Tencent Hy4 Preview Pricing: 770B Open Weights at $0.834/M

Published August 28, 2026By ABD Legacy LLC
Tencent Hy4 pricing open weight model pricing 1M context model best open model for agents

What is Tencent Hy4 preview and how much does it cost?

Tencent Hy4 preview is a 770-billion-parameter Mixture-of-Experts model (49B active per token) with a 1M-token context window, released and open-sourced on August 28, 2026 under Apache 2.0. On OpenRouter it lists at $0.834 per 1M input tokens and $2.501 per 1M output tokens (cached input $0.042), with Tencent Cloud as the sole provider. Tencent's domestic pricing is about ¥6 per 1M input and ¥18 per 1M output — roughly 25–36% below GLM-5.3 and 70–82% below Kimi K3. Bloomberg covered the release the same day, and Tencent positions the model as beating Z.ai's GLM-5.3 and Moonshot's Kimi K3 in its internal blind evaluation (2.99 vs 2.92 vs 2.94 out of 4.00).

On August 28, 2026, Tencent released and open-sourced Hy4 preview — its largest model yet, 53 days after Hy3. The launch post ended with an unusual sign-off: "Use it. Tell us what breaks." The headline numbers are 770B total parameters, 49B active per token, a native context window exceeding 1 million tokens, and Apache 2.0 weights with no field-of-use clause and no geo carve-out.

For AI agencies, this is a pricing-benchmark event, not just a model release. Hy4 preview lands mid-ladder in the open-weight class: cheaper than GLM-5.3 and Kimi K3, still pricier than DeepSeek V4 Pro's off-peak rates, and far cheaper than the paid frontier — with a 1M context window that only a handful of models offer at any price. That is exactly the shape of data point that changes what you quote for long-context agent work.

The pace: Hy3 to Hy4 in 53 days

Context for how fast this market moves: Tencent's Hy3 shipped on July 6, 2026 at 295B total parameters, 21B active, and a 256K context window. Fifty-three days later, Hy4 preview is 2.6x larger with 4x the context. Tencent frames the jump as "We scaled Hy4 preview on three fronts: model size, context length, and training data" — a scaling release, not a claimed algorithmic breakthrough. The iteration speed in Chinese open weights has not slowed, and pricing pressure on the whole open-weight ladder keeps compounding.

Hy4 preview specs at a glance

SpecValue
DeveloperTencent Hy Team (Tencent Hunyuan)
ReleaseAugust 28, 2026 — open-sourced same day
Model size770B total parameters, 49B active per token (MoE)
Architecture78 layers (1 dense FFN + 77 MoE layers); 256 routed + 1 shared expert per MoE layer, top-8 routed; Gated DeepSeek Sparse Attention with IndexCache; identity Hyper-Connections; built-in MTP layer (10B total / 0.7B active) for speculative decoding
Context window1,048,576 tokens (1M+); 64,000 max output
InputsText only — this release is not a unified multimodal checkpoint
WeightsHugging Face tencent/Hy4-preview (+ FP8 variant), ModelScope, GitCode, CNB — Apache 2.0
API routestencent/hy4-preview on OpenRouter; Tencent Cloud sole provider
Deployment pathsvLLM (vllm/vllm-openai:hy4-preview) and SGLang (lmsysorg/sglang:hy4-preview) prebuilt images, tensor-parallel 8, speculative decoding enabled

Hy4 preview pricing vs the open-weight class and frontier

The price is the adoption driver. Hy4 preview undercuts GLM-5.3 by roughly 40% on input and 43% on output, and undercuts Kimi K3 by about 6x on output. Cached input at $0.042/M is a ~20x discount off the input rate — the number that matters for agentic workloads that resend a large system prompt and a growing tool-call history on every turn.

ModelInput per 1MOutput per 1MNotes
Tencent Hy4 preview$0.834$2.501Apache 2.0; cached input $0.042; 1M context; OpenRouter + Tencent Cloud
DeepSeek V4 Pro (off-peak)$0.66$1.98Cache read $0.022 — still the price floor; peak rates ~2x
GLM-5.3-Flash (Z.ai)$0.15$0.50MIT weights; promo $0.075/$0.25 through Sep 9, 2026
GLM-5.3 (Z.ai flagship)$1.40$4.40Hy4 preview is ~40% cheaper on input, ~43% on output
Qwen 3.8 Max (Alibaba)$2$6Open weights since Aug 12; license threshold $50M/12mo
Kimi K3 (Moonshot)$3$15Open weights; Hy4 preview is ~6x cheaper on output
GPT-5.6 Sol (OpenAI)$4$20Promo through Nov 21, 2026; cached input $0.40

Rates as published Aug 28, 2026. Hy4 preview rates are OpenRouter's published list (Tencent Cloud sole provider). DeepSeek V4 Pro rates are DeepSeek's official off-peak rates effective Aug 16, 2026. Tencent's domestic pricing (~¥6/$18 per 1M) reported via KuCoin flash and Tencent materials; international rates on OpenRouter differ.

Run the per-task math on a typical agent workload — 10K input tokens, 2K output tokens per task:

On this shape, Hy4 preview is about 25% above DeepSeek V4 Pro off-peak, 6x below GPT-5.6 Sol, and the GLM-5.3-Flash promo keeps its price edge for now. The real Hy4 argument is the long tail: on a 100K-token input / 5K-token output task (the shape that stresses context, not price), Hy4 preview's 1M window removes the chunking and re-planning cost that short-context models add — which is often the larger line item than raw tokens. For a 100,000-task monthly delivery at the 10K/2K shape, the difference between Sol and Hy4 preview is about $6,670/month in raw token cost ($8,000 vs $1,330) before retries and overhead.

Benchmarks: tier parity on a vendor-selected test, mixed in public

Tencent ran a blind evaluation with 163 internal experts scoring 203 engineering tasks. Hy4 preview averaged 2.99 out of 4, against Kimi K3 at 2.94 and GLM-5.3 at 2.92. A 0.05-point margin, on tasks the vendor selected, graded by the vendor's own employees, is not a result — it is a statement that the three models are interchangeable on this workload. Read it as: Hy4 preview reached the GLM-5.3 / Kimi K3 tier, which is a real achievement, and stop there.

Tencent's published model card reports frontier-adjacent numbers: GPQA Diamond 92.3, HLE (High + Tools) 55.4, SWE-Bench Multilingual 82.9, SWE-Bench Pro 65.7, DeepSWE 64.3, Terminal-Bench 2.1 85.4, MCP-Atlas 83.7, CyberGym 78.4, Office QA Pro 66.2. The agentic column — Terminal-Bench 2.1 at 85.4 and MCP-Atlas at 83.7 — is the one worth caring about if you run tool-calling loops rather than chat.

Public benchmarks are mixed, and Tencent does not lead everywhere: coverage notes Hy4 preview still lags GLM-5.3 on code and cybersecurity tests like DeepSWE and CyberGym. The honest read: strong all-rounder with standout agentic scores, not a clean sweep. All numbers are vendor-reported until independent replication.

Architecture: borrowed strengths, 1M context that is native

Two things stand out for anyone evaluating long-context work. First, the attention module uses Gated DeepSeek Sparse Attention with an IndexCache — a direct borrowing from the lab whose pricing has set the floor for this market — so sparse indices are reused across layers rather than recomputed. That is what makes a 1M context economically serveable at all. Second, the MTP layer ships speculative decoding in the weights (vLLM and SGLang both enable it in Tencent's prebuilt images), so the speed story is partly pre-built rather than stack-dependent.

The 1M context is native per the model card, supported by the sparse-attention design rather than by rope-scaling a shorter model — a better sign than most 1M claims. It is still worth running a needle-in-haystack test at your actual working depth before trusting it.

Can you actually serve it? The 770GB answer

Short answer: not on hardware you own, unless you own a lot of it.

PrecisionWeights aloneFits on 8xH100 (640GB)?Fits on 8xH200 (1.1TB)?
BF16~1.54 TBNoNo (no room for KV cache)
FP8~770 GBNoYes, with room for KV cache

770B parameters at one byte each is 770GB before you allocate a single token of KV cache. An 8xH100 node tops out at 640GB of VRAM, so the FP8 checkpoint does not fit — you need H200-class cards or multi-node tensor parallelism. This is the practical difference between Hy4 and Hy3's single-GPU GGUFs: Hy3 at 295B was squeezable; Hy4 at 770B is not, and no amount of llama.cpp cleverness makes a 770B model a laptop model. Tencent recommends vLLM and SGLang for production, both with prebuilt images exposing an OpenAI-compatible endpoint. For most teams that means the API, not the weights. The Apache 2.0 license still matters — no vendor can revoke your right to run it, and you can fine-tune and redistribute — but "open weights" and "weights you can afford to serve" are different things at this scale.

The caveats Tencent published themselves

This is the most credible part of the release, and it deserves quoting in full:

"This is an early version of Hy4. There is real headroom left in both pre-training and post-training, and we are shipping with known issues — among them, spending longer than necessary reasoning through complex tasks, and a tendency to over-verify its own work."

Vendors rarely name their model's failure mode on the model card. Take it seriously, because it compounds with a second number: ~36 tokens/sec at P50 and 3.19s P50 latency as observed on OpenRouter, where Tencent Cloud is currently the sole provider. A model that over-reasons, at 36 tok/s, billed per output token, is a bad fit for a tight agent loop — the over-verification tax lands on your latency and your bill simultaneously.

Also worth knowing before you route production traffic: OpenRouter recorded roughly 86% availability over the model's first three days. That is a day-one capacity story, not necessarily a lasting one, but do not make Hy4 a single point of failure this week. Other release notes: this is a text model (Hunyuan's image/video/3D models are separate), and the "preview" label signals the checkpoint will be replaced — Tencent says it will iterate as quickly as it did on Hy3.

What this changes in the calculator

We have added Tencent Hy4 preview as a selectable model strategy in the AI Agency Pricing Calculator — modeled at OpenRouter list $0.834/$2.501 with the Apache 2.0 weights, the sole-provider and preview-status caveats flagged, positioned between the cache-aware DeepSeek V4 Pro and the plain open-weight stack. The practical takeaway for your quoting:

Price your AI work against current market benchmarks

Try the AI Agency Pricing Calculator →

Estimate setup fees, retainers, and margin in under a minute — then model per-agent workload costs with the AI Agent API Cost Calculator.

Frequently asked questions

What is Tencent Hy4 preview and how much does it cost?

Tencent Hy4 preview is a Mixture-of-Experts flagship model released and open-sourced on August 28, 2026: 770B total parameters with 49B active per token, a context window exceeding 1 million tokens, and Apache 2.0 weights. On OpenRouter it lists at $0.834 per 1M input tokens and $2.501 per 1M output tokens (cached input $0.042 per 1M), with Tencent Cloud as the sole provider. Tencent's domestic pricing is about ¥6 per 1M input and ¥18 per 1M output — roughly 25–36% below GLM-5.3 and 70–82% below Kimi K3.

Is Tencent Hy4 preview open source?

Yes — Tencent released Hy4 preview weights on August 28, 2026 under the Apache 2.0 license, with no field-of-use clause and no geo carve-out. The weights are available on Hugging Face (tencent/Hy4-preview plus an FP8 variant), ModelScope, GitCode, and CNB. "Open weights" is the precise description: you get the model weights, deployment recipes, and fine-tuning pipeline, not the full training data. Note that 770B parameters is a cluster-scale deployment — the FP8 checkpoint alone is about 770GB, which does not fit on an 8xH100 node (640GB); you need H200-class or multi-node tensor parallelism.

How does Tencent Hy4 preview compare to GLM-5.3 and DeepSeek?

Tencent's internal blind evaluation scored Hy4 preview 2.99/4.00 against GLM-5.3 at 2.92 and Kimi K3 at 2.94 — a 0.05-point margin on vendor-selected tasks, which reads as tier parity, not a decisive win. On price, Hy4 preview undercuts GLM-5.3 ($1.40/$4.40) by roughly 40% on input and 43% on output, and undercuts Kimi K3 ($3/$15) by about 6x on output. It is still more expensive than DeepSeek V4 Pro's off-peak rates ($0.66/$1.98 with $0.022 cache reads), so DeepSeek keeps the price floor; Hy4 preview's pitch is 1M context and strong agentic benchmarks at an open-weight price. Public benchmarks are mixed — Hy4 preview lags GLM-5.3 on code and cybersecurity tests like DeepSWE and CyberGym.

What is Hy4 preview's context window and architecture?

Hy4 preview has a native 1M-token context window (1,048,576 tokens) with 64,000 tokens of max output. Architecturally it is a sparse MoE with 78 layers (a dense first layer plus 77 MoE layers with 256 routed experts and 1 shared expert, activating the top-8), built on Gated DeepSeek Sparse Attention with an IndexCache for cross-layer sparse index reuse, identity Hyper-Connections for the residual pathway, and a built-in MTP layer (10B total / 0.7B active) for speculative decoding. The 1M context is native, supported by the sparse-attention design — not rope-scaling a shorter model.

Can agencies self-host Tencent Hy4 preview?

Only with serious hardware. The FP8 checkpoint is about 770GB, so it does not fit on an 8xH100 node (640GB VRAM); you need 8xH200 (1.1TB) or multi-node tensor parallelism. Tencent ships official prebuilt vLLM (vllm/vllm-openai:hy4-preview) and SGLang (lmsysorg/sglang:hy4-preview) images with tensor-parallel 8 and speculative decoding enabled, both exposing an OpenAI-compatible endpoint. For most agencies the API is the practical route — Apache 2.0 matters because no vendor can revoke the right to run it, but "open weights" and "weights you can afford to serve" are different things at 770B scale.

What are the caveats of Tencent Hy4 preview?

Tencent itself flags it as an early version: real headroom in pre-training and post-training, plus known issues including spending longer than necessary reasoning through complex tasks and a tendency to over-verify its own work. Independent observations add two more: roughly 36 tokens/sec at P50 with 3.19s latency on OpenRouter's sole provider (Tencent Cloud), and about 86% availability over the first three days. A model that over-reasons, at 36 tok/s, billed per output token, is a bad fit for a tight agent loop — the over-verification tax lands on latency and bill simultaneously. Do not make it a single point of failure this week.

Related reading

Sources

Accuracy note: Model specs (770B/49B, 1M context, 64K output, Apache 2.0) are from Tencent's official release, the GitHub README, and the Hugging Face model card, verified Aug 28, 2026. OpenRouter list pricing ($0.834/$2.501, cached input $0.042) is from OpenRouter's published provider page; Tencent's domestic pricing (~¥6/$18 per 1M) is per KuCoin flash reporting — international rates differ. The internal blind evaluation (2.99 vs 2.92 vs 2.94 on 203 tasks, 163 experts) is Tencent's own, vendor-selected and vendor-graded; public benchmarks are mixed and Hy4 preview lags GLM-5.3 on DeepSWE/CyberGym. Throughput (~36 tok/s P50, 3.19s latency) and first-three-days availability (~86%) are OpenRouter observations at release, subject to change as capacity scales. All benchmark scores are vendor-reported; no independent replication exists as of Aug 28, 2026. Per-task math uses 10K input / 2K output tokens; real agent workloads vary with retries, fan-out, and cache behavior — model your own volume in the calculator.