steady speeds on your own machine until official offload docs appear.
How to Run It With vLLM and SGLang
Both engines have day-one images. The production recipes below come straight from the model card and target 8 GPUs with the FP8 weights.
vLLM (official recipe)
- Install Docker with the NVIDIA Container Toolkit and make sure
nvidia-smishows all 8 GPUs. - Run the dedicated image. It downloads
tencent/Hy4-flash-FP8into your Hugging Face cache on first start:
docker run --gpus all \
-p 8000:8000 \
--ipc=host \
-v ~/.cache/huggingface:/root/.cache/huggingface \
vllm/vllm-openai:hy4-flash tencent/Hy4-flash-FP8 \
--tensor-parallel-size 8 \
--speculative-config '{"num_speculative_tokens":5,"method":"dfly"}' \
--attention-backend FLASHMLA_SPARSE \
--tool-call-parser hy_v4 \
--reasoning-parser hy_v4 \
--enable-auto-tool-choice \
--port 8000 \
--served-model-name hy4-flash
The dfly speculative config uses the built-in DFly2 drafter, and the hy_v4 parsers split tool calls and thinking from the answer. The simpler pip route (pip install vllm then vllm serve "tencent/Hy4-flash") loads the BF16 weights and needs more memory.
SGLang (official recipe)
docker pull lmsysorg/sglang:hy4-flash
docker run --gpus all \
-p 8000:8000 \
--ipc=host \
lmsysorg/sglang:hy4-flash \
sglang serve \
--trust-remote-code \
--model-path tencent/Hy4-flash-FP8 \
--tp 8 \
--cuda-graph-backend-prefill disabled \
--speculative-algorithm DFLY \
--speculative-dfly-block-size 6 \
--host 127.0.0.1 \
--port 8000
Tip: The SGLang recipe binds to 127.0.0.1, so only the host machine can reach it. Change it to 0.0.0.0 only behind a firewall or reverse proxy with authentication.
Transformers, Docker and the OpenAI-Style Client
Hugging Face Transformers support is not merged yet. Until it lands, install from the open pull request:
pip install "transformers[torch,vision] @ git+https://github.com/huggingface/transformers.git@refs/pull/50903/head"
Docker Model Runner users can also pull it with one line: docker model run hf.co/tencent/Hy4-flash. Remember this is still the full model, so the same memory rules apply.
Once a server is up, any OpenAI-compatible client works. This is Tencent’s example, with its recommended sampling (temperature 0.9, top_p 1.0):
from openai import OpenAI
client = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="EMPTY")
response = client.chat.completions.create(
model="hy4-flash",
messages=[{"role": "user", "content": "Hello! Can you briefly introduce yourself?"}],
temperature=0.9,
top_p=1.0,
extra_body={"chat_template_kwargs": {"reasoning_effort": "no_think"}},
)
print(response.choices[0].message.content)
Set reasoning_effort to "high" (the default) for harder tasks, or "no_think" for quick answers that skip the thinking step.
llama.cpp, Ollama and LM Studio: Current Status
| Tool | Status (21 Nov 2026) |
|---|---|
| llama.cpp (stock) | Not supported. PR #32517 “add hy_v4_flash support” is open |
| llama.cpp (patched) | Works with the PR branch or AngelSlim’s patches (hy4-flash-mac-minimal.patch for Mac, hy4-flash-spark.patch for Spark) |
| Ollama / LM Studio / Jan | Will not load the GGUFs until the llama.cpp support is merged and shipped |
| MLX / mlx-lm | No support yet |
| vLLM / SGLang | Supported (dedicated images) |
When the PR merges, the normal workflow from our llama-server guide will apply. If you have not upgraded in a while, our llama.cpp 1.0 guide covers the current install steps. We will update this post when support lands.
Benchmarks: Vendor vs Independent
Tencent-reported results (instruct model)
| Test (vendor-reported) | Hy4 flash | Comparison |
|---|---|---|
| WorkIndex (Tencent internal, 8 task sets) | 42.1 | Hy4 preview 39.7, Hy4 51.4 |
| DeepSWE | 64.6 | Hy4 74.9 |
| OSWorld 2.1 | 68.1 | Hy4 78.9 |
| Blind user test (85 WorkBuddy users, 259 tasks) | 2.79 / 4 | Claude Haiku 5.5 2.81 / 4 |
| Cost per WorkIndex task | 1x | Hy4 ~20x, Hy4 preview ~13x, DeepSeek V4.1 Flash (Max) ~8x |
Tencent-reported results (base model, few-shot, selected)
| Benchmark | Hy4 flash Base | Hy4 Base | DeepSeek-V4-flash Base |
|---|---|---|---|
| MMLU (5-shot) | 88.7 | 91.6 | 89.7 |
| MMLU-Pro (5-shot) | 71.1 | 73.9 | 68.6 |
| SuperGPQA (5-shot) | 57.8 | 62.7 | 52.5 |
| SimpleQA (5-shot) | 37.5 | 57.2 | 31.3 |
| U-MATH (5-shot) | 69.8 | 71.6 | 57.4 |
| MBPP-plus (3-shot) | 73.7 | 74.0 | 76.8 |
| MMMU-Pro (4-shot, vision) | 52.0 | 59.6 | n/a |
Independent results (Vals AI, OpenRouter)
| Model | Vals Index | Terminal-Bench 4.0 | Vibe Code Bench 1.1 | Vals Index run cost |
|---|---|---|---|---|
| Hy4 flash | 50.8% | 15.2% | 74.8% | $0.07 |
| DeepSeek V4.1 Flash (Max) | 61.3% | 31.8% | 87.7% | $0.66 |
| Claude Haiku 5.5 (max effort) | 58.7% | 36.4% | 90.4% | $3.17 |
| Hy4 | 59.4% | 39.4% | n/a | $1.72 |
The gap is clear. Tencent’s own tests put Flash close to Haiku 5.5 on office tasks, while independent agent and coding tests put it well behind. OpenRouter also measured about 70 tokens/s, half of Tencent’s “past 150 tokens/s” claim (which assumes megakernels, 4-bit weights and the DFly2 drafter).
Hy4 flash vs DeepSeek V4.1 Flash vs Haiku 5.5
| Hy4 flash | DeepSeek V4.1 Flash | Claude Haiku 5.5 | |
|---|---|---|---|
| Open weights | Yes (Tencent Hy License) | Yes | No |
| API price in / out (per 1M) | $0.07 / $0.281 | $0.051 / $0.567 | $0.10 / $0.50 |
| Image input | Yes | See vendor docs | Yes |
| Context | 1M | 1M class | See Anthropic docs |
| Vals Index | 50.8% | 61.3% | 58.7% |
| Terminal-Bench 4.0 | 15.2% | 31.8% | 36.4% |
| Best for | Cheap high-volume office, summarising, image Q&A | Cheap coding and agents | Reliable agents, managed service |
Prices and scores come from Vals AI and OpenRouter as reported by Neurohive on 18 Nov. In short: pick Hy4 flash when cost per request matters most and tasks are simple. Pick DeepSeek V4.1 Flash or Haiku 5.5 for coding and multi-step agents.
Use It Through the API Instead
For most readers the API is the sane way to try Hy4 flash. Steps:
- Create an OpenRouter account and an API key.
- Use the model slug
tencent/hy4-flashwith any OpenAI-compatible SDK and base URLhttps://openrouter.ai/api/v1. - Pass
reasoning_effortthroughchat_template_kwargsas shown above if the provider supports it. - Keep outputs under 65,536 tokens, the current OpenRouter cap.
Function calling and structured JSON output are supported. At list price, a million input tokens costs 7 cents, and Neurohive reported a 50% promotional discount with no stated end date at launch.
Tencent Cloud is the only provider behind the OpenRouter listing, so your data goes to Tencent’s servers (Singapore or Guangzhou regions were listed for the hosted API).
Licence: What You Can and Cannot Do
The weights use the Tencent Hy License, Version 1.0. It is derived from Apache 2.0 but the text says plainly that it “is not the Apache License, Version 2.0”.
- Allowed: personal, research and commercial use, fine-tuning, quantising and redistribution, with the licence and change notices included.
- Model-as-a-service limit: if you give third parties remote access to inference or fine-tuning (an API), you need a separate written licence once your company’s worldwide revenue passes US$50 million in any 12 months (from any source), or before using it in a product with over 100 million monthly active users.
- Not counted as MaaS: purely internal use, and end-user apps where the model only powers specific features.
- Derivatives: fine-tuned, quantised, distilled or merged weights count as derivatives and carry the same terms.
- Regions: unlike some earlier Chinese model licences, there is no EU, UK or South Korea exclusion in the text.
Note: This is a plain-English summary, not legal advice. Read the LICENSE file in the repo before a commercial launch, and contact Tencent (hy_partners@tencent.com) if you are near the thresholds.
Troubleshooting
“unknown model architecture: hy_v4_flash” in llama.cpp, Ollama or LM Studio
Support is not merged. Build llama.cpp from PR #32517 or apply the AngelSlim patch for your platform, or wait for an official release.
Transformers says the model type is not recognised
Install Transformers from PR 50903 with the command above. A normal pip install transformers does not include Hy4 flash yet.
Out of memory on fewer than 8 GPUs
The official recipe assumes TP 8 with FP8 (~287 GB of weights plus KV cache). Use FP8 rather than BF16, lower the max context length, or move to the API.
Thinking text mixed into the answer
Start vLLM with --reasoning-parser hy_v4, or set reasoning_effort to no_think.
Engram layer numbers do not match
The README lists Engram layers 1, 11, 21, 31 while config.json says 2, 12, 22, 32. It looks like 0-based vs 1-based counting. Trust the config file if you write custom loaders.
My Take
Hy4 flash is a pricing story more than a quality story. At $0.07 per million input tokens with image input and a 1M window, it undercuts almost everything with real vision support. For bulk jobs like tagging, summarising long PDFs or answering questions about screenshots, that matters a lot.
But the independent numbers are a reality check. A 15% Terminal-Bench score means it is not a coding agent, whatever the vendor charts say. Tencent itself recommends the full Hy4 for long software-engineering work.
For home users, the “open weights” label does not mean “runs on my PC”.
Even the 2-bit file needs more than 128 GB, and llama.cpp support is still a pull request. I would wait for the merge and for clear Engram SSD-offload docs before downloading 160 GB.
If you want a big MoE that already runs on high-RAM rigs today, DeepSeek V4.1 Flash is the safer pick.
FAQ
Is Hy4 flash really open source?
It is open weights under the Tencent Hy License 1.0. You can download, run, fine-tune and sell products with it, but large API providers need a separate deal, so it is not a pure open-source licence like Apache 2.0.
Can I run Hy4 flash on an RTX 5090 or 4090?
Not on its own. The weights are 135-440 GB depending on the build, far above a single card’s VRAM. Use the API, or a 256 GB+ Mac Studio or multi-GPU server.
Does Hy4 flash work in Ollama?
Not yet. Ollama depends on llama.cpp, and Hy4 flash support (PR #32517) has not merged as of 21 November 2026.
How much does the Hy4 flash API cost?
$0.07 per million input tokens and $0.281 per million output tokens on OpenRouter, with cache reads at $0.003 per million.
Is Hy4 flash better than DeepSeek V4.1 Flash?
On independent tests, no: Vals Index 50.8% vs 61.3%, and Terminal-Bench 15.2% vs 31.8%. Hy4 flash is cheaper on output tokens and has native image input.
What is the difference between Hy4 flash and Hy4 flash-base?
Hy4 flash is the chat/instruct model you talk to. Hy4 flash-base is the raw pre-trained model for researchers and custom fine-tuning; it does not follow instructions well.
Will the full Hy4 get open weights too?
Tencent’s X post says to “stay tuned for Hy4 coming up next”, and launch coverage says flagship weights will follow. No date was confirmed at the time of writing.
Sources
- Hugging Face model card: huggingface.co/tencent/Hy4-flash (specs, benchmarks, commands) and its LICENSE file
- Tencent Hy on X, 18 Nov 2026 (open-weight announcement)
- GitHub: github.com/Tencent-Hunyuan/Hy4-flash
- Neurohive, “Hy4 flash: Tencent’s Cheapest Hy Model Gets Open Weights”, 18 Nov 2026 (Vals AI and OpenRouter data)
- ModelFit, “Tencent Hy4 Flash Weights Are Out: What Hardware Runs It?”, 17-18 Nov 2026 (GGUF sizes, AngelSlim speeds, llama.cpp status)
- AlphaSignal, “Tencent Opens Hy4 Flash Weights With 1M-Token Context”, 18 Nov 2026
- DEV Community, “Tencent Hy4 flash: 220B parameters, 12B active, and what the KV cache and license mean”, 20 Nov 2026
- vLLM recipe: recipes.vllm.ai/tencent/Hy4-flash; SGLang cookbook: docs.sglang.io



