{"slug": "tencent-hunyuan-hi-4-preview-specs-benchmarks-and-pricing", "title": "Tencent Hunyuan Hi-4 Preview: Specs, Benchmarks, and Pricing", "summary": "Tencent released Hunyuan Hi-4 Preview, a 770B-parameter mixture-of-experts model with 49B active parameters and a 1M-token context window, under the Apache 2.0 license. The model scores 92.3 on GPQA Diamond, 85.4 on Terminal-Bench, and 82.9 on SWE-Bench Multilingual, and Tencent's API pricing ($0.834 per million input tokens, $2.001 per million output tokens, $0.042 for cache hits) makes it the cheapest flagship open-weight model on the market. Tencent claims the model helped optimize its own training pipeline, boosting throughput by about 31.8%.", "body_md": "# Tencent Hunyuan Hi-4 Preview: Specs, Benchmarks, and Pricing\n\nTencent's Hunyuan Hi-4 Preview is a 770B MoE model with a 1M context window and the cheapest API pricing among open-weight flagships.\n\n## What is Hunyuan Hi-4 Preview?\n\nHunyuan Hi-4 Preview is Tencent’s newest open-weight large language model, released under the Apache 2.0 license with weights posted on Hugging Face in both BF16 and FP8 formats. It’s a mixture-of-experts (MoE) model with 770 billion total parameters but only 49 billion active per token, using 256 routed experts plus one shared expert, with eight experts activated per forward pass. It also ships with a gated sparse attention mechanism that lets it handle a context window beyond 1 million tokens without inference costs spiraling out of control. Unlike a general chat model, Hi-4 Preview is built specifically for agentic work: coding, multi-step planning, tool use, and office tasks like spreadsheets and slide decks.\n\n## TL;DR\n\n- **Hi-4 Preview** is a 770B-parameter MoE model from Tencent with 49B active parameters, released free and open under Apache 2.0.\n- Its **gated sparse attention** architecture supports a context window over 1 million tokens while keeping inference costs manageable.\n- On public benchmarks it scores **92.3 on GPQA Diamond, 85.4 on Terminal-Bench, and 82.9 on SWE-Bench Multilingual** , putting it near the top of the open-weight pack.\n- Tencent’s own **API pricing** ($0.834 per million input tokens, $2.001 per million output tokens, $0.042 for cache hits) makes it the cheapest flagship open-weight model on the market, though DeepSeek V4 Flash is cheaper for pure budget use.\n- The model is too large to run at home: **BF16 weights run around 1.8TB and the FP8 build needs roughly 900GB** , meaning multiple high-end GPUs (think 16x B200-class cards) as a practical minimum.\n- Tencent claims **Hi-4 Preview helped optimize its own training pipeline** , including kernel work, boosting end-to-end throughput by about 31.8% against their baseline.\n- In real-world testing through third-party agent apps, the model handled coding, 3D rendering, and long-document audit tasks competitively against other leading open models like Kimi K3 and GLM 5.3.\n\n## Remy doesn't write the code. It manages the agents who do.\n\nRemy runs the project. The specialists do the work. You work with the PM, not the implementers.\n\n## How does Hi-4 Preview’s architecture work?\n\nThe core idea behind Hi-4 Preview is sparsity at scale. Instead of activating all 770 billion parameters for every token, the model routes each token through a small subset: one shared expert plus a selection from 256 routed experts, with eight active at any given time. That keeps the active parameter count at 49 billion, which is a fraction of the total, and is what makes a model this large usable for inference at reasonable cost.\n\nThe gated sparse attention mechanism is the other key piece. Standard attention gets expensive fast as context grows, since cost scales with the square of sequence length. A gated sparse approach selectively attends to the most relevant parts of a long context rather than computing full attention over everything, which is how Tencent gets past the 1 million token mark without inference costs becoming impractical.\n\nTencent has also said the model played a role in refining its own training infrastructure, including kernel-level optimizations, which they credit with a roughly 31.8% throughput improvement over their prior baseline. That’s a notable claim: a model contributing to the efficiency of the very pipeline that trains it, though it’s Tencent’s own internal report rather than an independently verified result.\n\n## How does Hi-4 Preview perform on benchmarks?\n\nOn public evaluations, Hi-4 Preview lands near the top of the open-weight category. It scores 92.3 on GPQA Diamond (a graduate-level science reasoning benchmark), 85.4 on Terminal-Bench (agentic terminal/command-line tasks), and 82.9 on SWE-Bench Multilingual (real-world software engineering issues across languages). It also leads the open-weight field on some newer research-oriented benchmarks, including Horizon Math and BioMystery Bench.\n\nTencent additionally ran an internal blind evaluation involving 163 expert reviewers judging 203 real engineering tasks. Hi-4 Preview scored 2.99 out of 4, compared to 2.94 for Kimi K3 and 2.92 for GLM 5.3. Because this is Tencent’s own evaluation rather than a third-party benchmark, it’s worth treating as a data point rather than a definitive ranking. Combined with the public benchmark numbers, the overall picture is a model that trades wins with other leading open models like Kimi K3 and GLM 5.3, while still sitting behind closed frontier models such as Opus and GPT-5-class systems on the hardest tasks. That gap is expected: closed labs generally hold an edge on the most demanding reasoning and coding benchmarks, and an open-weight model closing most of that gap while remaining freely licensed is still a meaningful result.\n\n## What does Hi-4 Preview cost to use?\n\nPricing is arguably the standout part of this release. Through Tencent Cloud’s Token Hub, Hi-4 Preview costs $0.834 per million input tokens and $2.001 per million output tokens, with cache hits priced at just $0.042 per million tokens. That undercuts other open-weight flagships like Qwen 3.8 Max and Kimi K3. The one model that beats it on raw cost is DeepSeek V4 Flash, which remains the cheapest option if price is the only consideration.\n\n## Other agents ship a demo. Remy ships an app.\n\nReal backend. Real database. Real auth. Real plumbing. Remy has it all.\n\nThe model is also available through OpenRouter, which means it can be plugged into coding agents and tools like Cline or other OpenCode-style environments without going through Tencent’s own infrastructure directly. Because the license is Apache 2.0, any provider can host and serve the weights, which tends to push prices down further as competition among hosts increases. That licensing detail matters more than it might seem: it’s a large part of why this model can be priced so aggressively compared to more restrictively licensed alternatives.\n\n## Can you run Hi-4 Preview locally?\n\nNot practically, no. The BF16 weights come in at roughly 1.8TB, and even the quantized FP8 build needs around 900GB of memory. That puts local deployment out of reach for anything short of a multi-GPU server setup, realistically something in the range of 16 high-end data center GPUs (B200-class or similar) as a rough minimum. For nearly everyone, that means the API or a hosted app is the only realistic way to use the model, rather than downloading and running the weights directly.\n\n## Is Hunyuan Hi-4 Preview worth using?\n\nFor agentic and coding-heavy workloads, the benchmark scores and demonstrated task performance suggest yes, particularly if cost is a factor. It scores competitively against other top open-weight models on coding and reasoning benchmarks, and its API pricing is aggressive enough to matter for anyone running high-volume agent workflows or coding assistants. Independent testing has shown it handling complex, multi-constraint coding tasks (like physics-heavy game builds) and long-horizon, multi-document tasks (like auditing dozens of expense claims against shifting policy documents) with a level of consistency that’s harder to find in open models.\n\nThe tradeoffs are real, though. It still trails top closed models like Opus and GPT-5-class systems on the hardest reasoning and coding challenges. It’s also not something individuals or small teams can self-host given the hardware requirements. For most builders, that means treating it as an API-based option rather than a model to download, which is a fairly common state of affairs for frontier-scale open-weight releases at this size.\n\n## Frequently Asked Questions\n\n### How many parameters does Hunyuan Hi-4 Preview have?\n\nIt has 770 billion total parameters in a mixture-of-experts architecture, with only 49 billion active per token thanks to routing across 256 experts plus one shared expert.\n\n### What is Hi-4 Preview’s context window?\n\nOver 1 million tokens, enabled by a gated sparse attention mechanism designed to keep inference costs manageable at that scale.\n\n### How much does the Hi-4 Preview API cost?\n\nThrough Tencent Cloud’s Token Hub, it’s $0.834 per million input tokens, $2.001 per million output tokens, and $0.042 per million tokens for cache hits, making it the cheapest flagship open-weight model, though DeepSeek V4 Flash is cheaper overall.\n\n### Can I run Hunyuan Hi-4 Preview on my own hardware?\n\nNot realistically. The BF16 weights are about 1.8TB and the FP8 build needs roughly 900GB, requiring a multi-GPU server setup rather than consumer or single-server hardware.\n\n### How does Hi-4 Preview compare to Kimi K3 and GLM 5.3?\n\nIt’s broadly competitive, trading wins across public benchmarks and scoring slightly higher in Tencent’s internal blind expert evaluation (2.99 vs 2.94 and 2.92 out of 4), though that evaluation is Tencent’s own and not independently verified.", "url": "https://wpnews.pro/news/tencent-hunyuan-hi-4-preview-specs-benchmarks-and-pricing", "canonical_source": "https://www.mindstudio.ai/blog/tencent-hunyuan-hi4-preview-moe/", "published_at": "2026-09-08 00:00:00+00:00", "updated_at": "2026-09-08 20:53:28.594179+00:00", "lang": "en", "topics": ["large-language-models", "artificial-intelligence", "ai-research", "ai-products"], "entities": ["Tencent", "Hunyuan Hi-4 Preview", "Hugging Face", "DeepSeek V4 Flash", "Kimi K3", "GLM 5.3"], "alternates": {"html": "https://wpnews.pro/news/tencent-hunyuan-hi-4-preview-specs-benchmarks-and-pricing", "markdown": "https://wpnews.pro/news/tencent-hunyuan-hi-4-preview-specs-benchmarks-and-pricing.md", "text": "https://wpnews.pro/news/tencent-hunyuan-hi-4-preview-specs-benchmarks-and-pricing.txt", "jsonld": "https://wpnews.pro/news/tencent-hunyuan-hi-4-preview-specs-benchmarks-and-pricing.jsonld"}}