CrossModel
Hy4 Preview
tencent/hy4-preview
Back to model catalog

Tencent · Model guide

Hy4 Preview

Compare
tencent/hy4-previewReleased Aug 28, 2026

Tencent's next-generation open flagship: a 770B-parameter mixture of experts that activates 49B per token across 78 layers, with a native multi-token-prediction layer for speculative decoding. Gated DeepSeek Sparse Attention with cross-layer index reuse carries a one-million-token context. This preview is text-only and defaults to deep chain-of-thought reasoning, with a direct-answer mode available, and Tencent built its training data around the work its own software, finance, game and security teams ship.

Modalities
TextText
Context
1M
Max output
66K
Price / 1M tokens
$ 0.96in$ 2.88out
StreamingTool callingJSON outputImage inputThinking mode · can be turned off

What this model is good at

What the vendor positions it for, and which of our models to reach for instead.

VendorTencent presents Hy4 preview as an early release at the open-source frontier, built for productivity work, and says it ships with known issues so that feedback can shape the full version.source

Best for

  • Agentic coding · Tencent reports 85.4 on Terminal-Bench 2.1 and 82.9 on SWE-bench Multilingual, and co-designs the model with its own CodeBuddy coding agent.source

  • Long-context work · The context window reaches one million tokens, four times the previous generation, and Tencent scores it on a with-tools million-token benchmark.source

  • Office and analysis · Tencent highlights turning scattered context into documents, spreadsheets and presentations, with data analysis and financial modelling.(vendor claim)source

  • Scientific research · Tencent claims gains on hard research questions across AI research, molecular dynamics, condensed matter physics and pure mathematics.(vendor claim)source

Neighbouring models
  • tencent/hy3predecessor —Hy3 is the previous generation: a much smaller and cheaper model with a 256K context, and the stable release rather than a preview.
Compare them side by side →

Pricing and billing

Billed per token. Cached input is charged at the cache-read rate.

Price / 1M tokensYou pay
Input$ 0.96
Output$ 2.88
Cache read$ 0.048
Cache write$ 0.96
Estimate a request
$ 0.0154
Estimated cost per request
10.0K × $ 0.96
+ 2.0K × $ 2.88
≈ $ 15.36 per 1,000 requests

Capabilities and limits

What CrossModel guarantees across every route this model can take right now.

Context window
1.0M tokens
Max output
65.5K tokens
Input / output modalities
Text → Text
Streaming
Supported
Tool calling
Supported
Structured output (JSON)
Supported
Image input
Not supported
Thinking mode
Supported· can be turned off
Reasoning effort
none · high
Available endpoints
/v1/chat/completions · /v1/responses · /v1/messages
First-party sources:Specsgithub.com 1huggingface.co 2Capabilitiesgithub.com 1cloud.tencent.com 2

Published benchmarks

Scores the vendor reported, with the evaluation setup each one came from.

Agentic coding
  • Terminal-Bench 2.1
    pass_rate·Tencent Hy · 2026-08-28effort highClaude Codesetup not fully disclosed

    Claude Code harness, up to 500 turns and a 12-hour timeout per trial, capped at 16 CPUs and 32 GB RAM. The same table reports Hy3 at 70.8 rather than the 71.7 in this catalogue; Tencent notes the harness, judge model and anti-hacking mechanisms changed, so the two are not the same run.

    85.4%
Agentic work
  • Agents' Last Exam (CLI)
    score·Tencent Hy · 2026-08-28effort highClaude Codesetup not fully disclosed

    ALE-CLI, 105 tasks under the official protocol with the Claude Code harness, up to 12 hours per task on 8 CPU cores and 32 GB, scored by the official ALE evaluators.

    22.8%
Coding
  • SWE-Bench Multilingual
    resolved_rate·Tencent Hy · 2026-08-28effort highSWE-agentsetup not fully disclosed
    82.9%
  • SWE-Bench Pro
    resolved_rate·Tencent Hy · 2026-08-28effort highSWE-agentsetup not fully disclosed
    65.7%
Document understanding
  • OfficeQA Pro
    score·Tencent Hy · 2026-08-28effort highClaude Codesetup not fully disclosed

    Claude Code was the agent harness for every model except the GPT series, which ran on Codex CLI.

    66.2%
Knowledge work
  • GDPval-AA v2
    elo·Tencent Hy · 2026-08-28effort highsetup not fully disclosed

    Elo, marked as the official evaluation in Tencent's table rather than one of its own reruns.

    1678
Long context
  • OneMillionBench (with tools)
    score·Tencent Hy · 2026-08-28effort highsetup not fully disclosed

    Tencent does not disclose which tools the run had, so the tool set is left unstated rather than guessed.

    65.4%
Reasoning
  • Humanity's Last Exam (with tools)
    accuracy·Tencent Hy · 2026-08-28effort highsetup not fully disclosed

    Text-only subset. Tencent does not disclose the tool set, so it is left unstated; the same table reports 43.4 without tools.

    55.4%
Scientific reasoning
  • GPQA Diamond
    accuracy·Tencent Hy · 2026-08-28effort highsetup not fully disclosed
    92.3%
Tool use
  • MCP Atlas Public
    score·Tencent Hy · 2026-08-28effort highsetup not fully disclosed

    Scale's April 2026 methodology on the 500-task public set, 100 tool-call budget per task, with the judge model moved from Gemini 2.5 Pro to Gemini 3.1 Pro Preview.

    83.7%
  • Toolathlon Verified
    pass_rate·Tencent Hy · 2026-08-28effort high3 runssetup not fully disclosed

    Tencent's internal agent scaffold with some MCP tools reimplemented, 2 CPU cores and 10 GB per task and a 2-hour timeout instead of 5400 seconds; the 108-task Verified set, official evaluators, 100-step budget and 3-run Pass@1 protocol are unchanged.

    74.1%

Vendor-reported numbers, not CrossModel measurements. Scores are only comparable when the benchmark version, metric and evaluation setup match, so nothing here is averaged or ranked.

Use it in your tools

Point the base URL at CrossModel and paste this model ID — every tool below has a setup guide.

Claude Code
Anthropic's official command-line coding assistant
Cursor
AI-first code editor
Cline
Open-source AI coding agent for VS Code
Codex
OpenAI's coding agent for ChatGPT desktop and the terminal
OpenCode
Open-source terminal coding assistant
Zed
High-performance, AI-native code editor
Open WebUI
Open-source, self-hosted AI chat interface
Continue
Open-source AI coding assistant for VS Code & JetBrains
See all integration guides →

Frequently asked questions

What is Hy4 Preview?
Hy4 Preview is available on CrossModel as tencent/hy4-preview. It has a 1M-token context window and can return up to 66K tokens per request.
How much does Hy4 Preview cost?
Currently $ 0.96 per 1M input tokens and $ 2.88 per 1M output tokens. Cached input is billed at $ 0.048 per 1M.
Does Hy4 Preview support tool calling and structured output?
Tool calling: Supported · Structured output (JSON): Supported · Image input: Not supported · Streaming: Supported · Thinking mode: Supported
Which endpoint do I call?
For text models, OpenAI Chat Completions (/v1/chat/completions), OpenAI Responses (/v1/responses) and Anthropic Messages (/v1/messages) all work — pick whichever your SDK already speaks.
Can I try it without writing code?
Yes. Sign in and open the Playground on this page to chat with the model directly — no API key needed. Usage is billed from your wallet at the normal rate.

Chat first, integrate later

Chat first, then wire it in once you like the answers.