Hy4 Preview replaces closed frontier APIs for teams that want an open-weight model with a 1,048,576-token context, not the single highest reasoning score. It suits cost-sensitive agentic coding and long-document work, not teams needing multimodal input or a published safety evaluation.
Hy4 Preview is Tencent's open-weight Hunyuan large language model, released August 2026 as a Mixture-of-Experts design with 49B parameters active per token out of a much larger total. It targets developers who want a self-hostable, permissively licensed model for agentic coding, not a closed API-only frontier model.
Provider: Tencent · Family: Hunyuan
Context window: 1,048,576 tokens · Max output: 64,000
Input modalities: text, tool-calls · Output: text, tool-calls
About Hy4 Preview
Hy4 Preview is Tencent's fourth-generation Hunyuan large language model, released as an open-weight preview on August 28, 2026, about seven weeks after Hy3's production launch. It is a Mixture-of-Experts model with 770 billion total parameters and 49 billion active per token, built on 78 layers: one dense feed-forward layer plus 77 MoE layers, each with 256 routed experts and 1 shared expert, top-8 routed per token. A native multi-token-prediction layer (10B total, 0.7B active) handles speculative decoding via Gated DeepSeek Sparse Attention with IndexCache. Tencent frames the release as an early version with headroom left in pretraining and post-training, not a finished flagship. On published benchmarks, Hy4 Preview scores 92.3 on GPQA Diamond, 82.9 on SWE-bench Multilingual (with tools), 65.7 on SWE-bench Pro (Public), and 85.4 on Terminal Bench. In Tencent's own blind evaluation, 163 expert graders scored it 2.99 out of 4.00 across 203 engineering tasks, narrowly ahead of Kimi K3 (2.94) and GLM-5.3 (2.92). No SWE-bench Verified, AIME 2025, MMLU-Pro, or ARC-AGI-2 score is published, and trackers like DataLearnerAI and eesel place it behind Claude Opus 5 and GPT-5.6 on most public evals, trading wins mainly with other Chinese open models. Context runs to 1,048,576 tokens, with a max output of 64,000 tokens. No published long-context recall evaluation was found for this release. Hy4 Preview is text-only, taking no image, audio, or video input. It supports tool and function calling, accepting tools and tool_choice parameters plus JSON-schema output via response_format on its OpenAI-compatible endpoint. Tencent's docs note this covers request shape only, not behavioral identity with the source APIs. The model is co-designed with CodeBuddy and WorkBuddy for coding agents and office-productivity work rather than conversational chat. Tencent Cloud's TokenHub API prices Hy4 Preview per token, well under $1 per million input tokens but several times pricier than the outgoing Hy3 model, with exact rates detailed below. No batch or provisioned tier is published. WorkBuddy and CodeBuddy gave free access at launch, and Hy3 stays free through Sept 30, 2026 as a migration window. TokenHub serves Singapore and Silicon Valley over an OpenAI-compatible Bearer endpoint. OpenRouter lists it but forwards requests to Tencent Cloud; no listing was found on Bedrock, Vertex, Azure, Together AI, or Fireworks. Weights ship under Apache 2.0 on Hugging Face, GitHub, ModelScope, and GitCode, with a same-day FP8 build. Self-hosting BF16 needs roughly 1,848GB of VRAM (about 924GB at MXFP8), targeting 8-way tensor-parallel on B200/B300-class GPUs via vLLM or SGLang. No system card, alignment writeup, or red-team list is published for this release. Trackers note Tencent has skipped the goal-misgeneralization, sandbagging, and refusal-consistency evals standard for frontier launches. The model card is candid about two quirks: it over-reasons on complex tasks, and tends to over-verify completed work. Hy4 Preview suits teams wanting a self-hostable, Apache-licensed model for agentic coding at large context, who can absorb an 8-16 GPU cluster or TokenHub's rate. It fits less well for teams needing vision/audio input, a published safety evaluation, or the top closed-benchmark score, where Opus 5 and GPT-5.6 lead. Training data and cutoff date are not disclosed. Tencent's FAQ does not document API retention; OpenRouter separately offers zero-data-retention routing. No SOC2, GDPR, or HIPAA claim is documented for this model. Hy4 Preview is the fourth Hunyuan release, after Hy1 (2023) and Hy2 (2024). Hy3 shipped as a 295B/21B-active preview on April 23, 2026, reaching production July 6, 2026 with SWE-bench Pro rising 46.0 to 57.9. Hy4 Preview roughly doubles Hy3's active parameters and quadruples its context; no production, non-preview Hy4 has been announced as of this research.
Pricing
Tencent Cloud TokenHub charges $0.834 per 1M input tokens, $2.501 per 1M output, and $0.042 per 1M cache-hit input tokens (Singapore pricing, Aug 28 2026); the domestic CNY listing is about ¥6 input and ¥18 output per 1M tokens. Tencent has said this runs 5 to 6 times above the outgoing Hy3 model, which keeps its own free run through Sept 30, 2026 as a migration window for existing users. No batch or provisioned-throughput tier is published.
Key Features
- Flagship-Class Context Window: One of the largest context budgets of any open-weight model released in August 2026, with a 64,000-token maximum completion length.
- Apache 2.0 Open Weights: Full-precision and FP8 checkpoints mirrored across four major open-model hosts within hours of release, all under the Apache 2.0 license.
- Native Tool and Structured-Output Calling: Accepts function-calling parameters and returns schema-constrained JSON, built for multi-step coding and productivity agents rather than open-ended chat.
- Speculative Decoding via Native MTP Layer: A dedicated 10B-parameter multi-token-prediction layer accelerates generation without a separate draft model, built directly into the checkpoint.
- Steep Cache-Hit Discount: Reusing cached input tokens costs a small fraction of the fresh-input rate, a meaningful saving for multi-turn agent sessions that resend the same context.
Pros
- Posts an 85.4 on Terminal Bench, its strongest published agentic benchmark to date.
- Ships as fully open weights on day one, letting teams fine-tune or self-host without a vendor API dependency.
- Cache-hit pricing runs a small fraction of the fresh-input rate, well suited to repeat-context agent loops that reuse the same system prompt.
Cons
- None of the current-generation reasoning benchmarks (SWE-bench Verified, AIME 2025, MMLU-Pro, ARC-AGI-2) have a published score, and third-party trackers rank it below Claude Opus 5 and GPT-5.6 across most other evals.
- Handles text only, with no image, sound, or video capability in either direction, and no red-team or safety evaluation has been published for this release.
- Self-hosting needs close to a terabyte of VRAM across a double-digit GPU cluster, and per-token pricing runs well above the outgoing Hy3 model's rate.
Benchmarks
- deep swe: 64.3
- gpqa diamond: 92.3
- terminal bench: 85.4
- humanitys last exam: 55.4
- toolathlon verified: 74.1
- swe bench pro public: 65.7
- swe bench multilingual: 82.9
- tencent internal blind eval: 2.99
- lmarena code webdev rank open models: 3
Frequently Asked Questions
What does Hy4 Preview cost per 1M tokens?
Running Hy4 Preview through Tencent's hosted API costs $0.834 per million input tokens and $2.501 per million output, with cache hits billed at just $0.042 per million. If you self-host the open weights instead, the model itself is free and you only pay for compute.
How does Hy4 Preview compare to GLM-5.3 and Kimi K3 on benchmarks?
On Tencent's internal 203-task expert review, Hy4 Preview edged both rivals, landing at 2.99 versus Kimi K3's 2.94 and GLM-5.3's 2.92. Independently, it scores 82.9 on SWE-bench Multilingual, but no SWE-bench Verified or AIME figure exists yet for a fully apples-to-apples comparison.
Is Hy4 Preview open source or proprietary?
Hy4 Preview is open-weight and Apache-licensed, with checkpoints mirrored across Hugging Face, GitHub, ModelScope, and GitCode from day one. Commercial use, fine-tuning, and redistribution are all permitted.
Does Hy4 Preview train on data sent through its API?
Tencent has not published a data-retention or training-on-inputs policy for the TokenHub API. OpenRouter separately offers a zero-retention routing option when you go through its platform. Fully local, self-hosted inference never sends data anywhere.
Who should use Hy4 Preview, and who should avoid it?
It fits teams that want a self-hostable, open-weight model for agentic coding at very large context and can run or rent a large GPU cluster. Skip it if you need vision or audio input, a documented safety evaluation, or simply the highest score on closed reasoning benchmarks.