Hy3 logo

Hy3

Tencent · Released Jul 2026

Conditional

Apache 2.0 open-weight MoE agentic model (295B total, 21B active) with strong tool-calling reliability and anti-hallucination behavior, but coding trails GLM-5.2 and self-hosting demands 8+ GPUs. The free OpenRouter tier that ran through Jul 21 is gone; OpenRouter now lists six providers at $0.13-$0.20 input and $0.53-$0.80 output per 1M tokens. Best for self-hosted coding agents and long-context document work where permissive licensing matters more than absolute benchmark scores.

Is it right for you?

Good for

  • Agentic coding with stable tool-calling across scaffoldings — SWE-Bench accuracy within 4% across CodeBuddy, Cline, and KiloCode
  • Self-hosted deployments under Apache 2.0 — no API dependency, permissive commercial use, FP8 variant available for lower memory footprint
  • Long-context (256K) document processing with anti-hallucination training — hallucination rate fell from 12.5% to 5.4% in internal evals
  • Beating GLM-5.1 on expert blind tests (2.67/4 vs 2.51/4) — strongest in frontend, CI/CD, data/storage workflows

Not good for

  • Coding benchmarks (SWE-Bench Verified 78.0) trail GLM-5.2 (84.2) and Claude Sonnet 5 — not the choice for pure coding excellence
  • Self-hosting cost — 295B total parameters needs 8 GPUs (H20-3e minimum); the free OpenRouter tier that ran through Jul 21, 2026 is no longer listed

How it performs by task

Code generation

Very Good

SWE-Bench Verified 78.0 — strong for an open model but trails GLM-5.2 by 6.2 points

Agentic workflows

Very Good

Tool-call stability across scaffoldings, scaffold-agnostic variance within 4%

Reasoning

Very Good

GPQA Diamond 90.4, USAMO 72.0, IMOAnswerBench 90.0 — competitive with larger models

Long-context tasks

Very Good

256K context, MRCR long-dialogue 75.1%, anti-hallucination improvements

Pricing

Input

$0.20 / 1M

Output

$0.80 / 1M

Context

256K tokens; OpenRouter providers $0.13-$0.20 in / $0.53-$0.80 out; free tier ended

View full pricing

Benchmarks

BenchmarkScoreSource
SWE-Bench Verified78.0 Source
SWE-Bench Multilingual75.8 Source
Terminal-Bench 2.171.7 Source
GPQA Diamond90.4 Source
USAMO 202672.0 Source
IMOAnswerBench90.0 Source
HLE (with tools)53.2 Source

Verdict history

Sep 30, 2026
conditional -> conditional (summary corrected: free tier expired, provider price range): Rating unchanged. The standing summary described a free OpenRouter tier through Jul 21 as current. OpenRouter's page, read 2026-09-30, lists no free variant and six providers at $0.13-$0.20 input and $0.53-$0.80 output per 1M tokens.

Verification log

  • Pricing· No changes

    Automated agent

  • Pricing· Updated

    Automated agent

    Free OpenRouter tier (through Jul 21) is gone; OpenRouter lists six providers at $0.13-$0.20 input and $0.53-$0.80 output per 1M. Stored $0.20/$0.80 is the top of that range and is kept. pricing_context, strengths, audiences and sources refreshed; summary corrected via verdict_set (conditional held). Context length not stated on the page; 256K is the stored value.

  • Pricing· Needs attention

    Automated agent

    The OpenRouter page returned no readable price or context rows, so the stored $0.20/$0.80 is kept and nothing was written. The Hugging Face tencent/Hy3 page (read 2026-09-26) confirms Apache 2.0 and 256K context but gives no hosted price. Flag: the summary and strengths still describe a 'free OpenRouter tier through Jul 21' as upcoming. That date is past, and the text needs a rewrite once the price is re-sourced.

  • Pricing· No changes

    Automated agent

  • Pricing· No changes

    Automated agent

  • Pricing· No changes

    Automated agent

  • Pricing· No changes

    Automated agent

  • Profile· No changes

    Imported at launch

  • Pricing· No changes

    Imported at launch

How we evaluate