Tencent HY4 Flash: 224B Parameters, 150+ Tokens/Sec, and Two Free Weeks in HermesNote · digest for November 19
Macro photo of a silicon chip with gold micro-traces and high-speed light streaks
Note3 sections≈ 4 min

Tencent HY4 Flash: 224B Parameters, 150+ Tokens/Sec, and Two Free Weeks in Hermes

A heavy 224-billion-parameter model should normally crawl on servers and burn through your budget in a few prompts. Yet the new Tencent HY4 Flash pushes over 150 tokens per second, and in a live test inside Hermes Agent the status bar held around 160 at rock-bottom API prices. It's a cargo truck moving at sports-car speed on compact-car fuel.

How 224 billion parameters run past 150 tokens per second

The trick is a sparse Mixture of Experts (MoE) architecture. Picture a library holding 224 billion books. A dense model without departments flips through every shelf from the lobby to the basement just to answer a simple email. HY4 Flash works differently: when a prompt arrives, it pulls only the 12 billion "books" needed for that task and leaves the other 200+ billion untouched. Those 12 billion are its active parameters.

Its Hugging Face model card lists 220 billion parameters, including 64 billion in an Engram memory module for frequent phrases. Across 41 of its 42 layers, LatentMoE routes each token—a small piece of a word—through just 16 specialists out of 512 plus one shared expert. IndexPool hybrid attention shrinks conversation memory (the KV cache) by 6x vs. Hy4 and 24x vs. Hy3, while a 0.85B-parameter DFly2 drafter guesses upcoming words for batch verification.

The context window holds 1 million tokens, or roughly 700,000 words—an entire folder of reports and spreadsheets at once. It reads text and images, and handles audio and video in office tests. Weights are open under the Tencent Hy License v1.0: free for regular projects, requiring special permission only for cloud providers earning over $50 million or services above 100 million monthly users.

What Tencent's 259-task blind test actually showed

Time to check the claims of parity with Claude Haiku 5.5 and 18x savings. Tencent ran an internal blind test among 85 employees using its WorkBuddy agent: staff didn't know which model they were using and ran 259 work tasks across 27 categories on HY4 Flash, Claude Haiku 5.5, DeepSeek V4.1 Flash, and GLM 5.3 Flash.

HY4 Flash scored 2.79 out of 4 on satisfaction vs. 2.81 for Claude Haiku 5.5. Quality was nearly a tie, its 6.7-minute median time was the fastest, and its 0.26 yuan task cost was roughly 18x lower than Anthropic's model. Official API pricing sits around $0.070 per million input tokens, $0.281 for output, and $0.003 for cache, while reinforcement learning cut answer length in half. On Nous Portal, paid access costs $0.04 and $0.14, and Tencent Cloud TokenHub offers 50% off through November 30, 2026.

There is a catch to those numbers. Tencent ran the test internally on its own staff with no independent audit. Even there, DeepSeek V4.1 Flash scored higher on satisfaction, though Tencent's newcomer came out 8x cheaper on the WorkIndex office benchmark. Also, Flash is weak on long, complex software engineering—Tencent itself recommends the larger Hy4 for that. In a live demo by Julian Goldie, building a 3D scene failed with a black screen because the model inserted a broken library URL, and it only fixed the scene after a follow-up error prompt.

Two free weeks in Hermes Agent: setup and real tasks

To let people test it free, Nous Research opened access for two weeks through Nous Portal, which feeds models to Hermes. Exact promo end dates and rate limits aren't listed on the page, though the zero-cost tier is active now.

Setup takes a couple of minutes using the Hermes Agent guide. Install the agent in your terminal, run hermes setup --portal to sign in via browser, and check status with hermes portal info. Inside a session, type /model tencent/hy4-flash:free. Watch out for the catalog trap: a paid tencent/hy4-flash entry sits right beside it, and forgetting the :free suffix will charge your balance.

Load this setup with routine work: score 12 B2B leads and draft opening email lines, turn consultation notes into a training module, sort 8 support tickets with draft replies, or build a call plan from CRM data for 15 accounts. And if your main expensive model hits a rate limit or fails mid-task in Hermes, just switch /model on the fly—free HY4 Flash picks up the stalled content plan from the exact same context and finishes the job.

Want a deep diveVotes help us choose what to break down in detail. This material has 0 votes.
Sources · 6