vs
1:0
overall capability wins (1 ties across 2 scored). Hy4 Preview leads overall.
- Pricing0:3
- Specs1:0
Verdict
InternLM2.5 20B is ~4.2x cheaper on a blended token basis than Hy4 Preview.
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- Hy4 Preview
- $1.46
- InternLM2.5 20B
- $0.38
InternLM2.5 20B is about 3.9x cheaper on this workload.
Price per 1M tokens
Full comparison
Organization
Tencent
Shanghai AI Lab
Family
Other
Other
License
Apache 2.0
Apache 2.0
Open weights
Yes
Yes
Release
Aug 28, 2026
Aug 1, 2024
Knowledge cutoff
-
-
API / provider
Tencent Cloud TokenHub
Together / Fireworks (ref.)
Modalities
text → text
text → text
Specs
Context window
1M
1M
Max output
-
8K
Parameters
770B (49B act.)
20B
Pricing
Input $/1M
$0.83
$0.30
Output $/1M
$2.50
$0.30
Blended $/1M (3∶1)
$1.25
$0.30
Speed
tok/s
-
70
TTFT (s)
-
0.4
Reasoning
MMLU-Pro
—
—
GPQA Diamond
92.3
—
Humanity's Last Exam
—
—
AIME 2025
—
—
MATH-500
—
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
—
—
AA-LCR v1.1
—
—
CritPt
—
—
MMMU-Pro
—
—
IFBench
—
—
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
65.7
—
SWE-bench Multilingual
82.9
—
LiveCodeBench
—
—
Terminal-Bench 2.1
85.4
—
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
—
—
SciCode
—
—
CursorBench
—
—
SWE-Rebench
—
—
NL2Repo-Bench
—
—
DeepSWE
64.3
—
WebDev Arena
1,624
—
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
—
—
| Metric | Hy4 Preview | InternLM2.5 20B | Delta |
|---|---|---|---|
| Identity | |||
| Organization | Tencent | Shanghai AI Lab | — |
| Family | Other | Other | — |
| License | Apache 2.0 | Apache 2.0 | — |
| Open weights | Yes | Yes | — |
| Release | Aug 28, 2026 | Aug 1, 2024 | — |
| Knowledge cutoff | - | - | — |
| API / provider | Tencent Cloud TokenHub | Together / Fireworks (ref.) | — |
| Modalities | text → text | text → text | — |
| Specs | |||
| Context window | 1M | 1M | tie |
| Max output | - | 8K | — |
| Parameters | 770B (49B act.) | 20B | A +750 |
| Pricing | |||
| Input $/1M | $0.83 | $0.30 | B +$0.53 |
| Output $/1M | $2.50 | $0.30 | B +$2.20 |
| Blended $/1M (3∶1) | $1.25 | $0.30 | B +$0.95 |
| Speed | |||
| tok/s | - | 70 | — |
| TTFT (s) | - | 0.4 | — |
| Reasoning | |||
| MMLU-Pro | — | — | — |
| GPQA Diamond | 92.3 | — | — |
| Humanity's Last Exam | — | — | — |
| AIME 2025 | — | — | — |
| MATH-500 | — | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | — | — | — |
| AA-LCR v1.1 | — | — | — |
| CritPt | — | — | — |
| MMMU-Pro | — | — | — |
| IFBench | — | — | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | 65.7 | — | — |
| SWE-bench Multilingual | 82.9 | — | — |
| LiveCodeBench | — | — | — |
| Terminal-Bench 2.1 | 85.4 | — | — |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | — | — | — |
| SciCode | — | — | — |
| CursorBench | — | — | — |
| SWE-Rebench | — | — | — |
| NL2Repo-Bench | — | — | — |
| DeepSWE | 64.3 | — | — |
| WebDev Arena | 1,624 | — | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | — | — | — |
Hy4 Preview
- Context
- 1M
- Parameters
- 770B (49B act.)
- Price
- $0.83 / $2.50
- Speed
- —
- Modalities
- text → text
- License
- Apache 2.0
Tencent Hy Team's Aug 28, 2026 open-weight flagship (770B MoE / 49B active, Apache 2.0) with a 1M context window and default high reasoning effort. Text-only instruct weights on Hugging Face; hosted API at $0.834/$2.501 per 1M tokens via TokenHub and OpenRouter.
InternLM2.5 20B
- Context
- 1M / 8K out
- Parameters
- 20B
- Price
- $0.30 / $0.30
- Speed
- 70 tok/s · 0.4s TTFT
- Modalities
- text → text
- License
- Apache 2.0
Open InternLM2.5 chat with up to 1M context — strong Chinese RAG and tool-use baseline.
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.
These two models have no overlapping published benchmarks in our dataset. Compare specs and pricing instead.