cd /news/artificial-intelligence/tencent-hy3-295b-moe-hits-swe-bench-… · home topics artificial-intelligence article
[ARTICLE · art-50698] src=byteiota.com ↗ pub= topic=artificial-intelligence verified=true sentiment=↑ positive

Tencent Hy3: 295B MoE Hits SWE-Bench 78 — Free API Ends July 21

Tencent released Hy3, a 295B Mixture-of-Experts model with Apache 2.0 weights, achieving a 78.0 SWE-bench Verified score and leading open-weight models on tool and search benchmarks. A free API on OpenRouter is available until July 21, after which pricing will be $0.20 per million input tokens and $0.80 per million output tokens, offering a cost-effective alternative to proprietary models like Claude Sonnet 5.

read3 min views1 publishedJul 8, 2026
Tencent Hy3: 295B MoE Hits SWE-Bench 78 — Free API Ends July 21
Image: Byteiota (auto-discovered)

Tencent dropped Hy3 on July 6 — a 295B Mixture-of-Experts model with Apache 2.0 weights, a 78.0 SWE-bench Verified score, and a free API on OpenRouter that expires July 21. If you’re choosing a coding agent backend or evaluating open-weight models for agentic workflows, you have 13 days to test it at zero cost.

What the SWE-Bench Score Actually Means #

Hy3 hits 78.0 on SWE-bench Verified. That number deserves context before you act on it.

GLM-5.2 scores 84.2 on the same benchmark. Claude Sonnet 5 leads the full BenchLM leaderboard at around 87 across agentic, coding, and reasoning tasks. So Hy3 is not the top coder — and Tencent doesn’t pretend it is.

Where Hy3 leads is tool and search workloads: 84.2 on BrowseComp, 91.0 on DeepSearchQA, and 79.1 on MCP-Atlas tool orchestration — best in the open-weight field on all three. If your agent crawls the web, calls MCP servers, or does long-context retrieval, Hy3’s benchmark profile fits that workload better than its SWE-bench number suggests.

Benchmark stability is also worth noting. Tencent tested Hy3 across KiloCode, Cline, and CodeBuddy scaffoldings and reported less than 4% accuracy variance. That kind of consistency across frameworks matters in production — a model that performs differently depending on your agent harness is a reliability problem.

The Free API Closes July 21 #

OpenRouter’s tencent/hy3:free

endpoint is live now. Cost: $0 per million tokens, input or output. On July 21, that flips to $0.20 per million input tokens and $0.80 per million output tokens.

After the free window closes, Hy3 is still substantially cheaper than alternatives — Claude Sonnet 5 runs around $3/$15 per million tokens, making Hy3 roughly 15x cheaper at the post-July-21 rate. But you won’t get another free evaluation window. The time to test it is now.

The API is OpenAI-compatible. Swap your base URL, change the model name, done:

from openai import OpenAI
import os

client = OpenAI(
    api_key=os.getenv("OPENROUTER_API_KEY"),
    base_url="https://openrouter.ai/api/v1"
)

response = client.chat.completions.create(
    model="tencent/hy3:free",
    messages=[{"role": "user", "content": "Review this function for bugs: ..."}],
    extra_body={"reasoning_effort": "high"}
)
print(response.choices[0].message.content)

The reasoning_effort

parameter controls chain-of-thought depth: "no_think"

for fast direct responses, "low"

for light reasoning, "high"

for complex tasks like multi-step code review or debugging. Set it wrong and you’re either wasting tokens or missing accuracy.

Apache 2.0 — No Regional Restrictions #

The April 2026 preview version excluded the EU, UK, and South Korea from its license. The July 6 official release fixes that. Hy3 is now Apache 2.0 with zero geographic exclusions — self-host it, fine-tune it, ship it in commercial products, anywhere.

That matters more than it sounds. GLM-5.2 outperforms Hy3 on SWE-bench, but carries usage constraints that matter to enterprise deployments. Hy3 with a clean Apache 2.0 license at $0.20/$0.80 per million tokens post-July-21 is a different risk profile for teams that can’t accept third-party data egress or unclear regional terms.

Self-Hosting: What You Actually Need #

The weights are on Hugging Face at tencent/Hy3 (BF16) and tencent/Hy3-FP8. Self-hosting a 295B MoE model isn’t casual — Tencent recommends 8 H20 GPUs or equivalent high-memory cards.

For vLLM, the relevant flags are --tool-call-parser hy_v3

and --reasoning-parser hy_v3

with --enable-auto-tool-choice

. SGLang uses --tool-call-parser hunyuan

and --reasoning-parser hunyuan

. Both support MTP speculative decoding for faster inference.

Most teams will start with the OpenRouter API and self-host only if the use case justifies the GPU spend. That’s a reasonable progression.

What to Do Before July 21 #

If you run agent workflows involving web search, MCP tools, or long-context retrieval: Hy3 is the right thing to test this week. Its SWE-bench score is real but not the whole picture — the tool and search benchmarks suggest a model tuned for exactly this class of workload.

If you need the best pure coding performance in the open-weight tier: GLM-5.2 still wins. Hy3 is not that model.

Start at OpenRouter’s free endpoint. The weights are at github.com/Tencent-Hunyuan/Hy3. The free window closes July 21.

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @tencent 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/tencent-hy3-295b-moe…] indexed:0 read:3min 2026-07-08 ·