Laguna S 2.1 vs Tencent Hy3: Two Open Models, Very Different Bets
This is the closest fight in the open-weight coding model space right now. Tencent Hy3 scores 71.7% on Terminal-Bench with 295B total parameters and 21B active per token. Laguna S 2.1 scores 70.2% on Terminal-Bench with 118B total parameters and 8B active per token.
That is a 1.5 percentage point difference in Terminal-Bench. Practically a rounding error. But Laguna achieves it with 2.5x fewer active parameters. These two models represent the two most efficient approaches to open-weight coding, and they arrive at nearly the same destination through completely different paths.
The Numbers
| Benchmark | Laguna S 2.1 | Tencent Hy3 |
|---|---|---|
| Terminal-Bench 2.1 | 70.2% | 71.7% |
| SWE-bench Pro | 59.4% | 57.9% |
| DeepSWE | 40.4% | N/A |
| SWE-bench Multilingual | 78.5% | N/A |
| Total parameters | 118B | 295B |
| Active per token | 8B | 21B |
| Context window | 1M | 256K |
| Input pricing | $0.10/M (free at 256K) | $0.50/M |
| Output pricing | $0.20/M (free at 256K) | $2.00/M |
Look at those numbers carefully. Tencent Hy3 leads by 1.5 points on Terminal-Bench. Laguna S 2.1 leads by 1.5 points on SWE-bench Pro. They are essentially tied on aggregate coding performance, but Laguna does it with dramatically less compute.
The Efficiency Story
Tencent Hy3 activates 21B parameters per token. Laguna S 2.1 activates 8B. That is a 2.6x difference in active compute per forward pass. Despite using less than half the compute, Laguna matches Hy3’s Terminal-Bench score within noise.
What does this mean practically?
- Inference speed: With 8B active params vs 21B, Laguna generates tokens faster on equivalent hardware.
- Cost: Laguna is free at 256K vs Hy3’s $0.50/$2.00. Even at 1M context, Laguna is 5x cheaper on input and 10x cheaper on output.
- Self-hosting: Laguna’s smaller active footprint means lower GPU memory requirements for inference. See our local deployment guide for specifics.
- Context window: 1M tokens vs 256K. Laguna has 4x more context, which matters for large codebases.
Where Laguna Wins: SWE-bench Pro
Laguna S 2.1 scores 59.4% on SWE-bench Pro vs Tencent Hy3’s 57.9%. This benchmark tests real GitHub issue resolution across popular open-source projects. It is arguably the most practical benchmark for day-to-day coding work.
The 1.5-point lead here suggests Laguna is slightly better at understanding existing codebases and producing correct patches. This aligns with Poolside’s training approach: RLCEF (Reinforcement Learning with Code Execution Feedback) teaches the model to write code that actually passes tests, not just code that looks correct.
For comparison, Claude Sonnet 5 hits 63.2% on SWE-bench Pro and Grok 4.5 hits 64.7%. Both Laguna and Hy3 are within striking distance of proprietary models at a fraction of the cost.
Where Hy3 Wins: Terminal-Bench (Slightly)
Hy3’s 71.7% vs Laguna’s 70.2% gives it a marginal edge on terminal-based agentic tasks. If Terminal-Bench accuracy is your single most important metric and you are willing to pay for the extra 1.5 percentage points, Hy3 delivers.
But consider: Laguna S 2.1’s thinking mode took it from 60.4% to 70.2% on Terminal-Bench. That thinking mode boost represents a 16% relative improvement. Without thinking, Hy3 wins by a wider margin. With thinking enabled, Laguna closes the gap to near-parity.
Context Window: 1M vs 256K
Laguna’s 1M token context is 4x larger than Hy3’s 256K. This matters more than people think:
- A medium-sized TypeScript project can easily have 200K+ tokens of relevant source code
- With Hy3’s 256K limit, you are already bumping against the ceiling
- With Laguna’s 1M, you can include full codebases, test suites, documentation, and conversation history
For monorepo work, large refactoring projects, or any task where you need the model to understand the full picture, Laguna’s context advantage is decisive.
Pricing Comparison
| Scenario | Laguna S 2.1 | Tencent Hy3 |
|---|---|---|
| Light use (256K context) | Free | ~$2-5/day |
| Heavy use (256K context) | Free | ~$15-30/day |
| 1M context tasks | ~$0.50-2/day | N/A (256K max) |
| Self-hosted | Hardware cost only | Higher hardware requirements |
The pricing gap is substantial. Laguna is free for most coding workflows. Hy3 costs $0.50/$2.00 per million tokens. Over a month of heavy use, that difference adds up to hundreds of dollars.
See our AI API pricing comparison for the full picture.
Training Approach Differences
Tencent Hy3 uses a more traditional scaling approach with 295B total parameters and 21B active. It benefits from Tencent’s massive data infrastructure and training resources. The result is a capable model that achieves excellent performance through parameter count.
Laguna S 2.1 uses RLCEF with 409K training environments and multi-harness rollouts. It trained in under 9 weeks on 4,096 H200 GPUs. The emphasis is on training methodology over raw scale. The model learns from code execution feedback rather than static data preferences.
The case studies illustrate this: Laguna built a browser engine from scratch (50 minutes, 181 steps), optimized its own training harness (5.2% faster, 70% less memory), and re-derived Erdos problem #397. These demonstrate the sustained reasoning that RLCEF training produces.
The Smaller Options: XS 2.1
If you want something even more efficient, Poolside offers Laguna XS 2.1 (33B total, 3B active, 262K context). It scores 33.4% on Terminal-Bench and 63.1% on SWE-bench Multilingual. While not competitive with either S 2.1 or Hy3 on raw benchmarks, it runs on consumer GPUs with 24GB VRAM. Nothing in Tencent’s lineup offers that kind of accessibility.
Check our VRAM requirements guide and best models for local use for hardware planning.
Integration and Tooling
Both models support standard inference frameworks. Laguna S 2.1 officially supports Ollama, vLLM, and SGLang. It also has the pool CLI agent for agentic coding workflows similar to Claude Code.
Both work with Aider and other third-party coding tools. The tooling ecosystem is comparable, though Laguna’s first-party pool CLI gives it a slight edge in developer experience.
Which Should You Choose?
Choose Laguna S 2.1 if:
- Cost matters (free at 256K vs $0.50/$2.00)
- You need 1M context for large codebases
- You want to self-host with lower hardware requirements
- SWE-bench Pro style tasks (issue resolution) are your primary use case
- You value the pool CLI and Poolside’s developer ecosystem
Choose Tencent Hy3 if:
- You need the absolute highest Terminal-Bench score in a reasonable price range
- Your tasks are specifically terminal-agent oriented
- You already have Hy3 integrated into your workflow
- The 1.5 point T-Bench advantage matters for your specific use case
For most developers, the 1.5 point Terminal-Bench advantage of Hy3 does not justify the 5-10x cost increase and 4x smaller context window. Laguna S 2.1 is the better default choice.
Read our best open-source coding models guide for how both models fit into the broader landscape.
FAQ
Is 1.5 percentage points on Terminal-Bench a meaningful difference?
At this level of performance (70-72%), 1.5 points is within run-to-run variance on many tasks. It is statistically meaningful in aggregate but unlikely to be noticeable on any individual task. The models are functionally equivalent in capability.
Which model is easier to self-host?
Laguna S 2.1 by a significant margin. With only 8B active params vs 21B for Hy3, and 118B total vs 295B, Laguna requires substantially less GPU memory for inference. Laguna XS 2.1 (33B/3B) runs on a single consumer GPU.
How do both compare to closed-source models?
Claude Fable 5 (88.0% T-Bench) and Muse Spark 1.1 (80% T-Bench) both outperform either model. But they cost $1.25-$2+ on input and cannot be self-hosted. The value proposition of open-weight models is cost control and deployment flexibility, not absolute peak performance.
Can I use both models together?
Yes. Some teams use a cheaper model (Laguna S 2.1, free) for initial drafts and iteration, then validate with a more expensive model for final review. This hybrid approach gets you most of the benefit of both at minimal cost.
Which has better multilingual support?
Laguna S 2.1 scores 78.5% on SWE-bench Multilingual, which is one of its strongest benchmarks. Tencent Hy3 does not publish this specific number. For polyglot codebases, Laguna is a proven choice.