Hy4 flash
tencent/hy4-flash
Back to model catalog

Tencent · Model guide

Hy4 flash

Compare
tencent/hy4-flashReleased Nov 16, 2026

The lighter, faster member of Tencent's Hy4 family, built for high-volume everyday office work. It was self-distilled from Hy4 and then trained with reinforcement learning to finish tasks with much shorter output; Tencent reports decoding above 150 tokens per second. It takes text and images over a one-million-token context with up to 64K output tokens, and reasons by default with a direct-answer mode available.

Modalities
TextImageText
Context
1M
Max output
66K
Price / 1M tokens
$ 0.080in$ 0.32out
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 recommends Hy4 flash for most everyday tasks, such as deep search, document processing, slide generation, simple front-end pages and computer use, and Hy4 for complex work such as long-horizon software engineering.source

Best for
  • Documents and slides · Tencent tuned and evaluated it on everyday WorkBuddy tasks: building and revising slide decks, editing Word and Markdown documents alongside the user, and repairing spreadsheets.(vendor claim)source

  • High-volume everyday tasks · TokenHub describes it as a lighter, more efficient everyday workhorse for high-frequency, large-scale use, and Tencent reports end-to-end task output about half as long after reinforcement learning.(vendor claim)source

  • Images and computer use · Natively multimodal: Tencent reports 75.8 on MMMU Pro and 68.1 on OSWorld 2.1 (partial credit).source

  • Long-context work · The context window is one million tokens, of which up to 960K can be input.source

Neighbouring models
  • tencent/hy4stronger —Hy4 is the full-size model of the same family; Tencent recommends it over Hy4 flash for complex work such as long-horizon software engineering.(vendor claim)source
  • tencent/hy3predecessor —Hy3 is the previous generation: a text-only model with a 256K context. Hy4 flash adds image input and a one-million-token context.
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$ 0.080
Output$ 0.32
Cache read$ 0.0032
Cache write$ 0.080
Estimate a request
$ 0.0014
Estimated cost per request
10.0K × $ 0.080
+ 2.0K × $ 0.32
≈ $ 1.44 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 + Image → Text
Streaming
Supported
Tool calling
Supported
Structured output (JSON)
Supported
Image input
Supported
Thinking mode
Supported· can be turned off
Reasoning effort
none · high
Available endpoints
/v1/chat/completions · /v1/responses · /v1/messages

Published benchmarks

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

Agentic coding
  • Terminal-Bench 4.0
    pass_rate·Tencent Hy · 2026-11-16effort highClaude Codesetup not fully disclosed

    Up to 500 turns and a 13-hour timeout per trial, capped at 16 CPUs and 32 GB RAM; the three GPU-dependent tasks additionally get one H100. The same table repeats Hy4 at 44.4, matching the Hy4 release page.

    19.2%
Agentic work
  • AutomationBench v1.0.6
    score·Tencent Hy · 2026-11-16effort highsetup not fully disclosed

    The 600 public tasks released by Zapier, scored by the official method.

    52.2%
Coding
  • DeepSWE 1.1
    resolved_rate·Tencent Hy · 2026-11-16effort highmini-swe-agentsetup not fully disclosed

    Each task runs in a sandbox with 8 CPU cores and 16 GB of memory. Tencent reports every model at its highest available reasoning setting, which for Hy4 flash is high.

    64.6%
  • ProgramBench
    score·Tencent Hy · 2026-11-16effort highmini-swe-agentsetup not fully disclosed

    Almost Solved score. This release's footnote names mini-SWE-agent as the harness, while the Hy4 release page ran ProgramBench with Claude Code; the same table repeats Hy4's 30.0 unchanged, so the harness is recorded as each page reports it.

    16%
Computer use
  • OSWorld 2.1 (partial credit)
    score·Tencent Hy · 2026-11-16effort highsetup not fully disclosed

    Tencent labels it OSWorld 2.1 (Full, partial score); the same table repeats Hy4 at 78.9, matching the Hy4 release page.

    68.1%
Multimodal reasoning
  • MMMU Pro
    accuracy·Tencent Hy · 2026-11-16effort highsetup not fully disclosed

    Tencent does not say whether tools were available, so this sits under the unqualified MMMU Pro rather than the no-tools or with-tools variants.

    75.8%
  • ZeroBench main (with tools)
    pass_at_5·Tencent Hy · 2026-11-16effort high10 runssetup not fully disclosed

    Network-isolated Jupyter environment that returns image outputs as its tool interface, up to 80 tool-call rounds per run; pass@5 estimated from 10 independent runs.

    45.8%
Reasoning
  • Humanity's Last Exam Full (with tools)
    accuracy·Tencent Hy · 2026-11-16effort highsetup not fully disclosed

    Tencent labels it HLE-Full (with tools). The tool set is not disclosed, so it is left unstated.

    55.2%

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.

See all integration guides →

Frequently asked questions

What is Hy4 flash?
Hy4 flash is available on CrossModel as tencent/hy4-flash. It has a 1M-token context window and can return up to 66K tokens per request.
How much does Hy4 flash cost?
Currently $ 0.080 per 1M input tokens and $ 0.32 per 1M output tokens. Cached input is billed at $ 0.0032 per 1M.
Does Hy4 flash support tool calling and structured output?
Tool calling: Supported · Structured output (JSON): Supported · Image input: 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.