Tencent Hy3: 295B Open-Source MoE Model for Agentic Coding — Apache 2.0, Free API, 256K Context
Tencent Hunyuan released Hy3 on July 6, 2026 — 295B MoE with 21B active params, 256K context, Apache 2.0 weights, and a two-week free API on OpenRouter. Specs, benchmarks, deployment, and how it compares to GLM-5.2.
On July 6, 2026, Tencent Hunyuan posted a launch thread that crossed 219K views within hours:
"🚀 Hy3 is here. 295B MoE. Best in its size class. Rivals trillion-scale flagships. Reliable and affordable for most agentic use cases. Apache 2.0. Friendly for commercial use. FREE API for 2 weeks."
Hy3 lands squarely in explainx.ai's China AI playbook thesis: ship capable open weights, undercut closed APIs on cost, and let enterprises route agent workloads through commodity inference.
Single-consumer-GPU local runs at full weight fidelity
Why Hy3 matters in July 2026
Three structural shifts make this release bigger than another benchmark slide:
Agent reliability is the product — Hy3's post-training focused on tool-call stability, output-format constraints, and scaffolding variance (SWE-bench accuracy within 4% across CodeBuddy, Cline, and KiloCode setups per Tencent).
Open weights + free API window — matches the free-models playbook: let developers habituate on zero marginal cost, then compete on hosted inference price.
Apache 2.0 at 295B scale — enterprises that cannot touch US export-controlled APIs get a permissive license alternative for non-regulated coding agents.
Nous Research framed the positioning clearly: "focused on cost-effective agentic use, and particularly strong on coding, tool-calling reliability, reasoning, and 256K long-context tasks."
Architecture — efficient MoE, not a dense giant
Hy3 is a sparse MoE, not a 295B dense transformer you pay for on every token.
Component
Detail
Experts
192 total, 8 active per forward pass
Attention
64 heads, GQA with 8 KV heads (dim 128)
Hidden size
4096
MTP layer
3.8B parameters for speculative decoding (vLLM/SGLang EAGLE-style)
Precision
BF16
Vocabulary
120,832 tokens
explainx.ai read: total parameter count is a marketing headline; 21B active is what shapes latency and per-token cost. That is why Tencent claims Hy3 "rivals trillion-scale flagships" — sparse activation lets a ~300B model punch like a much larger dense stack on agent benchmarks, without activating every weight.
For MoE context at even larger scale, compare NVIDIA Nemotron 3 Ultra (550B) — another open-weight agent-first design from June 2026.
What improved since Hy3 Preview (April 2026)
Tencent shipped Hy3 Preview in late April 2026, then ran a product feedback loop before the July 6 GA release.
Post-training scale-up
Feedback from 50+ internal Tencent products
Higher-quality SFT data and expanded RL training
Blind human eval: 270 experts, real work tasks, Hy3 scored 2.67/4 vs GLM-5.1 at 2.51/4
Production fixes (not just benchmark chasing)
Issue
Before → After (internal evals)
Hallucination rate
12.5% → 5.4%
Commonsense errors
25.4% → 12.7%
Multi-turn intent drift
17.4% issue rate → 7.9%
Tencent's stated principle for knowledge tasks: "answer when grounded, state when evidence is missing, do not conflate sources or fabricate data."
That matters for agents — a model that hallucinates file paths or API shapes breaks CI faster than one that scores lower on trivia QA.
Benchmarks — public leaderboards and reported agent scores
Hugging Face evaluation results (official)
Benchmark
Hy3 score
SWE-bench Verified
78%
SWE-bench Pro
57.9%
GPQA Diamond
90.4%
HLE
53.2%
Apex Agents
25.6%
SkillsBench v1.1
55.3%
WildClawBench Overall
53.6%
Reported in launch coverage (verify on your scaffold)
Grok's X summary and early reposts cited SWE-Bench Multilingual at 75.8% and Terminal Bench at 71.7%, with comparisons to GLM-5.2 and GPT-5.5-class models. Those numbers did not appear on the Hugging Face card at publish time — treat them as directional launch claims, not substitutes for your own eval harness.
Practical rule: if you run agents in Kilo, Cline, or Claude Code-style loops, benchmark your repo, your tools, your failure modes — not a single SWE-bench row.
Agentic strengths — where Hy3 is aiming
Tencent highlights five productivity lanes:
Coding — SWE-bench leadership tier for open weights
Frontend design — strongest blind-eval gains vs GLM-5.1
Game development — multi-file coherence
CI/CD and data/storage — operational tasks in the 270-expert eval
Tool-call reliability is the recurring theme. Hy3 ships dedicated parsers for vLLM (--tool-call-parser hy_v3) and SGLang (hunyuan), plus reasoning_effort modes (no_think, low, high) for chain-of-thought control.
Kilo Code's launch demo — one-shot "Create a 3D HTML visualizer of Earth spinning in space" across Opus, Fable, and Hy3 — is anecdotal, but it matches the frontend strength Tencent claims in blind evals.
For agent developers already comparing Claude Code vs Codex vs GLM-5.2 plans, the OpenRouter free window is the lowest-friction A/B test: same prompt, same tools, swap model route.
Self-hosting — hardware reality check
Tencent's deployment guide is explicit:
"Hy3 has 295B parameters in total. To serve it on 8 GPUs, we recommend using H20-3e or other GPUs with larger memory capacity."
SGLang ships an equivalent recipe with EAGLE speculative decoding.
The "Goldilocks for 96–256 GB VRAM" reply — nuance
One viral reply called Hy3 "just the right size for the 96~256 VRAM crowd." That is misleading for full-weight serving. At 299B parameters in BF16, you need datacenter multi-GPU clusters, not a single RTX 5090 or Mac Studio.
Where the reply has a point: 21B active parameters means each inference step is closer to a large dense model than a 300B dense monster — so per-token cost on hosted APIs can feel "Goldilocks" even when weight storage is not.
Hy3 vs GLM-5.2 — the comparison everyone is making
Neither replaces the other automatically. GLM-5.2 has deeper explainx.ai coverage as China's response to the Fable export ban. Hy3 is Tencent's bid for default agent backbone inside Chinese big-tech product stacks — then exported globally via OpenRouter.
If you need a security-benchmark lens, see Zhipu vs Claude Mythos on bug finding. Hy3's launch emphasized shipping velocity and tool stability, not CVE discovery leaderboards.
Quickstart — OpenAI-compatible API after vLLM boot
python
from openai import OpenAI
client = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="EMPTY")
response = client.chat.completions.create(
model="hy3",
messages=[
{"role": "user", "content": "Refactor this FastAPI route to async."},
],
temperature=0.9,
top_p=1.0,
extra_body={"chat_template_kwargs": {"reasoning_effort": "high"}},
)
print(response.choices[0].message.content)
Tencent recommends temperature=0.9, top_p=1.0. Use reasoning_effort: "high" for math, coding, and multi-step agents; "no_think" for direct responses.
Enterprise and compliance notes
Apache 2.0 removes license friction for commercial agents — but not data-sovereignty or export-control questions.
Regulated workloads (PII, HIPAA, defense) still need jurisdiction review before routing to Chinese-origin models — same framing as Asian AI alternatives post-Mythos ban.
Model-agnostic gateways (OpenRouter, LiteLLM) let you tier routes: US closed APIs for sensitive data, Hy3/GLM/DeepSeek for cost-sensitive coding sandboxes.
Tencent also ships AngelSlim quantization and a finetuning guide for teams that need domain adapters without full pretraining.
Who should switch, who should wait
Try Hy3 now if:
You run agent scaffolds (Kilo, Cline, custom MCP loops) and want a free two-week bake-off
You need 256K context for repo-scale or document-heavy agents
You want Apache 2.0 weights for on-prem experimentation (with 8-GPU budget)
Wait if:
You only have consumer GPUs — use API hosts instead of self-hosting
Your eval is security-critical and you have not red-teamed the model on your stack
You are standardized on GLM Coding Plan economics and off-peak multipliers already work for you
Specs, free-API windows, and benchmark scores reflect Tencent's July 6, 2026 release and Hugging Face leaderboard entries as of July 7, 2026. Free API promotions expire — confirm current pricing on OpenRouter before production routing. Self-hosting requirements may change as quantization recipes mature.
