Xylolabs

Hy4 flash GGUF

Base model Formats License Code

GGUF quants of tencent/Hy4-flash (220B parameters, 12B active) for llama.cpp, on CUDA, Metal or CPU, without the DFly2 speculative-decoding drafter.

Files

Each quant has two files:

  • Hy4-flash-<QUANT>-00001-of-00002.gguf is the trunk.
  • Hy4-flash-<QUANT>-00002-of-00002.gguf holds the Engram tables (63.4 GB, FP8 E4M3). It is the same file for every quant (sha256 c3589e8789026f7de689a5bd2258daf018deddcbd903a443713051a6d84351a9), and llama.cpp picks it up from the same folder.

Also in this repository: mmproj-Hy4-flash-BF16.gguf (1.04 GB) and mmproj-Hy4-flash-Q8_0.gguf (0.78 GB), the vision tower in BF16 and in Q8_0, and Hy4-flash.imatrix.gguf, the importance matrix.

Quant Trunk size
Q8_0 165.50 GB
Q6_K 129.89 GB
Q5_K_M 113.37 GB
Q4_K_M 97.81 GB
Q4_K_S 92.19 GB
IQ4_XS 86.94 GB
Q3_K_M 78.58 GB
IQ3_M 72.49 GB
IQ3_XXS 64.89 GB
Q2_K 61.79 GB
IQ2_M 56.26 GB

Smaller Engram files

For Q4_K_M and Q3_K_M there are also pairs whose second file holds the Engram tables in Q5_0 or IQ4_NL instead of FP8 E4M3: Hy4-flash-<QUANT>-EQ5_0-0000N-of-00002.gguf and Hy4-flash-<QUANT>-EIQ4_NL-0000N-of-00002.gguf. Their first file is the quant's trunk, byte for byte. The Q5_0 file is 43.56 GB and the IQ4_NL file is 35.64 GB.

How to run

Hy4-flash support in llama.cpp is pull request #32517 (commit 0a054d0b0) until it is merged. The same build runs on CUDA and on Metal. Tencent's patched build also loads these files.

git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
git fetch origin pull/32517/head:hy4-flash && git checkout hy4-flash
cmake -B build -DCMAKE_BUILD_TYPE=Release        # add -DGGML_CUDA=ON for NVIDIA GPUs
cmake --build build -j --target llama-cli llama-server

./build/bin/llama-cli -m Hy4-flash-Q4_K_M-00001-of-00002.gguf \
  --override-tensor '^per_layer_token_embd\.weight$=CPU' --load-mode mmap --lazy-mode on \
  -ngl 99 --jinja -c 8192

--override-tensor keeps the Engram tables in host memory, and --load-mode mmap --lazy-mode on reads them from disk only when a prompt needs them. llama-server takes the same flags.

Accuracy

WikiText-2 raw test (wiki.test.raw), chunks 0–23 of 512 tokens. Logits at positions 256–510 predict the next token (255 predictions per chunk, 6,120 total). The reference is the Hy4-flash BF16 GGUF trunk with FP8 E4M3 Engram tables. Its llama.cpp run used a logical batch of 2,048 and 4 parallel sequences. The launcher left ubatch at its default. The logs omit its effective value and the perplexity executable’s exact revision.

BF16 perplexity was 30.3937 from the original logits and 29.64825 from the saved reference, which floors logits 16 below each position’s maximum. The table uses each quant’s own logits for perplexity. Mean KL uses packed reference log probabilities above −16 without renormalization. Same top-1 is agreement with the packed reference across the full vocabulary. The GGUF quant runs also used a logical batch of 2,048 and 4 parallel sequences with the default ubatch.

Trunk Engram Mean KL vs BF16 Same top-1 Perplexity Answer check
Q8_0 E4M3 0.1741 87.859% 29.60 5/5
Q6_K E4M3 0.1930 87.190% 28.21 5/5
Q5_K_M E4M3 0.1868 86.895% 29.19 5/5
Q4_K_M E4M3 0.2441 85.049% 30.96 5/5
Q4_K_S E4M3 0.2660 84.820% 31.83 5/5
IQ4_XS E4M3 0.2745 83.840% 29.15 5/5
Q3_K_M E4M3 0.3726 80.866% 26.99 5/5
IQ3_M E4M3 0.4513 78.922% 34.95 5/5
IQ3_XXS E4M3 0.5094 77.222% 31.22 5/5
Q2_K E4M3 0.6766 73.775% 25.64 5/5
IQ2_M E4M3 0.7155 72.908% 28.65 5/5
Q3_K_M IQ4_NL 0.3728 81.013% 27.08 5/5
Q3_K_M Q5_0 0.3635 80.801% 27.47 5/5
Q4_K_M IQ4_NL 0.2584 84.592% 31.16 5/5
Q4_K_M Q5_0 0.2542 84.379% 31.10 5/5

The answer transcripts record llama.cpp b12473-0a054d0b0. The check used greedy, single-turn chat (temperature 0), an 8,192-token context and a 4,096-token generation cap per prompt. Each final response was checked for the expected text on five prompts covering a capital city, multiplication, Fibonacci code, Korean translation and elapsed time.

Answer-check prompts
  1. What is the capital of France? Answer in one word.
  2. Compute 17 * 23 and give only the number.
  3. Write a Python function that returns the n-th Fibonacci number.
  4. Translate to Korean: The weather is nice today.
  5. A train leaves at 3:15 pm and arrives at 5:40 pm. How long is the trip?

On the same text and BF16 reference, Tencent’s AngelSlim Q4_K_M measured 0.2814 mean KL, 84.44% same top-1 and 31.92 perplexity.

Mean KL and same top-1 against BF16 by complete weight size and effective bits per weight

Size counts the serialized bytes of each complete two-shard GGUF pair, including its Engram shard. Effective bits per weight is eight times that byte count divided by the 218,938,402,893 floating-point weights in the source text model, including Engram. Standalone vision, draft-model weights and integer lookup buffers are excluded from the denominator. Stars mark Q4_K_M trunks. Colors distinguish the Engram formats and the AngelSlim release.

Plotted configurations and file sizes
Release Configuration Weight-file bytes Decimal GB Effective bits/weight
Xylolabs Q8_0 / E4M3 228,859,223,360 228.859 8.3625
Xylolabs Q6_K / E4M3 193,250,525,728 193.251 7.0614
Xylolabs Q5_K_M / E4M3 176,726,016,544 176.726 6.4576
Xylolabs Q4_K_M / E4M3 161,173,537,312 161.174 5.8893
Xylolabs Q4_K_S / E4M3 155,548,975,648 155.549 5.6838
Xylolabs IQ4_XS / E4M3 150,301,901,344 150.302 5.4920
Xylolabs Q3_K_M / E4M3 141,940,556,320 141.941 5.1865
Xylolabs IQ3_M / E4M3 135,853,572,640 135.854 4.9641
Xylolabs IQ3_XXS / E4M3 128,247,202,336 128.247 4.6861
Xylolabs Q2_K / E4M3 125,151,805,984 125.152 4.5730
Xylolabs IQ2_M / E4M3 119,621,616,160 119.622 4.3710
Xylolabs Q3_K_M / IQ4_NL 114,220,286,848 114.220 4.1736
Xylolabs Q3_K_M / Q5_0 122,140,363,840 122.140 4.4630
Xylolabs Q4_K_M / IQ4_NL 133,453,267,840 133.453 4.8764
Xylolabs Q4_K_M / Q5_0 141,373,344,832 141.373 5.1658
AngelSlim Q4_K_M 168,347,133,632 168.347 6.1514

License and notices

These files are under the Tencent Hy License, Version 1.0, copied in LICENSE. Its terms apply to them, including the extra terms for large Model-as-a-Service operators. We quantized the weights and converted them to GGUF; the conversion patch is hy4-flash-release-convert.patch. NOTICE names the source model and its copyright holder. Xylolabs is not affiliated with or endorsed by Tencent.

Tencent's original model description

The text below is reproduced from the README of tencent/Hy4-flash (the model introduction, highlights, benchmarks, quickstart and deployment sections), copyright Tencent, under the Tencent Hy License (see above). It describes Tencent's original model, not these files, and its usage instructions (Transformers, vLLM, SGLang) are for Tencent's checkpoints; for the GGUF files use "How to run" above. Changes by Xylolabs: heading levels are one level deeper, and image and link paths point to Tencent's repository instead of copies; the wording is unchanged. Tencent's own README is the authoritative version.

Model Introduction

Hy4 flash is our most cost-effective Hy model to date, featuring 220B total parameters (including 64B Engram parameters) with just 12B active parameters per token. It natively accepts text and image inputs, generates text, and supports a context window of up to 1M tokens.

Its 42-layer language backbone consists of one dense feed-forward layer and 41 MoE layers, with identity Hyper-Connections (iHC) supporting inter-layer information flow. A vision encoder (adapted from SigLIP2-So400m-Patch16-Res512) is connected to the backbone through an MLP projector. In addition to the backbone, a built-in 5-layer drafter, DFly2 (0.85B parameters), reuses the target model's KV cache and sparse indices to support long-context speculative decoding at a low, constant cost.

Compared with Hy4, Hy4 Flash introduces three notable architectural changes:

  1. Hybrid attention. Sparse attention (IndexPool) and sliding-window attention (SWA) are interleaved at a 1:3 ratio.
  2. LatentMoE. Routed experts operate in a lower-dimensional latent space. Each token activates 16 of 512 routed experts, alongside one shared expert.
  3. N-gram embedding (Engram). The backbone is augmented with learned bigram and trigram representations, retrieved via multi-head hash lookups.
Property Value
Overview
Architecture Mixture-of-Experts (MoE)
Total Parameters (excluding Engram, ViT, and DFly2) 155.5B
Engram Parameters 63.5B
ViT Parameters 0.43B
DFly2 Parameters 0.85B
Activated Parameters per Token Text 11.3B + Vision 0.43B
Context Length 1M (1,048,576 tokens)
Language Backbone
Backbone Layers 42 — 1 dense layer and 41 MoE layers
Hidden Size 3,072
Vocabulary Size 120,832
Residual Connections iHC (identity Hyper-Connections) with 4 residual streams
Attention — Hybrid Layout
Attention Architecture Sparse attention + sliding-window attention (SWA)
Backbone Layer Pattern 2 × Sparse → 10 × (3 × SWA → Sparse)
Attention Formulation Multi-head Latent Attention (MLA)
Multimodal Position Encoding Interleaved MRoPE
Attention — Layers Sparse / SWA
Backbone Layers 12 / 30
Attention Heads 32 / 64
MLA q_lora_rank 1,024 / 1,024
MLA kv_lora_rank 512 / 512
QK Head Dimension (non-RoPE + RoPE) 192 + 64 / 192 + 64
Value Head Dimension 256 / 256
RoPE Base 10,000,000 / 10,000
Sparse Context Selection Up to 2,048 tokens via group-level selection
SWA Context Window 512-token causal window (511 preceding + current)
Feed-Forward Networks
Dense FFN Intermediate Size 13,824
Routed Experts per MoE Layer 512
Shared Experts per MoE Layer 1
Activated Routed Experts per Token 16
MoE Latent Dimension 1,536
Expert Intermediate Size 1,536
N-gram Embedding
N-gram Orders Bigrams and trigrams
Backbone Layer Indices (starting from 0) 1, 11, 21, 31
Hash Heads per N-gram Order 8
Hash Vocabulary Size per Head 10.3M
Embedding Dimension per N-gram Order 768 (8 heads × 96 dimensions)
Embedding Table Parameters 63.4B
Vision Encoder
Vision Encoder Layers 27
Vision Encoder Hidden Size 1,152
Vision Encoder Intermediate Size 4,304
Vision Encoder Attention Heads 16
Vision Patch Size 16 × 16
Vision Position Embedding Size 128 × 128
Visual Feature Merging 2 x 2

Pretrained model capability

Bits-per-byte (BPB) on internal code & agent validation, lower is better:

Bits-per-byte on internal code and agent validation versus active parameters; lower is better

Few-shot pretrain benchmark evaluation:

Category Benchmark #Shots Hy4 flash Base Hy4 Base Hy3 preview Base DeepSeek-V4-flash-Base DeepSeek-V4-Pro-Base
Knowledge MMLU† 5 88.7 91.6 87.5 89.7 91.1
MMLU-Pro† 5 71.1 73.9 65.8 68.6 73.4
MMLU-Redux† 5 87.6 90.6 87.0 88.8 90.2
SuperGPQA† 5 57.8 62.7 52.3 52.5 54.8
C-Eval† 5 92.0 95.0 90.1 91.9 93.4
CMMLU† 5 91.6 93.1 89.6 89.8 90.7
SimpleQA 5 37.5 57.2 26.4 31.3 52.1
Chinese-SimpleQA† 5 74.6 83.9 69.6 71.9 76.6
AGIEval (English)† 3 75.4 79.3 76.0 76.3 78.2
AGIEval (Chinese)† 3 90.6 91.2 87.8 88.5 89.8
Reasoning HellaSwag (3,000-item subset) 10 85.2 87.1 85.7 85.1 86.8
PIQA 4 84.4 85.3 84.6 83.8 84.9
ARC-Challenge 0 95.0 97.3 96.0 97.3 97.0
Social IQa 0 80.1 82.7 80.5 72.5 73.1
KOR-Bench 3 54.7 57.6 54.6 53.4 55.4
C³ 0 77.3 82.4 75.9 74.1 78.3
Math U-MATH 5 69.8 71.6 60.9 57.4 64.8
OCW Courses 5 64.0 66.7 62.6 63.3 68.2
Code MBPP-plus 3 73.7 74.0 77.4 76.8 79.1
Multilingual Belebele (5-language subset)† 5 93.8 95.0 94.2 94.4 95.2
INCLUDE† 5 82.8 88.1 84.3 84.0 84.3
Multimodal MMMU-Pro (Standard, 10 opts) 4 52.0 59.6 — — —
SimpleVQA 4 48.5 54.5 — — —
CVBench 4 83.4 86.7 — — —
RealWorldQA 4 71.1 72.3 — — —
TextVQA 4 73.4 78.2 — — —
CountBenchQA 4 86.6 90.6 — — —

† Macro-averaged accuracy for the marked text benchmarks. Other text rows use the accuracy metric reported by our evaluation suite. Bold marks the highest available numerical score in each row; it does not indicate statistical significance.

Multimodal results are reported only for Hy4 flash Base and Hy4 Base.

ARC-Challenge and HellaSwag use 299-item and 3,000-item subsets, respectively. INCLUDE uses a 5,000-example Lite subset covering 44 languages. Belebele covers German, Korean, Vietnamese, Spanish, and Russian (4,475 examples in total).

Disclaimer: Because our base model was exposed to a large amount of synthetic instruction data during mid-training, it occasionally fails to follow the expected answer format in few-shot evaluations, which may distort the scores.

Practicality and Cost Efficiency

Optimization Approach

Across training, evaluation, and inference, we aggressively optimized for cost efficiency while preserving intelligence:

  • Pre-training: Hy4 flash integrates hybrid attention, LatentMoE, and Engram into its architecture—reducing KV cache size to just 1/24 of Hy3 and 1/6 of Hy4, with extensive scaling ladder experiments ensuring performance meets expectations.
  • Post-training: Self-distillation from Hy4 establishes strong general intelligence, while reinforcement learning cuts average end-to-end output tokens by 50%.
  • Evaluation: We built WorkIndex, an internal benchmark grounded in typical WorkBuddy tasks, and paired it with human evaluation—focusing on deliverable quality, single-turn output length, and task turnaround time rather than public benchmark scores detached from everyday work.
  • Inference: Combining megakernels, 4-bit quantization, and DFly2 pushes decoding speed past 150 tokens/s, enabling low-cost serving at scale.

Task Performance and Cost

Hy4 flash delivers overall task performance slightly above Hy4 preview at a much lower cost per task. On WorkIndex, it scores 42.1, compared with 39.7 for Hy4 preview, while its average cost per task is just 1/13 of Hy4 preview, 1/20 of Hy4, and 1/8 of DeepSeek-V4.1-Flash-1017 (Max tier, peak-hour pricing).

Performance and cost per task on WorkIndex for Hy4 flash and comparison models Performance and cost per task on Open Benchmarks for Hy4 flash and comparison models

Figure 1: Performance and cost across internal and external tasks. WorkIndex (internal) comprises 8 evaluation sets covering slide creation, slide editing, human-AI slide co-writing, Markdown human-AI co-writing, document editing, spreadsheet repair, audio/video processing, and visual understanding.
Open Benchmarks includes 7 public agent benchmarks: DeepSWE, ProgramBench, Terminal-Bench 4.0, APEX-Agents, AutomationBench, HLE, and OSWorld 2.1.

Average cost per task on WorkIndex for Hy4 flash, Hy4 preview, and Hy4

Figure 2: Average cost per task on WorkIndex. Task scores are 42.1 for Hy4 flash, 39.7 for Hy4 preview, and 51.4 for Hy4.

We recommend Hy4 flash for most everyday productivity tasks, including deep search, document processing, slide creation, simple frontend demos, and computer use. For complex tasks such as long-horizon software engineering, we recommend Hy4 for its higher intelligence ceiling.

User Blind Evaluation

In the blind evaluation, Hy4 flash achieved user satisfaction close to Claude Haiku 5.5, while recording the lowest cost per task and shortest median end-to-end completion time among the evaluated models. Its user satisfaction score was 2.79/4 (Claude Haiku 5.5: 2.81/4), with a median end-to-end completion time of 6.7 minutes and a cost of RMB 0.26 per task.

Blind evaluation comparing user satisfaction, median end-to-end completion time, and cost per task

Figure 3: Blind evaluation with 85 internal WorkBuddy users across 259 tasks spanning 27 categories.

Appendix: Benchmark results

Full benchmark scores, task costs, and evaluation notes for Hy4 flash and comparison models

Quickstart

Use Transformers to load the model for local inference, or deploy a server and access it through the OpenAI-compatible API.

Reasoning mode (instruct models): reasoning_effort supports "high" (default, thinking enabled) and "no_think" (direct responses). Both examples below use "no_think".

Recommended sampling parameters: temperature=0.9, top_p=1.0.

Transformers

Support for Hy4 flash will be merged into the transformers main branch soon. Until then, install from the commit below.

pip install "transformers[torch,vision] @ git+https://github.com/huggingface/transformers.git@refs/pull/50903/head"

The following example generates a brief image description:

from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "tencent/Hy4-flash"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(model_id, device_map="auto")

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg",
            },
            {"type": "text", "text": "Describe this image in one sentence."},
        ],
    }
]

inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    reasoning_effort="no_think",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True))

OpenAI-compatible API

For API access, start a server using the vLLM deployment instructions, then run the following client example:

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)

Deployment

DFly2 for Speculative Decoding

DFly2 is Hy4 flash's speculative-decoding drafter, trained with AngelSpec and built directly on the target model's Top-k Sparse KV: each draft layer shares the KV cache (KV Share) and the compressed indexer's selection of up to 2,048 positions (Index Share) with its corresponding sparse attention layer. This yields three advantages:

  • Constant cost: each draft token attends to at most ~2K keys, regardless of context length.
  • Zero extra KV memory: no separate KV cache is maintained for the drafter.
  • PD-disaggregation ready: not involved in prefill; the prefill-to-decode transfer adds only one last-layer hidden state.

DFly2 architecture: each draft layer reuses the KV cache and the indexer-selected positions of a sparse attention layer of Hy4 flash

Quantization

We provide AngelSlim, a more accessible, comprehensive, and efficient toolkit for large model compression. AngelSlim supports a comprehensive suite of compression tools for large-scale multimodal models, including common quantization algorithms, low-bit quantization, and speculative sampling.

For this model, we have provided a GGUF compressed version for end-to-end testing purposes.

Inference Frameworks

For production serving, we recommend using vLLM or SGLang. Please refer to the recipes:

vLLM

Use official prebuilt image vllm/vllm-openai:hy4-flash:

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
SGLang

Use the official prebuilt image lmsysorg/sglang:hy4-flash (multi-arch, x86 and Arm):

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
Downloads last month
8,276
GGUF
Model size
219B params
Architecture
hy_v4_flash
Hardware compatibility
Log In to add your hardware

2-bit

3-bit

4-bit

5-bit

6-bit

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Xylolabs/Hy4-flash-GGUF

Quantized
(40)
this model

Collection including Xylolabs/Hy4-flash-GGUF