Teams deploying their own agent model
Run a controllable OpenAI-compatible endpoint for coding, tools, multi-step workflows, and long-context tasks on dedicated infrastructure.
Independent tool overview
Hy-3 at a glance
Tencent Hy3 is a 295-billion-parameter mixture-of-experts reasoning and agent model with 21 billion active parameters, a 256K context window, open weights under Apache 2.0, and official vLLM and SGLang deployment recipes.
Visit the official Hy-3 site ↗
- Developer
- Tencent Hy Team
- Architecture
- Mixture of experts with 295B total and 21B active parameters
- Experts
- 192 total, top 8 activated per token
- Context length
- 256K tokens
- Published checkpoints
- Hy3 BF16 and Hy3-FP8
- License
- Apache License 2.0
- Recommended serving
- vLLM or SGLang across 8 large-memory GPUs
- Last reviewed
- August 29, 2026
Overview
What Hy-3 is
Hy3 is Tencent's production-oriented text model for reasoning, coding, tool use, long conversations, and agent workflows. It follows the earlier Hy3 Preview and keeps the same broad architecture while adding more post-training, reinforcement learning, and product feedback.
The model is a large mixture of experts: it has 295B total parameters, activates 21B parameters per token, adds a 3.8B-parameter multi-token-prediction layer, and routes each token through eight of 192 experts. That can reduce compute per generated token compared with a dense 295B model, but the full expert weights still make deployment a substantial infrastructure project.
Tencent publishes a 256K context length, BF16 base weights, a separate FP8 checkpoint, tool-call and reasoning parsers, OpenAI-compatible serving examples, fine-tuning code, and reinforcement-learning guidance. Official production recipes target vLLM and SGLang with tensor parallelism across eight GPUs.
Tencent's reliability claims are promising but primarily self-reported. It says internal hallucination rates fell from 12.5% to 5.4%, commonsense error rates from 25.4% to 12.7%, and multi-turn issue rates from 17.4% to 7.9%. Those numbers describe Tencent's test sets, not a guarantee for a buyer's domain, tools, language mix, or agent scaffold.
Hy3 is best evaluated as a deployable model component, not a finished assistant. Teams still need inference infrastructure, prompt and tool design, identity and permissions, retrieval, moderation, observability, red-team testing, fallbacks, and human approval for consequential actions.
Use cases
Who Hy-3 is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Large-scale coding and productivity agents
Evaluate a model Tencent specifically post-trained for coding, office work, financial modeling, frontend work, game development, and tool use.
Organizations needing open weights
Inspect, host, fine-tune, quantize, and govern the model under the published Apache 2.0 license.
Inference teams optimizing MoE throughput
Use official vLLM, SGLang, MTP speculative decoding, FP8, and AngelSlim guidance as a starting point for deployment experiments.
Capabilities
Core Hy-3 features
295B mixture-of-experts architecture
Routes each token through eight of 192 experts, activating 21B parameters while retaining a much larger total capacity.
256K context window
Accepts long prompts and multi-turn histories, subject to memory, serving, latency, and real-task quality constraints.
Selectable reasoning effort
The official chat template exposes direct-response, low, and high reasoning-effort modes for different cost and task profiles.
Tool-call support
Official vLLM and SGLang recipes include Hy3-specific tool-call and reasoning parsers plus automatic tool selection.
OpenAI-compatible endpoint
Self-hosted examples expose a chat-completions interface that can reduce integration work for compatible clients.
BF16 and FP8 checkpoints
Tencent publishes the original instruct weights and a lower-precision FP8 variant for teams balancing memory, throughput, and quality.
Fine-tuning and RL pipeline
The repository includes supervised fine-tuning guidance and GRPO post-training support using verl, Megatron-LM, and vLLM rollout.
Compression tooling
Tencent points users to AngelSlim for quantization, low-bit compression, and speculative-sampling workflows.
Process
How the Hy-3 workflow works
Step 1
Define the deployment target
Specify languages, latency, concurrency, context length, tool count, quality thresholds, data residency, and cost per successful task.
Step 2
Choose a checkpoint and serving stack
Compare BF16 and FP8 on the actual hardware, then follow a pinned vLLM or SGLang recipe rather than an unversioned snippet.
Step 3
Benchmark the full agent
Test the chosen scaffold, prompts, tools, retrieval, and retry behavior; Tencent notes that even scaffold changes can move SWE-Bench accuracy.
Step 4
Constrain tools and outputs
Validate schemas, apply least-privilege credentials, sandbox code, cap time and spend, log calls, and require confirmation for consequential writes.
Step 5
Evaluate domain reliability
Measure hallucination, refusal, long-context retrieval, multilingual behavior, prompt injection, data leakage, and multi-turn constraint retention on private test cases.
Step 6
Load-test economics
Measure tokens per second, time to first token, batch efficiency, GPU utilization, failures, and cost per completed task under realistic concurrency.
Step 7
Roll out with fallbacks
Start with low-risk read-only tasks, monitor regressions by model and server version, and maintain human escalation plus a tested rollback path.
Cost
Hy-3 pricing and free plan
Tencent does not charge a license fee for the published Hy3 or Hy3-FP8 weights; both are released under Apache 2.0. The real price is infrastructure and operations. Tencent recommends eight H20-3e or other large-memory GPUs for serving the 295B model, while managed inference providers set their own token rates. Benchmark total cost on the selected precision, context length, concurrency, and agent workload.
Hy3 BF16 weights
No model license fee
The primary instruct checkpoint for teams prioritizing the published full-precision serving path.
- Apache 2.0 license
- 295B total parameters
- BF16 is the documented supported precision
- Requires substantial storage, memory, and serving capacity
- Infrastructure, staffing, and network costs are separate
Hy3-FP8 weights
No model license fee
A lower-precision official checkpoint intended to reduce the deployment burden.
- Apache 2.0 license is shown for the current model release
- Lower memory footprint than BF16
- Benchmark quality and throughput on target hardware
- Still a very large 295B-total-parameter model
- Infrastructure and operational costs remain material
Self-hosted production
Variable infrastructure cost
Operate vLLM or SGLang on owned, leased, or cloud GPU capacity.
- Tencent recommends 8 H20-3e or other large-memory GPUs
- Add storage, networking, orchestration, monitoring, and engineering
- Cost changes with precision, batching, context, and utilization
- Idle capacity can dominate low-volume economics
- No vendor SLA is included with the weights
Managed inference
Provider-specific
Use a third-party hosted endpoint instead of operating the model stack.
- Rates and available checkpoints vary by provider
- Check input, output, cache, and long-context pricing
- Review retention, training, residency, and subprocessor terms
- Confirm tool calling and OpenAI-client compatibility
- Benchmark provider latency and reliability separately
Pricing checked . Check current pricing at the source ↗
Assessment
Hy-3 strengths and limitations
Where it stands out
- Activates 21B parameters per token while retaining the capacity of a 295B-total-parameter MoE model.
- Publishes a long 256K context window and dedicated post-training for reasoning, agents, coding, and multi-turn reliability.
- Includes official BF16 and FP8 weights under a permissive Apache 2.0 license.
- Provides concrete vLLM and SGLang recipes with tool, reasoning, and multi-token-prediction settings.
- Exposes an OpenAI-compatible serving interface that can fit existing client patterns.
- Ships supporting material for fine-tuning, reinforcement learning, quantization, and compression.
- Tencent reports meaningful reliability improvements on internal product-oriented evaluations, providing hypotheses teams can reproduce on their own tests.
What to consider
- The 21B active-parameter figure does not make Hy3 a 21B-size deployment; all 295B expert weights must still be stored and served.
- Tencent recommends eight large-memory GPUs, putting self-hosting beyond a typical workstation and many small teams.
- A 256K advertised context window does not guarantee accurate retrieval, instruction retention, or economical inference across the full window.
- The headline benchmark and reliability results are predominantly reported by Tencent and may not transfer to a different domain, language, scaffold, or hardware stack.
- Tencent reports up to a four-percentage-point SWE-Bench accuracy range across agent scaffolds, underscoring that the surrounding system materially affects results.
- FP8 and third-party quantizations can change quality, compatibility, and throughput; every checkpoint needs separate evaluation.
- OpenAI-compatible means a similar API shape, not complete behavioral or feature equivalence with OpenAI-hosted models.
- Tool calling remains probabilistic and can produce malformed arguments, repeated actions, unsafe plans, or actions based on fabricated premises.
- Open weights do not supply moderation, identity, permissions, audit logs, privacy controls, uptime guarantees, or incident response.
- Model output can contain insecure code, false facts, bias, copyrighted material, or sensitive-data leakage and still requires application-level safeguards.
Compare
Hy-3 alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Consumer
GLM-5.1
Consider GLM-5.1 when comparing another open flagship explicitly designed for long-horizon agentic coding.
Explore GLM-5.1 →Coding
Qwen3-Coder-Next
Consider Qwen3-Coder-Next when a smaller open-weight model for agentic coding is more practical than a 295B-total-parameter deployment.
Explore Qwen3-Coder-Next →Business Operations
DeepSeek
Consider DeepSeek when you want a broader family of open models and hosted access options to compare against self-hosting Hy3.
Explore DeepSeek →Questions
Hy-3 FAQs
What is Tencent Hy3?
Hy3 is a text-generation, reasoning, coding, and agent model from Tencent's Hy Team. It uses a 295B-parameter mixture-of-experts architecture with 21B parameters active per token and a 256K context window.
Is Hy3 open source?
Tencent publishes the model weights and repository under the Apache License 2.0. Teams should still review the exact repository and checkpoint license files and account for third-party dependencies.
Can Hy3 run on a laptop?
Not in the official production configuration. Tencent recommends eight H20-3e or other large-memory GPUs. Community quantizations may broaden experimentation, but they are separate artifacts with different performance and support.
Why does Hy3 have 295B parameters but only 21B active?
It is a mixture-of-experts model. A router selects eight of 192 experts for each token, so only part of the network computes each token, while the complete expert set still needs to be available in memory or storage.
Does Hy3 support tool calling?
Yes. Tencent's vLLM and SGLang recipes include Hy3-specific tool-call parsers and automatic tool selection. Applications still need schema validation, least-privilege credentials, logging, and approval controls.
How much does Hy3 cost?
There is no model license fee for the Apache-licensed official weights. Cost comes from GPUs, storage, networking, operations, and engineering or from a managed inference provider's token rates.
What is the difference between Hy3 and Hy3-FP8?
Hy3 is the primary BF16 checkpoint, while Hy3-FP8 stores model values at lower precision to reduce memory and potentially improve serving efficiency. Benchmark both on your tasks before choosing.
Are Tencent's benchmark gains guaranteed in production?
No. They are useful published evidence, but production quality depends on the prompt, tools, agent scaffold, retrieval, languages, domain, serving configuration, and evaluation set.
Bottom line
Our Hy-3 verdict
Hy3 is a compelling open-weight candidate for organizations that want a capable agent and coding model with a much lower active-parameter count than its total capacity. Tencent has made deployment unusually concrete with official parsers, vLLM and SGLang recipes, FP8 weights, and training support. It is still enterprise-scale infrastructure: the right evaluation compares full-agent task success and total serving cost against smaller open models and managed APIs, not just headline benchmark scores.
Visit Hy-3 website ↗Similar Consumer Tools
View all Consumer AI tools →
Higgsfield Explainer
Higgsfield Explainer - Turn any topic into a faceless explainer video
Willow Frontier Mini
Willow Frontier Mini - Willow's free, unlimited AI dictation model
Claude Sonnet 5
Claude Sonnet 5 - Anthropic's mid-sized, cost-effective model with strong agentic and coding capabilities
Grok 4.5
Grok 4.5 - SpaceXAI and Cursor's strongest Grok model with powerful performance and efficiency at low costs
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.