Tencent Hy4 Preview: 770B Open MoE, Agentic Benchmarks, and Where It Fits in 2026

Last updated: August 31, 2026

Official Sources#

On August 28, 2026, Tencent open-sourced Hy4 preview, the biggest open-weights release of the month: 770B total parameters with 49B active per token, a 1M-token context window, Apache 2.0, and weights on Hugging Face, ModelScope, GitCode, and CNB. It is up 2.6x on total parameters from July's Hy3 (295B, 21B active) and roughly 2.6x on context (256K to 1M), and it is text-only - no vision - which is the first honest signal about what it is for.

This post is the decision-intent read: what the architecture actually is, what the benchmarks posted so far say and what they do not, what it costs on the API, what running it yourself really takes, and the honest cases for staying on the cheaper end of the open-weights field.

What Shipped#

Hy4 preview is a generation jump over Hy3 by every number that matters:

Architecture details from the model card: 78 layers, the first dense and the remaining 77 MoE with 256 routed experts plus one shared expert, activating the top-8 routed experts per token. Attention uses Gated DeepSeek Sparse Attention (the same sparse-attention family DeepSeek published) with IndexCache for cross-layer sparse index reuse, and the residual pathway uses identity Hyper-Connections. A native MTP layer (10B total, 0.7B active) is built in for speculative decoding. Tencent also ships Hy4-preview-FP8, the FP8-quantized instruct weights, as the primary deployment target.

The Apache 2.0 license matters for teams that DeepSeek's MIT license already brought into the open-weights camp: these weights can go into commercial products with zero attribution paperwork, and the FP8 release signals Tencent expects self-hosting to be the main path for most users.

Reasoning: Exactly Two Levels#

The single most interesting design choice is the reasoning control. The chat template accepts exactly two values - high (the default, deep chain-of-thought) and no_think - and raises an exception for anything else. There is no medium, no budget slider, no max-effort tier.

from openai import OpenAI
import os

client = OpenAI(
    base_url="http://127.0.0.1:8000/v1",
    api_key=os.getenv("OPENAI_API_KEY", "EMPTY"),
)

response = client.chat.completions.create(
    model="hy4-preview",
    messages=[{"role": "user", "content": "Explain the fix for this race condition."}],
    temperature=0.9,
    top_p=1.0,
    extra_body={"chat_template_kwargs": {"reasoning_effort": "no_think"}},
)
print(response.choices[0].message.content)

The recommended sampling parameters are temperature=0.9, top_p=1.0 per the model card, and the no_think path is for direct responses where a full reasoning trace is wasted. For agentic work, high is the mode that writes code and fixes tests; the flip side is the model card's own known-limitations note: Hy4 preview "spend[s] longer than necessary reasoning through complex tasks, and [has] a tendency to over-verify its own work." Simon Willison's first look found the reasoning trace itself runs in truncated, slightly imperfect English - legible, token-cheap, and a reminder that hidden reasoning text is not editorial prose.

Two levels means your routing layer gets a clean binary: reason or do not. That is coarser than Claude's five effort levels or Qwen's exhaustive thinking controls, and it is actually a reasonable default for agent inner loops - but plan for it, because there is no "medium" to step down to in between.

What the Benchmarks Posted So Far Say#

Treat every number below as vendor or community-run, on harnesses and sampling choices Tencent or third parties made, for a model explicitly labeled preview. The model card lists community evaluation results:

The coding rows are the story: Terminal-Bench 2.1 at 85.4 is at the top of the published open-weights pack in that category, and the SWE-bench Multilingual resolved rate of 82.9 is a strong repo-level figure for a text-only model. Tencent also ran a blind side-by-side with internal experts: 163 evaluators rated outputs on 203 engineering tasks, and Hy4 came out "slightly ahead" of GLM 5.3 (2.99 vs 2.92) and Kimi K3 (2.99 vs 2.94) on average score. Read "slightly ahead" literally - these are verdict-scale deltas on a curated task set, not a routing cliff.

The honest caveats: a preview model with admitted over-reasoning tendencies may look better in headless benchmark harnesses than in latency-sensitive production loops, and the real test is your own repo on your own tool surface. The only benchmark that matters is your own task distribution - the same rule that applies to every open-weights release this year.

Price and Where It Sits in the Open-Weights Field#

On OpenRouter, Hy4 preview is served by Tencent Cloud at $0.834 per million input tokens, $2.501 per million output, and $0.042 for cache reads (verified August 31, 2026). Context tops out at 1M tokens with up to 64K completion tokens, and tool calling plus structured outputs are supported through the OpenAI-compatible surface there. OpenRouter's activity tab already shows coding agents - Claude Code and Command Code among the top apps routing traffic to it - which is early evidence that agent harnesses adopt it fast when the API is this easy.

Where that lands against the rest of the field (prices per 1M tokens, verified August 31, 2026):

Price readout: Hy4 preview is not a budget model. It costs about 2.5x DeepSeek V4 Pro's off-peak output rate and sits in GLM-5.2 list-price territory, in exchange for a stronger posted agentic-terminal score and the biggest context of the group. If your workload is high-volume bounded inner loops, the DeepSeek economics post still shows Flash-class models winning that lane on price by an order of magnitude. Hy4's case is the top of the open-weights tier: hardest open-model work with a 1M context, at a quarter to a twentieth of frontier closed-model prices.

What Self-Hosting Actually Takes#

The weights are Apache 2.0, so self-hosting is legal and cheap-in-money once you have the iron - but "the iron" is the point. The BF16 weights are ~1.56TB; the FP8 release roughly halves that, and even FP8 is an 8-GPU class deployment on current high-memory accelerators. This is not a workstation model. Tencent's own serving path is vLLM or SGLang with tensor-parallel 8, the FLASHMLA_SPARSE attention backend, and a speculative-decoding configuration on top of the built-in MTP layer:

docker run --gpus all \
  -p 8000:8000 \
  --ipc=host \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  vllm/vllm-openai:hy4-preview tencent/Hy4-preview-FP8 \
    --tensor-parallel-size 8 \
    --speculative-config '{"num_speculative_tokens":3,"method":"mtp"}' \
    --attention-backend FLASHMLA_SPARSE \
    --tool-call-parser hy_v4 \
    --reasoning-parser hy_v4 \
    --enable-auto-tool-choice \
    --port 8000 \
    --served-model-name hy4-preview

At this scale the break-even math is exactly what our self-hosting economics post works through: fixed hardware cost beats a per-token bill only once utilization is high and steady, and an under-utilized 8-GPU node loses to the API almost every time. Use the API for evaluation and occasional volume; buy the GPUs only when the workload is constant and the data-egress or privacy case is real.

The serving ecosystem matters as much as the GPU count: the official vLLM prebuilt image (with the hy_v4 tool-call and reasoning parsers) and an SGLang prebuilt image both exist, which is what makes "drop it into your existing OpenAI-compatible toolchain" true on day one. Finetuning is supported through a pipeline in the repo, with the AngelSlim toolkit for further quantization.

Decision Guide#

Use Hy4 preview when:

• You want the strongest posted open-weights agentic-terminal numbers with a 1M context window.
• Your workload is text-only, long-context, and hard: repo-scale refactors, multi-file synthesis, long-horizon planning.
• You can tolerate (or even want) heavy reasoning on every hard task, and latency is not the binding constraint.
• The 1.56TB weight footprint is fine because you can self-host on an 8-GPU node or buy the API.

Use something else when:

• Speed and cost of bounded inner loops matter. DeepSeek V4 Flash at $0.22/$0.66 off-peak is the floor this tier's economics are built on - see our budget model comparison for the lane-by-lane math. Hy4 preview's price is 3x Flash on input and 4x on output.
• You need vision. Hy4 preview is text-only; if screenshots or diagrams are part of the task, this model is out by construction.
• Your hardware budget is a single workstation. A 27B dense or a 284B-class MoE quantization fits there; 770B-class weights do not. The laptop-sized end of the market is served by models like the Qwen 3.8-27B instead.
• You need a proven, preview-free production model. Hy4 is explicitly early - Tencent says both pre-training and post-training have "real headroom," and the over-verification habit is a latency tax on long agent loops. For production traffic today, the open-weights showdown field has models with months of production mileage.

When to Skip the Switch#

Even if the benchmarks look right, hold off on restructuring your stack around Hy4 preview if:

• Your eval set is not ready. Switching the top of your routing ladder without a golden set on your own repos is how "slightly ahead on 203 tasks" becomes "mysteriously worse on my codebase." Run the side-by-side first.
• Your workload is cache-heavy and repeat-identical. Hy4 preview's cache-read rate ($0.042) is real but does not approach DeepSeek's near-free cache hits; a workload whose input bill is 90% cached pays a different model's economics.
• The preview label scares your ops team. Weights can change, serving behavior can change, and the FP8 release you deploy today may not be the version Tencent tunes next month. That is normal for the open-weights release cadence - just do not build a frozen contract on top of it.

FAQ#

What is Tencent Hy4 preview?#

Hy4 preview is a text-only Mixture-of-Experts model released under Apache 2.0 on August 28, 2026: 770B total parameters with 49B active per token, a 1M-token context window, 256 routed experts with top-8 activation, DeepSeek-style gated sparse attention, and a native MTP layer for speculative decoding. Weights ship in BF16 and FP8 on Hugging Face, ModelScope, GitCode, and CNB, and it is served on OpenRouter by Tencent Cloud.

How much does Hy4 preview cost per token?#

$0.834 per million input tokens, $2.501 per million output, and $0.042 per million for cache reads, verified August 31, 2026 on the OpenRouter page. It supports up to 64K completion tokens and is served by one provider (Tencent Cloud) as of the verification date.

Can I run Hy4 preview locally?#

Technically yes - the weights are Apache 2.0, and the official vLLM/SGLang paths target tensor-parallel 8 GPU nodes with the FP8 release. Realistically no for a workstation: the BF16 weights are about 1.56TB, and even FP8 needs an 8-GPU class box. Self-hosting only wins above high, steady utilization, per the standard open-weights break-even math.

Is Hy4 preview good for coding?#

The posted numbers are strong for a text-only model: Terminal-Bench 2.1 at 85.4, SWE-bench Multilingual resolved at 82.9, SWE-bench Pro at 65.7, and DeepSWE at 64.3 on community evaluations listed on the model card, with coding agents already the top traffic sources on OpenRouter. Treat every figure as vendor or community-run on a preview model, and validate on your own repos before routing production work.

How do I control Hy4 preview's reasoning?#

Exactly two modes: high (default) and no_think, set via extra_body={"chat_template_kwargs": {"reasoning_effort": "no_think"}} on the OpenAI-compatible API. Any other value raises an exception, so the routing contract is a clean binary: deep reasoning or direct response.

How does Hy4 preview compare with DeepSeek V4?#

Same active-parameters class (49B both for Hy4 preview and V4 Pro) and same 1M context, but different economics: DeepSeek V4 Pro is $0.66/$1.98 off-peak versus Hy4's $0.834/$2.501, DeepSeek is MIT-licensed and costs less, and Hy4's posted agentic-terminal benchmark (85.4) is the headline open-weights number of the month. Divide the two by which runs your evals better, with price as the tiebreaker - the full lane math is in our DeepSeek V4 economics post.

What context window does Hy4 preview support?#

1M tokens of context with up to 64K completion tokens (per the OpenRouter model page, verified August 31, 2026). At this size the 1M window is the differentiator against every 256K-class open model, and it matches the context ceiling DeepSeek V4 and Qwen3.8 Max set for the category.

Continue Reading#

• DeepSeek V4 Economics: The Cost-Quality Frontier for Agentic Coding - the cost math Hy4 preview enters at, re-verified August 31
• GLM-5.2 vs DeepSeek V4 vs Qwen3: The Open-Weights Coding Showdown - the wider field Hy4 preview joins
• Self-Hosting Open-Weights Models: Break-Even Math - when 1.56TB of weights earns its own GPUs
• Frontier Model API Pricing, August 2026 - the verified rate card for every tier Hy4 preview competes in
• Tencent Hy3: A 295B Open MoE That Punches Above Its Weight - the previous generation this model doubles

Sources#

All sources fetched and verified August 31, 2026:

• Tencent Hy4-preview model card on Hugging Face - architecture table, benchmark appendix, quickstart, vLLM/SGLang recipes, license, known limitations
• OpenRouter: Tencent Hy4 preview - $0.834/$2.501 pricing, $0.042 cache read, 64K completion cap, provider status, top apps
• Simon Willison: Introducing Hy4 Preview - August 29 look at the chat template and reasoning behavior
• Gated DeepSeek Sparse Attention and IndexCache - the attention papers cited by the model card
• DeepSeek API Pricing - off-peak/peak rates used in the field table, verified August 31
• House-verified reference prices: DeepSeek V4, GPT-5.6 family, Sonnet 5, and Gemini rows cross-checked against the frontier pricing tracker (verified August 27 and re-confirmed in this run)