Tencent Hy4 Flash Free for 2 Weeks on OpenCode: 1M Context, Multimodal & ZDR

Tencent Hy4 Flash Free for 2 Weeks on OpenCode: 1M Context, Multimodal, and ZDR

At 15:31 UTC+8 on November 16, 2026, OpenCode announced on X the launch of Tencent Hunyuan's newly released lightweight multimodal model Hy4 Flash, listed under the identifier hy4-flash-free, with a two-week free trial window running roughly through November 30. Later at 17:19 UTC+8, the team followed up to confirm that OpenCode Go subscribers also get 14 days of free, near-unlimited usage that will not count against their Go monthly quota.

Built around a 1M context window, native image understanding, and a zero data retention policy, Hy4 Flash is well suited for everyday coding, long document analysis, and frontend screenshot debugging. Before jumping in, however, it is important to understand the access boundaries: without a paid subscription, the zero-cost Console tier works only inside OpenCode clients and returns HTTP 403 if called from external editors with a Console API key; only existing OpenCode Go subscribers can route it across external tools without consuming Go quota, and "near-unlimited usage" still carries practical rate limits. This guide breaks down the full rules, real user feedback, and client-agnostic alternatives.

💡 Key Highlights

  • 🗓️ Two-Week Free Window: Open for two weeks starting November 16, 2026, expected to run through roughly November 30, with no regional restrictions listed in the docs.
  • 🧠 224B MoE Lightweight Architecture: 224B total parameters with 12B active per token; OpenCode exposes the full 1M context window and 64K max output, with text/image input and tool calling.
  • 🛡️ Zero Data Retention Policy: OpenCode confirms the model is hosted in US and EU nodes with 0-day data retention and no training on user prompts or code.
  • 🔒 Dual Access Rules for Free Tier vs. Go: The Console free tier runs strictly inside OpenCode clients and rejects external harnesses with HTTP 403, while OpenCode Go subscribers can invoke it free across external tools.
  • ⚖️ Objective Alternatives: If you prefer not to change your coding setup for a temporary model, check Tencent WorkBuddy's parallel two-week free offer or evergreen API-key providers like NVIDIA NIM and Xiaomi MiMo.

🔥 1. Model Specs and Promotion Rules

1. What Is Hy4 Flash?

Hy4 Flash corresponds to tencent/hy4-flash, a cost-efficient companion model released by Tencent Hunyuan on November 16, 2026. It uses a Mixture-of-Experts and latent attention design with 224B total parameters, including 64B Engram parameters, activating only 12B parameters per forward pass.

With hybrid attention, a KV-cache footprint reduced to one-sixth of the standard Hy4 model, 4-bit quantization, and megakernel execution, Hy4 Flash reaches decoding speeds above 150 tokens per second. It supports text alongside JPEG, PNG, WEBP, and GIF image inputs, full tool calling, and configurable reasoning effort between none and high.

Context length is one of the highlights here. OpenCode configures hy4-flash-free with a full 1M context window, specifically 1,048,576 total tokens, up to 983K input tokens, and 65K output tokens. Community users noted that some other platforms cap the same model at 300K context, whereas OpenCode exposes the full 1M capacity.

2. Free Mechanism and Timeframe

In its X announcement, OpenCode stated that Hy4 Flash will offer near-unlimited free usage for the next two weeks, and the documentation marks it as free for a limited time. Counting from the November 16 announcement, the window runs until around November 30.

This operates as a limited-time free inference window rather than a pre-claimed token balance you can save for later. Once the two-week promo ends, normal billing or endpoint removal may take effect at any time.

3. Zero Data Retention Privacy Terms

For developers feeding proprietary repositories and logs into an AI assistant, training opt-out guarantees matter. According to OpenCode's Console models documentation and Go privacy table, Hy4 flash Free follows a strict Zero Data Retention policy. Hosted on US and EU infrastructure, prompts and code are retained for zero days and are never used for model training.

🛠️ 2. Getting Started and Access Boundaries

1. Zero-Subscription Access: Inside OpenCode Clients

If you do not subscribe to OpenCode Go and want to try hy4-flash-free at zero cost, run an official OpenCode client directly, including the CLI, desktop app, or web console. Select Hy4 flash Free in the model picker or set opencode/hy4-flash-free in your configuration to start using it without a credit card or paid plan.

2. Using Cursor, Claude Code, Codex, Pi, and ZCode

Many developers look for an API key first to plug into their existing editor. OpenCode's documentation states a strict rule for the Console free tier: free chat models are only available from within OpenCode itself, and any external harness or custom script receives HTTP 403 even when authenticating with a valid Console API key.

To call hy4-flash-free over a standard OpenAI-compatible endpoint from Cursor, Claude Code, Codex, Pi, or ZCode, you need an OpenCode Go subscription using the Go gateway https://opencode.ai/inference/go/openai/v1 and model ID opencode-go/hy4-flash-free. Per OpenCode's follow-up announcement, Go, Go Plus, and Go Teams subscribers can use it free for 14 days without deducting from their Go usage quota. For existing Go subscribers, that is a handy zero-quota backup; for everyone else, buying a $10 subscription just for a two-week flash model makes little sense when the native OpenCode client is free.

⚠️ 3. Real-World Limits and Caveats

1. Near-Unlimited Still Hits Rate Limits, and Maxed-Out Go Accounts Report Blocks

The announcement promises near-unlimited usage rather than truly uncapped throughput. Shortly after both posts went live, replies surfaced concrete friction: one user hit a usage cap in just 13 minutes on a single thread without invoking subagents, and multiple developers reported HTTP 429 errors. Because OpenCode does not publish exact RPM or daily request limits for hy4-flash-free, heavy automated loops and concurrent agent runs can still trip anti-abuse throttling.

There is also an important caveat for Go subscribers. Even though the follow-up tweet states usage will not count against Go quota and the Go docs mention free models remain accessible after hitting limits, a Go user whose monthly quota was already exhausted reported getting blocked within one second when trying to call hy4-flash-free. If your Go monthly allowance is already depleted, do not count on this promo as a guaranteed bypass.

2. Latency Expectations Near 1M Context

While a 1M context window looks generous on paper, time-to-first-token and long-range recall still degrade as prompts approach full capacity, a concern developers raised immediately in the replies. In practice, feed only relevant source folders, type definitions, and error logs instead of dumping build artifacts or dependencies.

3. Understanding Where a 12B-Active Flash Model Fits

With 12B active parameters per token, Hy4 Flash is built as a fast, economical workhorse rather than a frontier reasoning flagship. Based on Tencent's published numbers, Hy4 Flash scores 19.2 on Terminal-Bench 4.0 compared to 44.4 for the full 770B-total, 49B-active Hy4 model, and Tencent itself recommends Flash for everyday documents, slides, and light coding while reserving standard Hy4 for long coding runs. Use this promo for code completion, unit tests, log triage, screenshot debugging, and document summaries, and switch back to heavier models for deep architectural work.

4. Post-Promo Pricing Reference

OpenCode has not listed post-promo pricing for hy4-flash-free. As a reference, Tencent's native API pricing for Hy4 Flash sits at roughly $0.07 per 1M input tokens, $0.281 per 1M output tokens, and $0.003 per 1M cached input tokens, placing it firmly in the low-cost flash tier even after the two-week window closes.

🛡️ 4. Reliable Free Alternatives Without Client Lock-In

If you prefer not to switch clients for a two-week promo and want workflows tied cleanly to Cursor, Claude Code, Codex, Pi, or ZCode, consider these options.

First, if your goal is specifically to test Tencent Hy4 Flash, Tencent's desktop agent workspace WorkBuddy is also running a simultaneous two-week free promotion.

Second, if you need standard OpenAI-compatible API keys that work across any editor at zero cost, prioritize NVIDIA NIM. New developer accounts receive 1,000 free inference credits with no credit card required and a 40 RPM serverless limit across 1M-context models like DeepSeek V4.1 Flash and GLM-5.3-Flash. Full setup details and endpoint comparisons live in our Free LLM API Directory and Guide.

Finally, for production workloads that cannot rely on temporary promos, stable commercial-grade backends provide essential continuity. Xiaomi's MiMo open platform offers consistent concurrency, low latency, and standard tool compatibility. Registering via our Xiaomi MiMo partner link with referral code Z7EJ2L unlocks developer trial perks as a dependable daily driver.

📬 Subscribe to the Free AI API Newsletter

Limited-time promos and fine-print rate limits shift quickly. Visit Free AI API to subscribe to our developer newsletter for fact-checked rules, real-world caveats, and reliable zero-cost compute updates.

Frequently Asked Questions

Community Feedback

Comments are tied to your GitHub account — sign in to join the discussion.