Laguna S 2.1 vs Tencent Hy3: Cheap Open Coder vs Tencent's Efficient MoE

Poolside's Laguna S 2.1 (July 21, 2026) is an open-weight 118B/8B-active coding MoE at $0.10 / $0.20 per 1M tokens. Tencent's Hy3 (released July 6, 2026) is Tencent Hunyuan's production-grade Mixture-of-Experts model — 295B total parameters with only 21B active per pass — priced at $0.18 / $0.59 per 1M with a 262K-token context. Both are efficiency-first MoE models — this guide compares them on coding, general capability, price, openness, and fit.

Every figure is labeled by source. Laguna's coding scores are Poolside-reported; Hy3's specs and price are from its OrcaRouter model page. Benchmarks and prices move — verify before committing.

TL;DR verdict. Tencent Hy3 is the broader generalist: a 295B/21B-active MoE with a 262K context, strong reasoning, long-context and agentic capability, and a reported 84.2 on BrowseComp, at $0.18 / $0.59. Laguna S 2.1 is the coding specialist: Poolside reports it leading open disclosed-size models on SWE-Bench Multilingual (78.5%) at $0.10/$0.20, with open weights you can self-host — but its numbers are vendor-run and coding-only. Choose Hy3 for broad, long-context work; choose Laguna S 2.1 for cheap, self-hostable, coding-focused work.

Key takeaways

• Tencent Hy3 is the broader generalist. 295B/21B-active MoE, 262K context, reasoning + agentic strength.

• Laguna is a coding specialist and slightly cheaper. $0.10/$0.20 vs Hy3's $0.18/$0.59, open weights, self-hostable.

• Laguna's scores are vendor-run and coding-only; Hy3's specs come from its live model page.

• Both are efficiency-first MoEs; Laguna activates ~8B, Hy3 ~21B per pass.

• Split: broad long-context generalist (Hy3) vs best-per-dollar open coder (Laguna).

The specs and price, side by side

• Vendor / license — Laguna S 2.1: Poolside, open-weight; Tencent Hy3: Tencent, hosted API (OpenAI-compatible).

• Price per 1M (in / out) — Laguna S 2.1: $0.10 / $0.20; Tencent Hy3: $0.18 / $0.59.

• Architecture — Laguna S 2.1: 118B total / ~8B active MoE; Tencent Hy3: 295B total / 21B active MoE (192 experts, top-8 routing).

• Context — Laguna S 2.1: 1M-class; Tencent Hy3: 262K tokens (text in, text out).

• Coding (vendor) — Laguna S 2.1: SWE-Bench Multilingual 78.5% (leads open disclosed-size per Poolside); Tencent Hy3: strong general coding (BrowseComp 84.2 reported).

Benchmark deep-dive: broad MoE generalist vs weight-class coder

Tencent Hy3's strength is breadth and efficiency at scale: a 295B-parameter MoE that activates only 21B per pass, built on the Hy3-preview line with expanded RL training for reasoning, long-context, and agentic tasks — Tencent positions it as reaching results comparable to flagship models several times its parameter size, with a reported 84.2 on the BrowseComp browsing/long-context benchmark. For broad work over long documents, it's a strong, cost-effective choice.

Laguna S 2.1's case is coding efficiency and openness: Poolside reports 78.5% on SWE-Bench Multilingual (leading open disclosed-size models) and strong DeepSWE/Terminal-Bench results from a ~8B-active model at a fraction of a cent per call. The caveats: vendor-run numbers, no independent Intelligence Index, coding-only focus, and Poolside concedes frontier models still lead several benchmarks. So Hy3 is the broader generalist with far more active parameters and a longer real-world context; Laguna is the leaner, coding-focused, open-weight weight-class leader. The question is whether your workload is broad and long-context (Hy3) or coding-heavy and self-hosted (Laguna).

Thinking mode and cost

Both are efficiency-first MoEs and both offer configurable reasoning effort — Laguna via thinking/no-thinking, Hy3 via its reasoning-effort setting. On price, Laguna is a bit cheaper ($0.10/$0.20 vs $0.18/$0.59): a 15,000-in / 3,000-out call is about $0.0021 on Laguna versus about $0.0045 on Hy3 — both tiny. The bigger differences are structural: Laguna is open-weight and self-hostable (removing per-token cost entirely for coding), while Hy3 offers far more active parameters and a 262K context for broad, long-document work. For a coding-dominant, cost- or data-sensitive workload Laguna wins; for broad long-context tasks Hy3's capacity earns its modest premium.

How to access each

Laguna S 2.1 is open-weight — hosted at $0.10/$0.20 or self-hosted from Hugging Face (poolside/Laguna-S-2.1). Tencent Hy3 is available via an OpenAI-compatible API (model ID tencent/hy3), and through OrcaRouter. Route both behind one OpenAI-compatible endpoint — Laguna for coding-heavy traffic, Hy3 for broad and long-context work.

Three real-world scenarios

1. High-volume coding automation

Laguna S 2.1 is purpose-built and cheapest here; its SWE-Bench Multilingual and DeepSWE numbers target exactly this workload, and self-hosting removes per-token cost.

2. Long-context and agentic workloads

Tencent Hy3's 262K context, 21B active parameters, and reported BrowseComp strength make it the safer single model when tasks need big inputs, browsing, and tool use.

3. Self-hosted, data-resident deployments

Laguna's open weights let you keep proprietary code in-house — Hy3 is a hosted API, so for data-resident coding Laguna is the enabler.

When not to use each

Don't pick Laguna S 2.1 for broad, long-context, or browsing-heavy general tasks — Hy3's larger active model and 262K context cover more. Don't pick Hy3 if pure coding-per-dollar or self-hosting is the goal — Laguna is cheaper and open-weight. And treat Laguna's vendor coding numbers as provisional.

Decision checklist

• Choose Laguna S 2.1 if: coding is the workload, cost is critical, or you need a self-hostable coder that beats its weight class.

• Choose Tencent Hy3 if: you want a broad, long-context, efficient MoE generalist at a low hosted price.

• Run both if: route coding to Laguna and broad/long-context work to Hy3, behind one endpoint.

FAQ

Is Laguna S 2.1 better than Tencent Hy3?

For coding efficiency and openness, Poolside reports Laguna leading its open weight class, and it's self-hostable. For broad and long-context work, Hy3 is a larger MoE (295B/21B active) with a 262K context. Coding-specialist vs generalist.

Which is cheaper?

Laguna S 2.1 ($0.10/$0.20) is a bit cheaper than Hy3 ($0.18/$0.59), and Laguna can be self-hosted to remove per-token cost.

Can I self-host either?

Laguna S 2.1 yes (open weights on Hugging Face); Tencent Hy3 is a hosted API.

What is Tencent Hy3?

Hy3 is Tencent Hunyuan's production-grade MoE — 295B total parameters, 21B active per pass (192 experts, top-8 routing), 262K context, text in / text out, released July 6, 2026.

Does Laguna have an independent score?

Not yet — its coding scores are Poolside-reported. Hy3's specs and price come from its live OrcaRouter model page.

Which should a coding startup pick?

For cost-sensitive or self-hosted coding features, Laguna S 2.1; add Hy3 if you also need broad, long-context capability.

Can I run both together?

Yes — route coding to Laguna and general/long-context work to Hy3 through one OpenAI-compatible endpoint.

Bottom line

Tencent Hy3 is an efficient, broad MoE generalist — 295B total / 21B active, a 262K context, and a reported 84.2 on BrowseComp — at a low hosted price of $0.18/$0.59, the safer single model when your workload needs breadth and long context. Laguna S 2.1 is the coding specialist: Poolside reports it leading open disclosed-size models on SWE-Bench Multilingual (78.5%) at $0.10/$0.20, with open weights you can self-host — though its numbers are vendor-run and coding-only. For cheap, self-hostable coding, Laguna; for broad long-context work, Tencent Hy3. Route both — and every rival flagship — through one OpenAI-compatible endpoint at OrcaRouter.