TL;DR: Tencent’s Hy4 Preview (released August 28, 2026) is a 770B-total / 49B-active MoE under plain Apache 2.0 with a 1M-token context — on paper the largest permissively licensed open-weight model you can download. In practice it is not a home-lab model and will not become one soon: the official FP8 build is ~814GB, no GGUF quants exist as of October 27, 2026, and even an 8×H100 80GB node (640GB) cannot load it. For almost everyone, the right way to use Hy4 is the API at roughly $0.75–0.83 in / $2.25–2.50 out per million tokens. If you want a big Apache 2.0 Hunyuan model you can actually run, that is Hy3 295B, not Hy4.

	Hy4 Preview (770B-A49B)	Hy3 (295B-A21B)	GLM-5.3-Flash (320B-A18B)
License	Apache 2.0, no strings	Apache 2.0, no strings	MIT
Context	1M tokens	256K	200K-class
Smallest official build	FP8, ~814GB	FP8, ~300GB	>100GB-class at useful quants
GGUF / llama.cpp	None as of Oct 27, 2026	Yes, mainline since Jul 2026	Yes
Realistic self-host floor	16×B200 / 8×B300 node	96–128GB unified-memory box	128GB+ rig or rented GPU
Verdict	API model	Self-hostable	Rent or big rig

Honest take: Hy4 is an API model wearing an open-weight license. The license still matters — you can audit it, fine-tune it, and nobody can take it away — but “can download” and “can run” are 814GB apart.

Is Tencent Hy4 really Apache 2.0?

Yes, by every report we could check — with one verification caveat. Hy4 Preview shipped August 28, 2026 on Hugging Face as tencent/Hy4-Preview, and multiple independent trackers describe the license as the standard Apache License 2.0 text with no acceptable-use appendix, no revenue threshold, and no geographic restrictions (ibl.ai, DataLearner). That matches Tencent’s pattern: Hy3’s April 2026 preview carried geo limits that were lifted at the July full release, and the full Hy3 we verified in September is clean Apache 2.0.

The caveat: huggingface.co was unreachable from our verification environment in October 2026, so we could not read the LICENSE file in the repo directly. Before building a commercial product on Hy4, open huggingface.co/tencent/Hy4-Preview yourself and confirm the LICENSE file is the unmodified Apache 2.0 text — it takes two minutes. Recent big-MoE releases have been a licensing minefield (GLM-5.3’s $10B-revenue clause, MiniMax’s non-commercial terms, Kimi’s modified MIT), so the two-minute check is never wasted. Our open-weight license breakdown covers how those clauses bite.

What is the Hy4 architecture?

Hy4 Preview is a 770B-parameter mixture-of-experts model that activates ~49B parameters per token. From the model card specs reproduced across trackers: 78 transformer layers (the first uses a dense FFN, the remaining 77 are MoE), 256 routed experts with top-8 routing plus 1 shared expert per MoE layer, a native multi-token-prediction (MTP) layer (~10B parameters total, ~0.7B active) for speculative decoding, a ~1M-token context window, and 64K max output (eesel.ai, Bifrost model library).

The MoE math cuts the same two ways it did for Hy3, and only one of them is friendly:

• Compute scales with active parameters. Per token, Hy4 does roughly the work of a 49B dense model. On hardware that can hold it, it is fast for its class, and the MTP layer adds speculative-decoding speedups on top.
• Memory scales with total parameters. The router can pick any of the 256 experts for the next token, so all 770B weights must be resident. BF16 weights are ~1.42 TiB across 131 safetensors shards; Tencent’s own FP8 build is ~814GB. There is no router trick that shrinks this.

How much memory does self-hosting Hy4 actually take?

More than any single machine you can buy has. The official vLLM recipe for Hy4 Preview lists 16×B200, or 8×B300 with MTP enabled (vLLM recipes) — a data-center node in the mid-six-figures. The comparison everyone actually cares about: an 8×H100 80GB node is 640GB of VRAM, and the FP8 build alone is ~814GB, so the most common rentable GPU node cannot load Hy4 at all, before you spend a single byte on the 1M-token KV cache.

What about GGUF quants on a big-RAM box, the way Hy3 became viable? As of October 27, 2026, no Hy4 GGUFs exist. Neither Unsloth nor bartowski has published quants, and we found no evidence that llama.cpp supports the Hy4 architecture — Hy3’s hy_v3 support took a dedicated PR, and Hy4’s layout (dense first layer, new MTP config) would need its own. If quants do land, scaling Hy3’s published GGUF sizes by the 2.6× parameter ratio gives rough estimates: ~225GB for a 1-bit quant, ~280GB for IQ2-class, ~410–470GB for Q4-class. Treat those as estimates, not downloads — and note that on Hy3, the 1-bit quant already degraded reasoning noticeably. A 1-bit 770B model on a 256GB server is an experiment, not a deployment.

The boundary cases, stated plainly:

• 24GB GPU (RTX 3090/4090): not in the same universe. Nothing to configure.
• 128GB unified memory (Strix Halo, Mac Studio): cannot run Hy4 even at 1-bit. These boxes top out around Hy3 295B, which is the model Tencent built for them.
• 512GB–1TB RAM server, CPU inference: theoretically possible once GGUFs exist; at ~49B active parameters, expect low single-digit tokens/second. For 1M-context work, unusable.
• 16×B200 cluster: you are a lab or an inference provider, and you already have the vLLM recipe.

For what the realistic hardware tiers cost in late 2026 — and why DRAM pricing makes big-RAM builds worse value than a year ago — see the local AI hardware guides at runaihome.com.

What does the Hy4 API cost compared to self-hosting?

As of October 2026, OpenRouter trackers list Hy4 Preview at $0.75–0.83 per million input tokens and $2.25–2.50 per million output tokens (AnotherWrapper, Bifrost — the two trackers disagree slightly, so check the live OpenRouter page). That is frontier-tier capability at roughly a tenth of proprietary frontier pricing, which is exactly the leverage an Apache 2.0 release is designed to create.

Run the break-even and the verdict writes itself. A heavy solo user burning 50M input + 10M output tokens a month pays roughly $60–66/month on the API. The cheapest hardware that could even theoretically hold a hypothetical Q4 GGUF is a ~512GB-RAM server; the rentable node that actually runs it (16×B200) costs more per day than that user’s API bill for a year. There is no utilization level at which a home lab wins this one — the math that works for Hy3 on a 128GB box you already own never starts for Hy4.

The exception that justifies self-hosting anyway is the usual one: data that cannot leave the building. If that is your constraint, you are shopping for a multi-node GPU cluster, and the practical middle step is renting one first — Vast.ai lists multi-GPU H100-class nodes at floating market rates, which lets you benchmark Hy4 under vLLM for a few hours before anyone signs a purchase order.

When should you not use Hy4 at all?

• You want a local coding backend. 1M context sounds perfect for codebases, but you will be using it through an API anyway — at which point latency, tooling, and price matter more than parameter count. Compare it against the BYOK options in the coding-tool guides at aicoderscope.com before defaulting to the biggest model.
• Your workload fits a model you can own. Hy3 295B (Apache 2.0) and GLM-5.3-Flash (MIT) cover most of the quality gap at a fraction of the footprint, and both have real GGUF ecosystems today.
• You need stable behavior. This is a Preview checkpoint. Tencent’s own Hy3 preview-to-release cycle changed the license terms and the weights; expect Hy4’s final release to differ. Pin the exact revision if you build anything against it.
• Benchmark-driven decisions. The “beats GLM-5.3 and Kimi K3” claim circulating in coverage traces to Tencent’s internal blind evaluation (163 in-house raters, 203 tasks). Internal evals are not independent benchmarks; wait for third-party numbers before re-architecting anything.

What to actually do

Prices as of October 2026, all verified in the sections above:

Your situation	The move	Cost	Where
Want Hy4’s quality today	API via OpenRouter	~$0.75–0.83 in / $2.25–2.50 out per 1M tokens	OpenRouter
Want the biggest Apache 2.0 model you can own	Hy3 295B on a 96–128GB box	hardware you may already have	Hy3 guide
Data can’t leave the building — evaluating a cluster buy	Rent a multi-GPU node first, benchmark under vLLM	floating market rates, by the hour	Vast.ai
Just need a strong MIT model with GGUFs	GLM-5.3-Flash	free weights	GLM-5.3 guide

FAQ

Can I run Tencent Hy4 on a single GPU? No. The smallest official build (FP8) is ~814GB, and no GGUF quantizations exist as of October 27, 2026. Even a hypothetical 1-bit quant would need roughly 225GB of memory — ten 24GB GPUs’ worth. The smallest hardware Tencent’s vLLM recipe supports is a 16×B200 or 8×B300 node.

Is Hy4 the same as Hy3 but bigger? No. Hy3 is 295B total / 21B active with 192 experts and 256K context; Hy4 Preview is 770B / 49B active with 256 experts, a 1M-token context, and a larger MTP layer. Hy3 has mainline llama.cpp support and a full GGUF ecosystem; Hy4 has neither yet. For self-hosters the practical difference is binary: Hy3 runs on a 96–128GB machine, Hy4 does not run outside a data center.

Will Hy4 get GGUF quants like Hy3 did? Probably eventually, but it changes less than you’d hope. llama.cpp would first need support for the Hy4 architecture (Hy3’s took a dedicated PR in July 2026), and even then the estimated 1-bit footprint (~225GB) exceeds every consumer platform including 128GB Strix Halo and 192GB Mac Studio configurations. Watch the Unsloth and bartowski Hugging Face pages — if quants appear, the viable audience is 256GB+ RAM servers accepting single-digit tokens/second.

Sources

• ibl.ai — Tencent Hy4: open-source 770B and vendor pricing leverage (license analysis, FP8/BF16 sizes, 8×H100 math)
• DataLearner — Hy4 Preview model page (Apache 2.0 for code and weights, shard count)
• vLLM recipes — tencent/Hy4-preview (official serving configuration: 16×B200, 8×B300 with MTP)
• eesel.ai — Tencent Hy4 specs and benchmarks (architecture details, internal-eval caveat)
• AnotherWrapper — Hy4 Preview pricing (OpenRouter pricing, October 2026)

What self-hosting actually costs

Real cost breakdowns for self-hosted AI: hardware floors, power, maintenance hours, and the honest comparison against paying for it. No spam, unsubscribe anytime.

Email address Subscribe