I spent last night reading the Tencent Hy4 model card, its LICENSE file and Tencent Cloud's pricing page, because the headline going around ("770B open model, Apache 2.0") is half wrong. The weights are out. The license is not Apache 2.0. If you build products on open models, that difference matters more than any benchmark chart.
I wrote this with help from AI tools and checked every price and claim against the linked sources myself. The cover image is AI-generated.
Watch the video version: Tencent Hy4 Open Weights: Price, What Changed, and When to Use It on my channel Daily AI News By Ali. Earlier context: Hy4 Flash and Qwen4-Max news from Nov 16 and my Gemini 4 Neon price breakdown.
Tencent Hy4 open weights in brief
Short answer: Tencent released Hy4 weights on November 20, 2026: 770B total and 49B active parameters, a 1M-token context, on Hugging Face. It ships under the Tencent Hy License, which is derived from Apache 2.0 but is not Apache 2.0. The API costs $0.834 in and $2.501 out per 1M tokens, half price until Dec 31.
1. What Tencent actually released
What happened. According to the Hy4 model card on Hugging Face, Hy4 is a mixture-of-experts model with 770B total and 49B active parameters for text, plus a 1.23B vision encoder. It has 78 layers (1 dense, 77 MoE), 256 routed experts plus 1 shared expert with top-8 routing, a 1,048,576-token context and a 3.3B draft model (DFly2) for speculative decoding. Three checkpoints are listed: Hy4, Hy4-FP8 and an Hy4 base model that is not instruction-tuned.
AI Frontier Post reports the weights were announced at 03:52 UTC on November 20. Hy4 had launched API-only on November 4, and Hy4 Flash weights came out on November 17.
Tencent's own claims (internal evals, not independent tests): hallucination down from 15.76% to 5.45% versus the preview, speculative errors down from 23.1% to 5.6%, and #1 among open-weight models on a 579-task finance eval graded by 213 professionals. On Tencent's own chart, Hy4 still trails GPT 6.1 Sol on OSWorld 2.1 (78.9 vs 81.8) and on Terminal Bench 4.0 (44.4 vs 66.4).
Why it matters. This is one of the largest open-weight models you can download today with a 1M context. For teams in finance or document-heavy work, the model card explicitly targets co-editing Word, Excel and PowerPoint files.
What to do. Treat the benchmark numbers as vendor claims. If agentic coding is your main use case, Tencent's own chart says a closed model is still clearly ahead on Terminal Bench.
2. The license trap: "open weights" is not "Apache 2.0"
What happened. TechNode wrote that the flagship is "available under the Apache 2.0 license." The LICENSE file itself says the Tencent Hy License v1.0 is "derived from the Apache License, Version 2.0, with modifications" and, in plain words, "This License is not the Apache License, Version 2.0."
The key modification is Section 10, on model-as-a-service (MaaS). You need a separate written license from Tencent if your worldwide gross revenue goes above US$50M in any 12 months, or your product passes 100M monthly active users. Internal use is exempt (10.3), and MaaS excludes end-user products where the model is confined to specific features, and mere relaying.
Why it matters. The preview was Apache 2.0, per AI Frontier Post. Many teams will assume the final release kept that license. Legal review processes that whitelist "Apache 2.0" will wave this through by mistake.
What to do. Add the actual LICENSE file to your model registry, not the headline. If you sell an API or platform built on Hy4 and you are anywhere near those thresholds, talk to Tencent before you launch, not after.
3. Price, real bill math, and hardware
What happened. Tencent Cloud TokenHub pricing (updated Nov 19) lists Hy4 at $0.834 input, $2.501 output and $0.042 cached input per 1M tokens, with 50% off until 23:59:59 on Dec 31, 2026, Beijing time. Hy4 Flash is listed at $0.07 / $0.281 / $0.003, with a 50% discount running until noon on Nov 30, Beijing time. LiteLLM's post confirms model ID hy4, up to 960,000 input and 64,000 output tokens.
Here is my own math with those official prices, for a workload of 10M input and 2M output tokens a month, no caching:
| Model | Input / output per 1M | Monthly bill (10M in, 2M out) |
|---|---|---|
| Hy4 Flash (listed rates) | $0.07 / $0.281 | ~$1.26 |
| Hy4 (promo, to Dec 31) | $0.417 / $1.25 | ~$6.67 |
| Hy4 (list) | $0.834 / $2.501 | ~$13.34 |
| Claude Sonnet 5.5 (per VentureBeat, Nov 18) | $2 / $10 | $40 |
| Gemini 4 Neon | $3 / $15 | $60 |
At list price, Hy4 costs about a fifth of Gemini 4 Neon for this workload.
Self-hosting reality check. AI Frontier Post puts the download sizes at about 1.55 TB (Hy4), 0.79 TB (FP8) and 1.54 TB (base). The model card's vLLM and SGLang recipes use tensor parallel 8. Even the FP8 weights (~0.79 TB) are bigger than the 512 GB maximum memory of a Mac Studio, so "run it on one box under my desk" is off the table. The AngelSlim GGUF is labelled "for end-to-end testing purposes", not production.
Developer takeaway. The model card's quickstart uses an OpenAI-compatible API with model "hy4", temperature 0.9, top_p 1.0, and a reasoning_effort switch of "no_think" or "high" (default high). That means most existing OpenAI-client code only needs a new base URL and model name. Example (illustrative; check the card for exact field placement):
from openai import OpenAI client = OpenAI(base_url="http://localhost:8000/v1", api_key="local") # your vLLM/SGLang server resp = client.chat.completions.create( model="hy4", messages=[{"role": "user", "content": "Summarize this contract clause."}], temperature=0.9, top_p=1.0, extra_body={"reasoning_effort": "no_think"}, # "high" is the default ) Use "no_think" for cheap classification and extraction calls, and keep "high" for multi-step work.
4. Dragos: 2 million bait prompts found 3 breached AI providers
What happened. In "Three breaches, endless leaks" (Nov 19), HD Moore at Dragos describes sending about 2 million prompts containing unique URLs and canary credentials through AI providers and routers between July and November 2026. They found evidence of compromise at three or more AI providers, "a couple hundred" AWS keys stolen in the largest breach, and bait keys tested "in some cases within a few hours." Marketplace resellers logged and kept prompts through standard logging, and some stolen requests were marked zero data retention. As the post puts it, "a privacy flag only binds the first hop."
Why it matters. If your agents paste config files, logs or .env snippets into prompts, every hop between you and the model is a place those secrets can leak.
What to do. Use a first-tier provider directly where you can, be careful with routers, avoid reseller marketplaces, turn on secret filtering, and seed your own free canary tokens (Dragos mentions Thinkst Canarytokens) so you find out fast. Short version: AI Prompts Leaked API Keys.
5. GPT-6.2 drops reasoning when you edit chat history
What happened. OpenAI's "Using GPT-6.2" guide says that starting with GPT-6.2, the Responses API drops reasoning items after mid-conversation edits to the history (prefix edits). The "Keep the prefix unchanged" guide adds that the model may redo reasoning, costing more tokens, money and latency, and quality may drop.
Why it matters. A lot of agent frameworks rewrite earlier messages, for example to trim context or change reasoning effort mid-session. On GPT-6.2 that silently throws away work, and per the prompt caching guide cache reads on GPT-6.2 Astra/Sol cost 0.05x input, so a broken prefix is expensive.
What to do. Only append. To change reasoning effort, append a configuration_update item and keep the request-level setting the same. OpenAI's input_transformations and prompt_cache_options.comparison_response_id diagnostics help you spot where a prefix changed. One-minute version: GPT-6.2 One-Line Fix.
What to watch next
- [ ] Independent benchmarks for Hy4 against Gemini 4 Neon and Claude Sonnet 5.5
- [ ] Whether Tencent clarifies the license wording after the Apache 2.0 headlines
- [ ] Hy4 Flash discount ending Nov 30 and Hy4 promo ending Dec 31 (rerun your cost math)
- [ ] Audit your agents for secrets in prompts and for prefix edits before moving to GPT-6.2
FAQ
Is Tencent Hy4 open source under Apache 2.0?
No. The weights are downloadable, but the Tencent Hy License v1.0 says it is not the Apache License 2.0. Large model-as-a-service providers (over US$50M revenue or 100M MAU) need a separate written license.
How much does the Hy4 API cost?
$0.834 input, $2.501 output and $0.042 cached input per 1M tokens on Tencent Cloud TokenHub, with 50% off until Dec 31, 2026 (Beijing time).
Can I run Hy4 locally?
Not on a single consumer machine. Even the FP8 checkpoint is about 0.79 TB, and the official vLLM/SGLang recipes use tensor parallel 8.
What changed in GPT-6.2 for chat history?
Editing earlier messages drops reasoning items. Append a configuration_update item instead of editing the request when you change reasoning effort.
Where can I watch daily AI news?
On my YouTube channel Daily AI News By Ali.
Follow & Connect
- X: @AliMughal887
- GitHub: Muhammad Ali Mughal (@ali111887)
Watch on YouTube: Daily AI News By Ali · Resume Boost AI · PixelValut
