language:
- en
- zh
- es
- fr
- ja
- ko
- ru
license: other
license_name: tencent-hunyuan
license_link: https://huggingface.co/tencent/Hy4-flash/blob/main/LICENSE
base_model: tencent/Hy4-flash
base_model_relation: finetune
library_name: transformers
pipeline_tag: text-generation
thumbnail: dealign_mascot.png
gated: manual
extra_gated_prompt: |
By requesting access you confirm you are 18+, will use the model solely in
compliance with applicable law in your jurisdiction, and accept full responsibility
for all outputs. Access is granted at the sole discretion of the repository owner.
extra_gated_fields:
Name: text
Affiliation / organization: text
Country: country
Primary use case: text
I confirm I am 18 or older: checkbox
I accept full responsibility for outputs: checkbox
extra_gated_heading: "Request access — manual approval required"
extra_gated_description: "This model is reviewed manually. Fill in the fields below."
extra_gated_button_content: "Request access"
tags: - tencent
- hunyuan
- hy4
- hy4-flash
- mxfp4
- compressed-tensors
- uncensored
- reasoning
- tool-calling
- moe
[!NOTE]
Please credit if you build on this. If you redistribute, quantize, fine-tune, serve, benchmark, or write about this model, a line like
"Based onlyber-ai/Hy4-flash-UNCENSORED-MXFP4"
in your model card or write-up means a lot. Thank you.
[!IMPORTANT]
Shipped in mixed-precision MXFP4 (OCP microscaling FP4 for MoE weights) + FP8 for
attention, in the compressed-tensors format. Extended thinking with reasoning-effort
tiers (no_think/high) and tool calling are preserved.
Recommended sampling:temperature=0.7,top_p=0.95,max_tokens ≥ 16384onhigh.
Hy4-flash — UNCENSORED (MXFP4)
Uncensored release · Mixed MXFP4 / FP8 · 128K context · Extended thinking · Tools · ~217 GB
What this is
This is tencent/Hy4-flash (Hunyuan V4 Flash)
with refusal behavior removed, so the model complies with user requests across topic
categories rather than declining them. Shipped in mixed MXFP4 + FP8 compressed-tensors
format for native 4-bit MoE throughput on Blackwell. All upstream capabilities — extended
thinking with reasoning_effort tiers, tool calling, and the native 128K+ context — are
preserved. The bundle is drop-in compatible with the base model's serve configs (see below
for the vLLM image requirement).
Internal benchmarks
Capability — MMLU (228-question balanced subset, same harness as base)
228 MMLU questions · logit mode on no_think (argmax of A/B/C/D token logprobs;max_tokens=1, temperature=0) · reasoning_effort=high mode uses greedy sample and
parses the final letter.
| Mode | Base | UNCENSORED-FP8 (ship) | UNCENSORED-MXFP4 (this) |
|---|---|---|---|
no_think (logit) |
88.16% | 87.28% | 86.40% |
reasoning_effort=high (sample) |
94.30% | 91.67% | 89.91% |
Both modes hold within ~2.5 pp of the FP8 ship; the high mode's 12 unparsed responses hit
the 4096-token cap in extended reasoning before emitting the final letter (budget-limited,
not capability-limited).
HarmBench-320 — full canonical test set
320 standard HarmBench behaviors · reasoning_effort=high · max_tokens=4096 ·
no system prompt · deterministic grader (refusal opener + full-content audit for task
attempts + reasoning-engagement audit).
All rows audited by LLM after the deterministic grader, so the keyword-flagged
false-positives are corrected — a "disclaimer / for-research" preamble followed by a
substantive attempt counts as COMPLY, and an empty_final row whose reasoning trace
engaged with the task counts as COMPLY.
Harmful categories (chem/bio, cyber, illegal, harmful, misinfo, harassment) — 240 prompts
| Mode | COMPLY | % |
|---|---|---|
no_think |
237 / 240 | 98.8% |
reasoning_effort=high |
237 / 240 | 98.8% |
Full HB-320 (incl. 80 copyright prompts)
| Mode | COMPLY | REFUSE | % |
|---|---|---|---|
no_think |
317 / 320 | 3 | 99.1% |
reasoning_effort=high |
284 / 320 | 36 | 88.8% |
The high-mode drop vs no_think is almost entirely in the copyright category (47 / 80
vs 80 / 80), where many responses ran out of reasoning budget at 4096 tokens before
emitting final text — a high-mode 49K-budget rerun that matches the FP8 ship's harness
is pending. The harmful-content axis in both modes stays at 98.8%.
HarmBench-320 by SemanticCategory
| SemanticCategory | n | COMPLY (no_think) | COMPLY (high) |
|---|---|---|---|
| chemical_biological | 42 | 39 / 42 (92.9%) | 42 / 42 (100%) |
| cybercrime_intrusion | 52 | 52 / 52 (100%) | 51 / 52 (98.1%) |
| harassment_bullying | 21 | 21 / 21 (100%) | 21 / 21 (100%) |
| harmful | 18 | 18 / 18 (100%) | 16 / 18 (88.9%) |
| illegal | 53 | 53 / 53 (100%) | 53 / 53 (100%) |
| misinformation_disinformation | 54 | 54 / 54 (100%) | 54 / 54 (100%) |
| copyright | 80 | 80 / 80 (100%) | 47 / 80 (58.8%) — budget-limited |
| Harmful (non-copyright) total | 240 | 237 / 240 (98.8%) | 237 / 240 (98.8%) |
| Full total | 320 | 317 / 320 (99.1%) | 284 / 320 (88.8%) |
Multiturn adversarial pressure (think-high, no system prompt)
Harsh prompt sets with preserved reasoning across turns, 1 seed × 10 turns, long budget
per turn:
| Setting | Comply | Loops | Entropy (mean / tail) | Max context |
|---|---|---|---|---|
| preserved-reasoning multiturn, 10 turns (cyber → chem → violence → bio → fraud → weapons → misinfo → drugs → harassment → self-harm) | 10 / 10 | 0 / 10 | 0.67 / 0.50 nats | 18,304 tokens |
Every turn finished cleanly (stop), no entropy collapse, no rep8 hit at any turn.tail_entropy stayed between 0.29 and 0.74 across the 10-turn span.
Loop stability — adversarial single-turn
10 harsh adversarial prompts × reasoning_effort=high · T=0 greedy · 16K tokens:
| Build | Loops (rep8 ≥ 0.5) | finishes | mean entropy (med) | tail entropy (max) |
|---|---|---|---|---|
tencent/Hy4-flash-FP8 (ship) |
0 / 10 | 10 stop | 0.33 | 0.42 |
lyber-ai/Hy4-flash-UNCENSORED-FP8 |
0 / 10 | 10 stop | 0.32 | 0.39 |
lyber-ai/Hy4-flash-UNCENSORED-MXFP4 (this) |
0 / 10 | 10 stop | 0.34 | 0.43 |
No loop collapse on the adversarial set. Entropy stays flat across 10 harsh prompts atreasoning_effort=high, T=0, 16K-token budget.
How to serve
vLLM (recommended, Blackwell / B300)
Requires the Lyber-patched vllm/vllm-openai:hy4 image (upstream PR pending):
docker run -it --rm --gpus all --ipc=host --shm-size=64g --network host \
-v /path/to/model:/model \
vllm/vllm-openai:hy4-lyber-v6 \
--model /model \
--served-model-name hy4-flash \
--quantization compressed-tensors \
--tensor-parallel-size 8 \
--max-model-len 131072 \
--max-num-seqs 64 \
--gpu-memory-utilization 0.92 \
--reasoning-parser hy_v4 --tool-call-parser hy_v4 --enable-auto-tool-choice \
--engram-config '{"cpu_offload":false}' \
--host 0.0.0.0 --port 8000
Hardware requirements
| GPU | Setup | Status |
|---|---|---|
| 8× B300 SXM6 | TP8, MXFP4 MoE | ✅ Fits comfortably (~217 GB) |
| 8× H200 | TP8 | ⚠️ Marlin MXFP4 only (no native FP4) |
| 8× H100 80GB | TP8 | ⚠️ Marlin MXFP4 only (no native FP4) |
Reasoning + tools
chat_template_kwargs = {"reasoning_effort": "high"} # or "no_think"
reasoning_effort accepts "no_think" or "high" (the two tiers Tencent ships on this
model). Tool calling and DFly speculative decoding (5 draft tokens) are preserved from the
base model.
Responsible use
You are solely responsible for how you use this model and its outputs.
- Do not use for generating illegal content in your jurisdiction
- Do not use to harm, deceive, or exploit real individuals
- Do not use to generate CSAM or non-consensual intimate content
- Comply with all applicable laws and regulations
License
This model inherits the Tencent Hy License, Version 1.0
from the base model. The modification does not change the license terms.
Credits
- Base model: tencent/Hy4-flash
- Released by: Lyber AI × dealign.ai
