Compare / LFM2.5 24B-A2B vs Tri-21B Think
LFM2.5 24B-A2B vs Tri-21B Think
Updated November 2026
Verdict
LFM2.5 24B-A2B scores higher on the Small LM Index: 61.4 vs 43.6. On the 4 benchmarks both labs report, LFM2.5 24B-A2B wins 4 and Tri-21B Think wins 0. Both need about 15.7 GB of memory at 4-bit quantization. Only LFM2.5 24B-A2B has a hosted API on OpenRouter ($0.21 per million tokens); the other has to be self-hosted.
Pick LFM2.5 24B-A2B if…
- you want the strongest model overall
- you need precise instruction following
- you want a hosted API rather than running it yourself
- you work with long documents
- you want the more widely used model (more quantized builds and community support)
Pick Tri-21B Think if…
- you need a permissive license (Apache-2.0)
Specifications
| LFM2.5 24B-A2B | Tri-21B Think | |
|---|---|---|
| Small LM Index | 61.4 (likely 57–65) | 43.6 (likely 40–48) |
| Lab | Liquid AI · United States | Trillion Labs · Rest of Asia |
| Released | 2026-10-13 | 2026-02-18 |
| Parameters | 24B (2B active, MoE) | 20.7B (dense) |
| Memory at 4-bit | ~15.7 GB | ~15.1 GB |
| Context window | 128K tokens | 32K tokens |
| Input | text | text |
| Reasoning | No thinking | Always thinks |
| License | LFM Open License | Apache-2.0 |
| API price / 1M tokens | $0.21 ($0.050 in, $0.70 out) | Self-host only |
| Hugging Face downloads (30 days) | 32.8K | 459 |
Benchmarks
Higher is better. Bold marks the winner where both labs report a score. ≈ marks the index model's estimate for a score that wasn't reported, ✓ an independent measurement and * a score measured by another lab.
| Benchmark | LFM2.5 24B-A2B | Tri-21B Think |
|---|---|---|
| MMLU-Pro Knowledge | ≈ 79 | 74.3 |
| GPQA Diamond Knowledge | 74.4 | 62.6 |
| Humanity's Last Exam Reasoning | 7.4 | 5.5 |
| AIME Math | ≈ 85 | 56.7 |
| LiveCodeBench Coding | ≈ 71 | 53.7 |
| SWE-bench Verified Coding | 51.0 | ≈ 24 |
| IFBench Instruction following | 74.7 | 56.7 |
| IFEval Instruction following | 96.8 | 84.7 |
| BFCL Agentic | 65.0 | — |
| MATH-500 Math | 96.0 | — |
| AA-LCR Long context | — | 11.0 |
| SciCode Coding | 32.7 | — |
Scores are mostly reported by each lab with its own settings, so small gaps are noise. See the methodology. Open in the interactive comparison to add more models.
Questions
Which is better, LFM2.5 24B-A2B or Tri-21B Think?
LFM2.5 24B-A2B scores higher on the Small LM Index: 61.4 vs 43.6. On the 4 benchmarks both labs report, LFM2.5 24B-A2B wins 4 and Tri-21B Think wins 0.
Which is cheaper to run?
Only LFM2.5 24B-A2B has a hosted API on OpenRouter ($0.21 per million tokens); the other has to be self-hosted. Both need about 15.7 GB of memory at 4-bit quantization.
Which has the longer context window?
LFM2.5 24B-A2B has the longer context window: 128K tokens against 32K.
Can I use LFM2.5 24B-A2B and Tri-21B Think commercially?
LFM2.5 24B-A2B is released under LFM Open License and Tri-21B Think under Apache-2.0. Apache 2.0 and MIT allow commercial use; other licenses can add conditions, so read the license before shipping a product.