Tencent Hy4 Preview: 770B MoE, 1M Context, Pricing and Benchmark Reality Check

Research checked on August 28, 2026. Tencent released Hy4 preview on August 28 and published the model weights. Its headline specification is unusually large for an openly available model. The backbone has 770B parameters, with 49B active per token. It advertises a 1M-token context window. Tencent Cloud’s API documentation is more precise: 1024k context, 960k maximum input and 64k maximum output. [S1][S2][S9]

The benchmark caveat is central: most of the eye-catching launch benchmarks are still vendor-reported Tencent results. They are useful evidence, but they are not yet independent replications. Tencent also labels Hy4 as an early release and openly lists over-long reasoning and excessive self-verification among its known limitations. [S2]

Key facts

• Release: August 28, 2026
• Architecture: 770B total / 49B active MoE
• Backbone: 78 layers; 256 routed experts + 1 shared expert in each MoE layer; top-8 routed experts active
• Native MTP: 10B total / about 0.7B active
• Context: 1M model context; Tencent Cloud exposes 1024k / 960k max input / 64k max output
• License: Apache 2.0
• APIs: Tencent Cloud TokenHub and OpenRouter
• OpenRouter list price: $0.834/M input, $2.501/M output, $0.042/M cache read
• Tencent Cloud list price shown for Guangzhou: 6 / 18 / 0.3 CNY per M tokens for input / output / cache hit
• Self-hosting: possible, but official serving guidance is firmly in multi-GPU datacenter territory

Architecture: where the 770B number comes from

Hy4 preview is a Mixture-of-Experts model, so “770B parameters” does not mean all 770B participate in every token. Tencent’s backbone uses 78 layers. The first feed-forward block is dense; the remaining 77 are MoE layers with 256 routed experts and one shared expert. Each token selects the top eight routed experts plus the shared expert. The published active-parameter figure is 49B per token. [S2]

A native Multi-Token Prediction (MTP) layer adds 10B total parameters, roughly 0.7B active. Tencent pairs this with Gated DeepSeek Sparse Attention, IndexCache for cross-layer sparse-index reuse, and an iHC residual pathway with four residual streams. [S2][S19]

IndexCache has a separate research paper describing cross-layer reuse of sparse-attention indexing. That paper is valuable for understanding the mechanism, but its speedups should not be pasted onto Hy4 as if they were production end-to-end measurements. The experiments and the released service are not the same object. [S19]

The jump from Hy3 preview

Hy3 preview shipped with 295B total parameters, 21B active parameters and a 256K context window. Hy4 moves those figures to 770B, 49B and 1M. On the raw specifications, that is roughly 2.61× the total parameter count, 2.33× the active parameter count and 4× the nominal context. [S16][S2]

For the earlier generation and its OpenRouter angle, see Tencent Hy3 on OpenRouter.

Tencent attributes the generational improvement to more than scale: it says pre-training, post-training and training-data volume all increased. There is not enough public ablation data to assign a clean share of the gain to each factor. [S1][S2]

“1M context” is real, but deployment limits still matter

OpenRouter lists an exact 1,048,576-token context window and up to 64,000 completion tokens. Tencent Cloud documents 1024k context, 960k maximum input and 64k maximum output for hy4-preview. [S7][S9]

Those numbers can coexist. Model context is a total budget; a service can cap prompt tokens below the window to preserve room for generation and internal behavior.

Self-hosting introduces another constraint: KV-cache memory. SGLang explicitly recommends sizing the context length to the available KV budget and gives substantially smaller reference settings such as 262K on several GPU configurations. Supporting a 1M maximum does not mean a given server configuration can economically serve 1M at high concurrency. [S11]

API compatibility and agent features

Tencent Cloud states that Hy4 preview is compatible with OpenAI Chat Completions, OpenAI Responses and Anthropic Messages. Its guide includes streaming, reasoning controls, function calling and JSON-schema structured output. Reasoning defaults to high; a direct-response/no-think path is available. [S9]

OpenRouter separately lists tool calling and JSON-schema structured outputs for its route. [S7]

Those capabilities make Hy4 a candidate for coding agents and tool-heavy workflows. Compatibility still needs application-level testing: parser conventions, reasoning fields, tool-call edge cases and retry behavior can differ even when the top-level protocol looks familiar.

What Hy4 preview means for Mac users

For a Mac, Hy4 preview is primarily a cloud model or remote server, not a local desktop model. The Mac can still be the useful control plane: it selects files, runs the agent, executes approved shell commands and records costs and failures. Inference happens at the API provider or on a remote multi-GPU host.

Goal	Practical Mac path	Check before use
Coding agent over a local repository	API client on the Mac, sending only selected files	privacy, tool permissions, retries and token cost
Offline or confidential work	smaller local model through Ollama, MLX or llama.cpp	whether web search, plugins and cloud fallbacks are truly disabled
Self-host Hy4	Mac as development and control machine, serving on remote GPUs	SSH/API security, model version, KV cache and infrastructure cost
Test the 1M context	start with a controlled document against the API limit	actual input cap, cache rules, latency and retrieval quality

The key data path is the prompt, not the weight download: source code, screenshots, terminal output and tool results leave the Mac as soon as the API route is used. A local agent can select and redact that data, but it does not make Hy4 a local model. For a fair comparison, run the same Mac workflow once with a small local fallback and once with Hy4, measuring success rate, total tokens, tool-call errors, elapsed time and cost per completed task.

Pricing: the cache rate is the number to watch

Tencent Cloud’s displayed Guangzhou pricing is 6 CNY/M input, 18 CNY/M output and 0.3 CNY/M cache hit. OpenRouter lists $0.834/M input, $2.501/M output and $0.042/M cache read. [S8][S7]

The following scenarios apply the published rates. Uncached input is billed at the input rate, and successful cache reads at the cache-read rate.

OpenRouter cost = input_M × 0.834 + output_M × 2.501 + cache_read_M × 0.042

Workload	Uncached input	Output	Cache reads	OpenRouter	Tencent Cloud
Light	2M	0.5M	10M	$3.34	24 CNY
Medium	20M	5M	100M	$33.39	240 CNY
Heavy	200M	50M	1,000M	$333.85	2,400 CNY

For document-heavy agents, cache economics can dominate the bill. Do not project these savings until you have verified the target provider’s cache semantics, eligibility and retention behavior.

Benchmark results: impressive, but label them correctly

Tencent’s launch material publishes a broad set of strong scores. Hugging Face’s automated evaluation-metadata discussion, for example, records 65.7 on SWE-Bench Pro and points back to the model card as the source. A same-day analysis transcribes additional Tencent figures including 85.4 on Terminal-Bench 2.1, 82.9 on SWE-bench Multilingual, 64.3 on DeepSWE, 92.3 on GPQA Diamond, 74.1 on Toolathlon-Verified, 83.7 on MCP-Atlas and 55.4 on Humanity’s Last Exam with tools. [S31][S32]

Benchmark	Hy4 preview	Evidence status on Aug. 28
GPQA Diamond	92.3	vendor-reported
Terminal-Bench 2.1	85.4	vendor-reported
MCP-Atlas	83.7	vendor-reported
SWE-bench Multilingual	82.9	vendor-reported
Toolathlon-Verified	74.1	vendor-reported
SWE-Bench Pro (public)	65.7	vendor-reported
DeepSWE	64.3	vendor-reported
HLE with tools	55.4	vendor-reported
APEX-Agents	37.1	vendor-reported

“Vendor-reported” is not a synonym for false. It is an evidence label. Agentic scores can move with the harness, tools, budget, model settings and benchmark version. Terminal-Bench 2.1 itself revised tasks that had drift or environment problems, illustrating why version and setup matter. [S25]

Tencent’s 163-expert blind test

Tencent also ran a blind internal comparison: 163 internal experts rated outputs on 203 engineering tasks. Hy4 preview averaged 2.99/4, versus 2.92 for GLM 5.3 and 2.94 for Kimi K3. Against GLM 5.3 Tencent reports 46.8% wins, 12.8% ties and 40.4% losses; against Kimi K3, 51.2% wins, 7.9% ties and 40.9% losses. [S2]

The appropriate interpretation is a narrow lead in Tencent’s internal workload, not a decisive universal win. The raters were Tencent employees and the tasks came from Tencent productivity scenarios. That is useful real-work evidence, but not independent evidence.

Open weights, Apache 2.0 — and datacenter-scale serving

Hy4 preview is licensed under Apache 2.0. [S5] The open license is developer-friendly, but the model’s size changes what “self-hosting” means.

SGLang estimates approximately 1.5 TB of BF16 weights and 760 GB for the MXFP8 checkpoint. Its hardware table discusses H200, B200, B300 and GB300 configurations and explicitly budgets KV-cache memory on top of the weights. [S11]

The vLLM recipe’s headline describes 16×B200 or 8×B300 with MTP, and the example serves the FP8 checkpoint with tensor parallelism. [S10]

In other words, Hy4 is downloadable, modifiable and deployable, but it is not a casual local model. Quantization can shrink weight storage; it does not make long-context KV memory disappear, and it may introduce its own quality/kernel trade-offs.

Known limitations are unusually concrete

Tencent lists two launch-time issues: Hy4 can reason longer than necessary on complex tasks and can over-verify its own work. The team also says this is an early version with room left in both pre-training and post-training. [S2]

For production agents, those weaknesses translate directly into metrics that benchmark tables often hide: task completion time, reasoning-token spend, repeated tool calls, redundant checks and cost per successful job.

Is Hy4 preview good?

On August 28, the evidence supports a strong but bounded conclusion: Hy4 preview is a serious open-weight frontier contender with unusually large context, aggressive API pricing and a design aimed at coding, tools and long-horizon productivity work. Its architecture, distribution, API limits and pricing are well documented. Its relative quality ranking is less settled because most launch-day scores are still first-party.

That makes it easy to recommend for controlled API evaluation. It is much harder to recommend a self-hosting project without datacenter hardware, and too early to declare it the best open model solely from the published benchmark chart.

What a reproducible comparison should measure

A reproducible comparison should compare Hy4 preview, GLM 5.3 and Kimi K3 under one harness, with identical tools, token budgets, timeouts and tasks. Report pass rate alongside total tokens, wall-clock time, tool errors, retries, verification loops and cost per completed task. That would turn launch-day benchmark claims into evidence that can support an actual deployment decision.