Strata on GitHub

Hy4-Flash for Strata

Run Tencent's 220-billion-parameter Hy4-Flash on your own computer. It starts at 64 GiB of RAM with an 8 GiB graphics card, with packs for 64, 96 and 128 GiB. Free and open source. No account needed.

5-step 1,300-voxel pagoda garden evolution

Pick your pack · Get started · Quality · Speed · Support Strata

220B parameters, 12B active per token. Runs from 64 GiB of RAM with an 8 GiB graphics card (GPU). Up to 1M tokens of context, with text and images in.

Hy4 flash is Tencent's most cost-effective Hy model to date. It has 220B parameters in total (64B of them in its Engram lookup) and uses just 12B per token. It reads text and images, writes text, and supports up to 1M tokens of context (1,048,576).

Pick your pack

The number in the name is your RAM. Find your computer's memory and take the pack with the same number.

Three packs. 64 GiB of RAM: SQ-64, with an 8, 12 or 16 GiB graphics card, 130 GB on disk, the smallest download. 96 GiB: SQ-96, with an 8 or 12 GiB graphics card, 157 GB. 128 GiB: SQ-128, with an 8 or 12 GiB graphics card, 181 GB, the closest to the original.

Your RAM Graphics card Pack file Memory in use Total on disk
64 GiB 8, 12 or 16 GiB Hy4-Flash-SQ-64.strata 59.4 GiB 129.61 GB
96 GiB 8 or 12 GiB Hy4-Flash-SQ-96.strata 85.1 GiB 157.16 GB
128 GiB 8 or 12 GiB Hy4-Flash-SQ-128.strata 107.5 GiB 181.23 GB

"Graphics card" is the card's memory. "Memory in use" is RAM plus graphics memory, approximately. "Total on disk" is the pack plus the shared lookup file every pack needs, Hy4-Flash-Engram-Q8.strata, for text. File sizes and checksums are below.

  • 64 GiB of RAM and an 8 GiB graphics card are enough to start. Take SQ-64.
  • 128 GiB of unified memory? SQ-128 is the closest match to the original. SQ-96 leaves about 22 GiB free for other work.

Get started

  1. Install Strata. Download the app for Windows, Linux or Mac. It needs no account.
  2. Choose Hy4-Flash and the pack that matches your RAM.
  3. Start chatting.
Prefer the terminal?

Linux or Mac:

curl -fsSL --connect-timeout 15 --max-time 120 https://github.com/Niko1221/Strata/releases/latest/download/install.sh -o install.sh && sh install.sh

Windows PowerShell:

irm -TimeoutSec 120 https://github.com/Niko1221/Strata/releases/latest/download/install.ps1 | iex

On Windows, open a new terminal afterwards. Then run strata launch (or strata launch --family hy4-flash --model SQ-64) and follow the prompts.

How close is it to the original?

SQ-128 stays closest to the original model on both test sets. Each smaller pack gives up some closeness to fit in less memory.

Drift from the original model, mean KL divergence, lower is better. Reasoning prompts: SQ-128 0.088, SQ-96 0.136, SQ-64 0.232. Tool-calling prompts: SQ-128 0.180, SQ-96 0.237, SQ-64 0.518. SQ-128 is the closest on both.

How to read this. Each pack's next-word predictions are compared with the full-precision model (BF16). The score is the average gap between the two, called KL divergence. Zero means identical, so a shorter bar is better.

Pack Reasoning prompts Tool-calling prompts
SQ-128 0.08789 0.17969
SQ-96 0.13591 0.23661
SQ-64 0.23206 0.51828
How this was measured

Each set contains eight held-out prompts of 512 tokens. The evaluation scores all 4,096 prompt positions using the packed model and its required Engram file across all 42 layers. The tool-calling column measures token distributions on tool-related prompts. End-to-end tool success requires a separate evaluation.

Performance with Strata

Up to 240 tokens a second. Speeds measured with Strata on eight machines, from two V100 cards to an RTX PRO 6000.

Output speed with Strata, in tokens per second, with prompt speed beside it. RTX PRO 6000: 102.8 to 240, prompt 4,015. RTX 5090: 90 to 111, prompt 1,694 at 128K context. M5 Max with Metal 4: 69 to 110, prompt 1,215. R9700 or RX 9070 XT: 62, prompt 1,122. Strix Halo 128 GB: 51 to 60, prompt 728. RTX 5080 16 GB: 44, prompt 1,627. DGX Spark GB10: 43.9, prompt 924. Two V100 cards: 41.8, prompt 1,068.

How speeds were measured

Measured with Strata 0.2.18 (strata-hy4 0.1.64, built-in DFly2 speculative decoding on): SQ-96 and SQ-64 on RTX PRO 6000 (4K–104K context); SQ-64 on RTX 5090 (128K context), Apple M5 Max 128 GB (Metal 4, 4K–32K context), Radeon AI PRO R9700 / RX 9070 XT (4K context), Ryzen AI Max+ 395 Strix Halo 128 GB (4K–128K context), RTX 5080 16 GB + 64 GB RAM (4K context), DGX Spark GB10 128 GB (4K context), and 2× Tesla V100 32 GB (4K context).

Recommended settings

Setting Value
temperature 0.9
top_p 1.0
top_k off

Use the same values with thinking on (reasoning_effort high, the default) and for direct answers (no_think).

Files and checksums

File Size
Hy4-Flash-SQ-64.strata 62.03 GB
Hy4-Flash-SQ-96.strata 89.58 GB
Hy4-Flash-SQ-128.strata 113.65 GB
Hy4-Flash-Engram-Q8.strata, needed by every pack 67.58 GB
Hy4-Flash-Vision.strata, for reading images 0.93 GB
Pack file checksums (SHA-256)
6e8ba2729465f38ed42b5bfa7db02c92beddc25ca0413b7f0c24ab9f1e8b51f4  Hy4-Flash-SQ-128.strata
3b1c5755a220c69892b3d2dbad506b8593a30be24b17a390ecbd1642c0c1e14c  Hy4-Flash-SQ-96.strata
b2cc38fe4bded35df82a701f9e868ea09c6fcd9be52c8446289e394f804417b2  Hy4-Flash-SQ-64.strata
ab0cf9b3f9c64bd25f6212d4508e7443ab8b9bc133fbccc7bdfb83510a20c2c1  Hy4-Flash-Engram-Q8.strata
ef09bceff8f01b8d226d2dc56aef09cec719b21b45783ee9b5d64e7241b72a09  Hy4-Flash-Vision.strata

Support Strata

Help keep local AI fast, independent, and open to everyone.
Strata is built independently and released free under the MIT license. Sponsoring on GitHub or buying a coffee directly funds new GPU kernels, model quantizations, and hardware testing.

Sponsor Strata on GitHub   Buy Strata a coffee

Licence

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for strata-engine/Hy4-Flash-Strata

Quantized
(45)
this model