What Is Hunyuan 3? Tencent's 295B Open-Source Agentic Model Explained (2026)

Most open-source models launch with a caveat: you need a small server room to run them.

Tencent's Hunyuan 3 — officially called Hy3 — broke that pattern on July 14, 2026, when the team released 1-bit and 4-bit GGUF quantizations that run on a single GPU via llama.cpp. The full-precision model is 598 GB. The FP8 version fits in under 300 GB on a single Nvidia H20-3e node. For context, GLM-5.2 needs eight H200 GPUs and roughly 744 GB just to load.

If you have been evaluating open models for agentic workflows — tool-chaining, multi-step search, long-context orchestration — and kept hitting the "great benchmarks, impossible hardware" wall, Hy3 is the first model that seriously challenges that trade-off.

That single-GPU story is what pushed Hy3 to #1 on the OpenRouter LLM leaderboard on July 13, 2026 — and sent Tencent's stock up 4.82% on the day of the full release. But hardware accessibility is only half the picture. Hy3 also claims the top open-source scores in agentic search, tool orchestration, and long-context retrieval — while honestly trailing GLM-5.2 in every coding benchmark.

We cross-referenced the official Tencent announcements, the Hugging Face model card (tencent/Hy3), five independent reports, and X community reactions to separate confirmed facts from marketing claims. This guide gives you the architecture, the real benchmark picture (strengths and weaknesses), pricing at every access tier, a 5-minute test path, and the common pitfalls people hit when deploying Hy3.

Hunyuan 3 in 30 seconds

Hunyuan 3 is a 295-billion-parameter Mixture-of-Experts (MoE) model with only 21 billion parameters active per token — a 7% sparsity ratio that is among the most aggressive in the open-source field. It supports a 256K native context window, ships under the Apache 2.0 license with no geographic or field restrictions, and is available as full-precision weights, FP8 quantization, and GGUF formats on Hugging Face under tencent/Hy3.

The model was built with an explicit anti-hallucination training philosophy: answer only when there is evidence, and state clearly when there is none. Tencent reports that hallucination rates dropped from 12.5% to 5.4% between the April preview and the July full release.

The 30-second version tells you what Hy3 is. The architecture section below explains why those numbers translate into the single-GPU deployment story.

How 295B parameters fit on one GPU: architecture breakdown

Hy3 uses a sparse Mixture-of-Experts architecture that routes each token through 8 of 192 available experts. This is what makes it possible to pack 295B parameters of knowledge into a model that only computes with 21B at inference time.

Why the MTP layer matters for speed

The 3.8B multi-token prediction (MTP) layer is the underappreciated part of Hy3's design. Standard autoregressive models predict one token at a time. Hy3's MTP layer predicts several tokens ahead in a single forward pass, then verifies them — a technique called speculative decoding. The practical effect: faster generation without sacrificing accuracy. On Tencent's internal benchmarks, this produces a 1.8–2.4× throughput improvement compared to standard decoding.

Both vLLM and SGLang support Hy3's speculative decoding out of the box, and the GGUF builds use it via llama.cpp's MTP support.

Architecture determines what is theoretically possible. The comparison table below shows what that means in practice — especially for your hardware budget.

Hunyuan 3 vs GLM-5.2, Inkling, and Kimi K3: the real comparison

Hardware requirements tell a practical story that benchmark tables often hide.

Hy3 is the smallest model in this group by total parameters — and that is the point. Tencent traded raw scale for deployment efficiency. A developer with a single high-end GPU can run Hy3 locally. That same developer cannot run GLM-5.2, Inkling, or (likely) Kimi K3 without multi-GPU infrastructure.

For a deeper look at Inkling, see our Inkling explainer. For Kimi K3, see our Kimi K3 guide.

The hardware comparison tells you which models you can actually deploy. The next question is whether the performance justifies the deployment.

Hunyuan 3 benchmarks: where it leads, and where it honestly does not

Tencent claims Hy3 is the #1 open-source model in agentic search, tool orchestration, and long-context retrieval. The coding scores tell a different story. Here is the full picture — with the critical caveat that all numbers are self-reported.

Strengths: agentic tasks and long context

These are the scores that got Hy3 to #1 on the OpenRouter leaderboard. BrowseComp and MCP-Atlas are particularly relevant for agentic workflows — the kind where a model needs to search the web, call APIs, and chain multiple tools together to complete a task.

Weakness: coding trails GLM-5.2 across the board

The DeepSWE gap is striking — GLM-5.2 scores nearly twice as high. If your primary use case is automated code repair or complex multi-file engineering tasks, GLM-5.2 is the stronger choice today. Hy3's strength is agentic orchestration: coordinating tools, retrieving context, and reasoning over long conversations — not writing or fixing code.

Reliability improvements: April preview → July full release

The July release is not just the April preview with a new license. Tencent reports measurable reliability gains:

The MRCR improvement — from 42.9% to 75.1% — is the number that matters most for agentic use. It means the model is significantly better at using information from earlier in a long conversation, which is critical when context accumulates across dozens of tool calls.

Benchmarks tell you what the model can do in controlled conditions. The next section shows you how to verify that on your own workload in under five minutes.

Before you commit: a 5-minute Hy3 quality test

Before choosing a deployment path, run this quick test to verify Hy3 fits your use case. It takes under five minutes and costs nothing.

Step 1. Go to OpenRouter and select tencent/hy3. The free tier is available through July 21, 2026 — no credit card required.

Step 2. Send it a task that matches your actual workload. Pick one:

Step 3. Evaluate the response against your current model. If Hy3 matches or beats your current solution on the task you actually care about, proceed to a deployment option below. If it falls short on the coding test specifically, that is expected — see the benchmark section above.

How to run Hunyuan 3: four paths from zero-setup to self-hosted

Once the quality test checks out, choose the access path that matches your constraints.

Option 1: OpenRouter API (fastest start)

OpenRouter routes Hy3 at approximately $0.06–0.20 per million input tokens and $0.21–0.80 per million output tokens. A free tier is available through July 21, 2026. No local hardware required — just an API key.

Best for: quick evaluation, prototyping, low-volume production where you do not want to manage infrastructure.

Option 2: Tencent Cloud API

Tencent's own API prices input at 1 yuan (~$0.14) per million tokens. Hy3 is integrated into Tencent Cloud's model serving infrastructure and supports the same speculative decoding pipeline as the self-hosted version.

Best for: production workloads with consistent volume, especially if you are already in the Tencent Cloud ecosystem or serving Chinese-language users.

Option 3: Self-hosted full precision or FP8

Download tencent/Hy3 from Hugging Face. The full-precision checkpoint is 598 GB; the FP8 version is under 300 GB and fits on a single Nvidia H20-3e node. Serve with vLLM or SGLang for speculative decoding support.

Best for: data-sensitive workloads where tokens cannot leave your infrastructure, or when you need to fine-tune.

Option 4: GGUF on a single GPU (released July 14)

The 1-bit and 4-bit GGUF quantizations run through llama.cpp with multi-token prediction support. This is the option that went viral — a 295B-parameter model running on consumer-accessible hardware.

Best for: local experimentation, privacy-first use cases, developers who want to run the model on hardware they already own.

Tool and IDE integrations

Hy3 is already available in several developer tools: Cline, OpenCode, Cherry Studio, Nous Portal, Hermes, Kilo Code, and OpenClaw. On the consumer side, it powers Tencent's Yuanbao AI assistant and is integrated into WeChat.

Hunyuan 3 pricing: what it actually costs at each tier

For comparison, GLM-5.2 API pricing is approximately $1.40 input / $4.40 output per million tokens. Hy3 is roughly 7–20× cheaper on input and 5–20× cheaper on output through OpenRouter. Even at the upper end of OpenRouter's range, it undercuts GLM-5.2 significantly.

Common pitfalls when deploying Hunyuan 3

These are the issues people hit most often in the first week of deploying Hy3, based on community reports and X threads.

Pitfall 1: Using the April preview weights instead of the July release

Symptom: High hallucination rates, poor multi-turn coherence, unexpectedly weak long-context retrieval.

Root cause: The April preview (2026-04-23) had restrictive terms, geographic limitations, and significantly worse reliability metrics. Some early tutorials and cached Hugging Face links still point to the preview weights.

Fix: Verify your checkpoint is from the July 6 or later release by checking the Hugging Face model card date on tencent/Hy3. The July weights use Apache 2.0 — if your license file mentions geographic restrictions, you have the wrong version.

Pitfall 2: Running GGUF without MTP enabled

Symptom: The model runs but generation is 2–3× slower than expected.

Root cause: llama.cpp supports Hy3's multi-token prediction, but it is not enabled by default. Without MTP, you lose the speculative decoding speedup that is central to Hy3's performance story.

Fix: Pass the MTP flag when launching llama.cpp. Check the tencent/Hy3 Hugging Face page for the current recommended flags — they are updated as llama.cpp versions change.

Pitfall 3: Expecting coding performance at the level of agentic benchmarks

Symptom: Hy3 produces mediocre code fixes or fails on complex multi-file engineering tasks, despite being "OpenRouter #1."

Root cause: Hy3's leaderboard ranking reflects agentic and conversational performance, not coding. SWE-bench Verified is 78.0 vs GLM-5.2's 84.2, and the gap widens on harder benchmarks (DeepSWE: 28.0 vs 46.2).

Fix: This is not a deployment error — it is a model limitation. For coding-heavy workloads, use GLM-5.2 or a coding-specialized model. Use Hy3 for the orchestration layer that calls the coding model, not as the coding model itself.

Pitfall 4: Hitting context limits on long agentic chains

Symptom: Quality degrades noticeably after ~200K tokens in a long agentic conversation.

Root cause: While Hy3 supports 256K native context, real-world agentic chains accumulate tool-call outputs, error messages, and intermediate results that consume context faster than conversational text. A 50-step agentic chain can easily hit 200K+ tokens.

Fix: Implement context summarization between major agentic steps. Compress completed sub-task outputs into summaries before passing them as context for the next step. If you consistently need 500K+ token context, consider Inkling (1M) or Kimi K3 (1M+) instead.

The bigger picture: Hunyuan is a family, not just an LLM

Hy3 is the language model, but it sits inside a broader Tencent AI ecosystem that spans three modalities:

This matters because Hy3 is designed to work as the orchestration layer — the "brain" that can coordinate creative pipelines. An agentic workflow might use Hy3 to plan a video project, call HunyuanVideo to generate clips, and use HunyuanImage 3.0 for thumbnail generation, all through MCP-style tool calls. Tencent reports a 90% task completion rate for Hy3's agent capabilities on internal applications.

For teams already using Hunyuan's video or image models (including through tools like Wan 2.6), Hy3 adds a reasoning layer that speaks the same ecosystem.

When to use Hunyuan 3 — and when to pick something else

The decision depends on two dimensions: what your workload looks like, and what infrastructure you have.

Timeline: from preview to production

The shift from April to July was not cosmetic. Tencent dropped all geographic restrictions, switched to Apache 2.0, released GGUF quantizations, and shipped a model with materially better reliability — hallucination down 57%, context recall up 75%. If you evaluated the April preview and passed, the July release is a different model in practice.

Frequently asked questions

What is Hunyuan 3? Hunyuan 3 (officially Hy3) is Tencent's flagship open-source large language model — a 295B-parameter Mixture-of-Experts model with 21B active parameters, Apache 2.0 license, and single-GPU deployment support.

Is Hunyuan 3 open source? Yes. The full weights are on Hugging Face under tencent/Hy3 with an Apache 2.0 license. There are no geographic, commercial, or field-of-use restrictions.

How many parameters does Hunyuan 3 have? 295 billion total, but only 21 billion are active per token due to the Mixture-of-Experts architecture. This 7% sparsity ratio is what enables single-GPU inference.

Can I run Hunyuan 3 on a single GPU? Yes. The 1-bit and 4-bit GGUF quantizations (released July 14, 2026) run on a single GPU via llama.cpp with multi-token prediction. The FP8 version fits on a single Nvidia H20-3e node (<300 GB). Use the 4-bit GGUF or FP8 for serious work — the 1-bit version is a proof-of-concept.

How does Hunyuan 3 compare to GLM-5.2? Hy3 leads in agentic tasks (BrowseComp 84.2, MCP-Atlas 79.1) and costs 7–20× less. GLM-5.2 leads in coding (SWE-bench 84.2 vs 78.0, DeepSWE 46.2 vs 28.0) but requires 8× H200 GPUs. Choose based on whether your workload is agentic orchestration or code engineering.

What does Hunyuan 3 cost? Through OpenRouter: $0.06–0.20 input / $0.21–0.80 output per million tokens, with a free tier until July 21, 2026. Through Tencent Cloud: ~1 yuan ($0.14) per million input tokens. Self-hosted: hardware cost only.

What is Hunyuan 3 good at? Agentic search (BrowseComp), multi-tool orchestration (MCP-Atlas), long-context retrieval (AA-LCR), and deep search (DeepSearchQA). It is not the strongest choice for coding tasks.

Is Hunyuan 3 the best open-source model? It reached #1 on the OpenRouter leaderboard, which weights agentic and conversational performance. On coding benchmarks, GLM-5.2 is stronger. On context length, Inkling and Kimi K3 offer 1M tokens vs Hy3's 256K. "Best" depends entirely on your use case — see the decision table above.

How do I download Hunyuan 3? The weights are on Hugging Face at tencent/Hy3. Full precision is 598 GB. FP8 is under 300 GB. GGUF quantizations (1-bit and 4-bit) are available for llama.cpp. Make sure you download the July 6 release, not the April preview.

Core summary

Hy3 is the rare model where the headline ("295B parameters on a single GPU") actually holds up under scrutiny. It is not the biggest, not the best at coding, and not independently benchmarked yet — but it offers a combination no other open model currently matches.

• Confirmed strength: Agentic orchestration (BrowseComp 84.2, MCP-Atlas 79.1) and long-context retrieval are genuinely strong — strong enough to reach #1 on the OpenRouter leaderboard.
• Confirmed weakness: Coding performance trails GLM-5.2 by 6–18 points across every benchmark. This is not a rounding error — it is a gap wide enough to affect production decisions.
• Confirmed advantage: Single-GPU deployment via GGUF and FP8 is real, not a marketing claim. No other model at this performance tier offers it.
• Unconfirmed: All benchmarks are Tencent-reported. Independent verification from Artificial Analysis, Chatbot Arena, or academic evaluations has not been published yet.

The one thing worth doing right now: open OpenRouter, use the free tier (available through July 21, 2026), and send Hy3 a real agentic task — a multi-step search, a tool-chaining workflow, or a 50K-token context retrieval problem. Five minutes with your own data will tell you more than this entire article. We will update it as independent benchmarks arrive.