The cost-oriented member of Tencent Hy Team's fourth generation, released on November 17 two weeks after the API-only Hy4 general release. Unlike Hy4, it shipped as open weights: an instruct model, an FP8 checkpoint and a pre-trained Hy4 flash-base on HuggingFace, ModelScope and GitCode. Tencent's model card gives 220B total parameters (including 64B Engram parameters) with 12B active per token. Broken down, that is 155.5B backbone parameters, 63.5B Engram n-gram embedding tables, a 0.43B vision encoder (adapted from SigLIP2-So400m) and a 0.85B built-in drafter; activation is "Text 11.3B + Vision 0.43B". It takes text and image input natively, outputs text, and supports a 1M-token context (max_position_embeddings 1,048,576).

It is a different architecture from the 770B Hy4 preview, not a smaller copy of it. The 42-layer backbone (hidden size 3,072, one dense layer then 41 MoE layers, iHC residuals with 4 streams) adds three changes that Tencent lists against Hy4. Hybrid attention interleaves sparse attention ("IndexPool", top-2,048 tokens by group-level selection) with 512-token sliding-window attention at 1:3 (12 sparse / 30 SWA layers, all MLA). LatentMoE runs the routed experts in a 1,536-dimensional latent space, with 16 of 512 routed experts plus one shared active per token (cf. NVIDIA's Nemotron 3 Super). An Engram n-gram embedding module (cf. DeepSeek's Engram) injects hashed bigram and trigram lookups at four layers. Tencent says these cut the KV cache to 1/24 of Hy3's and 1/6 of Hy4's, and that scaling-ladder experiments backed the choices. Post-training is self-distillation from Hy4 followed by RL that halved average end-to-end output length. The DFly2 drafter (5 layers) reuses the target model's KV cache and sparse indices, so each draft token attends to at most about 2K keys at any context length. With megakernels and 4-bit quantization, Tencent reports decoding above 150 tokens/s. Training tokens, hardware and duration are not disclosed. The weights are under the Tencent Hy License, an Apache-2.0 derivative that requires a separate licence for model-as-a-service operators with over US$50M revenue or products with over 100M MAU. This is a change from the preview's plain Apache-2.0.

All results are Tencent's, at the highest reasoning setting. Tencent's framing is cost per task: on its internal WorkBuddy-based WorkIndex the model scores 42.1 against 39.7 for Hy4 preview and 51.4 for Hy4, at 1/13 of the preview's and 1/20 of Hy4's cost per task. Public agent rows (Hy4 flash / Hy4 / DeepSeek-V4.1-Flash-1017): DeepSWE v1.1 64.6 / 74.9 / 73.8, Terminal-Bench 4.0 19.2 / 44.4 / 40.0, ProgramBench 16.0 / 30.0 / 29.0, APEX-Agents v1.1 63.1 / 75.2 / 72.9, HLE-Full with tools 55.2 / 59.1 / 57.1, OSWorld 2.1 68.1 / 78.9 / 68.4, MMMU-Pro 75.8 / 78.6 / 76.7. It trails Qwen3.8-Flash-Next and GLM-5.3-Flash on Terminal-Bench 4.0 (29.5, 35.4) and leads both on APEX-Agents and OSWorld. Tencent recommends Hy4 for long-horizon software engineering. The base model reports MMLU-Pro 71.1 and SimpleQA 37.5 (5-shot), against 73.9 and 57.2 for Hy4 Base, with a warning that heavy synthetic instruction data in mid-training distorts its few-shot formatting. In a blind test with 85 WorkBuddy users (259 tasks), satisfaction was 2.79/4 against Claude Haiku 5.5's 2.81. List price is about $0.070/$0.281 per M input/output tokens ($0.003 cached), with TokenHub at 50% off through November 30. Not yet scored on the Artificial Analysis Intelligence Index as of 2026-11-19: the AA leaderboard lists Hy4 but no Hy4 flash.

Model Details

Architecture MOE
Parameters 220B
Active params 12B
Experts 512 (top-16)
Context window 1,048,576
License Tencent Hy License

Benchmark Scores

Benchmark Score Mode
DeepSWE v1.1 64.6 mini-SWE-agent
Terminal-Bench 4.0 19.2 Claude Code harness, 500 turns
ProgramBench 16.0 almost-solved, mini-SWE-agent
APEX-Agents v1.1 63.1 pass@1
AutomationBench v1.0.6 52.2 —
Humanity's Last Exam (full) 55.2 with tools
OSWorld 2.1 68.1 full, partial score
MMMU-Pro 75.8 —
MMLU-Pro 71.1 base model, 5-shot

Variants

Name Parameters Notes
Hy4-flash 220B BF16 instruct checkpoint (incl. 64B Engram tables)
Hy4-flash-FP8 220B FP8-quantized instruct checkpoint
Hy4-flash-base 220B Pre-trained base model
moeopen-weightmultimodalefficiencylong-contextagentic

Related