Verdict first
● Inkling 2 Small — 594B total / 17B active, agents and coding
● Hunyuan Hy4 — 770B total, vision and computer use
● Weights — both promised, neither delivered
● The difference — one named a week, the other said soon
Two very large mixture-of-experts models landed three days apart, from labs with nothing else in common. Both were announced with open weights promised and neither has released them. That coincidence is the most interesting thing about the pair, so it is worth being precise about what each one actually published.
Side by side
|
Inkling 2 Small |
Hunyuan Hy4 |
| Announced |
5 November 2026 |
4 November 2026 |
| Lab |
Thinking Machines |
Tencent |
| Total parameters |
594B |
770B |
| Active parameters |
17B |
not published |
| Context |
1M tokens, with compaction |
not published |
| Input |
Text and image |
Text and image |
| Audio |
No |
not published |
| Computer use |
not claimed |
Yes, GUI and command line |
| Coding benchmark |
86.3% SWE-Bench Pro v2 Hard |
not published |
| Availability |
API only, via Tinker |
API, Tencent Cloud |
| Weights |
Apache 2.0, stated for the following week |
Open-sourced "soon", no date |
Note how many cells read not published on one side. These two labs disclose very differently, and that is a finding rather than a gap in the research: Thinking Machines published a benchmark table and an architecture; Tencent published pricing and internal evaluations. Neither published the other's numbers, so most direct comparisons are not available at all.
What each is actually for
Inkling 2 Small: agents and code
The published scores are coding scores — 86.3% on SWE-Bench Pro v2 Hard and 72.1% on DeepSWE v1.1 — and the context design names agent trajectories explicitly, with compaction for long runs. The 1M window matters more for an agent that accumulates tool output over hours than for a chat. The lab also reports an internal misbehaviour rate of 14.9 against 66.8 for the first Inkling, which is a metric nobody outside can reproduce but which tells you what they were optimising.
Hunyuan Hy4: seeing and clicking
Hy4 is the first in the Hunyuan line to support vision and computer use together, with specialised training for Blender and CAD. Tencent published no coding benchmark against the field, reporting instead internal evaluations: 83.18 average with a 96.15% task pass rate on its WorkBuddy office benchmark, and a crowdsourced financial test where it scored 3.04 against 2.97 for DeepSeek-V4.1-Flash-1017. It also published a weak result, 30.4 out of 100 on its own Hy-BioExperimentBench, which is worth crediting.
The open-weight promise is the real comparison
Both labs announced weights they had not released. The difference is falsifiability. Thinking Machines said Apache 2.0 the following week, after external safety testing and red-teaming — a claim that fails visibly if the week passes. Tencent said soon, which cannot fail.
That distinction is worth more than any benchmark in the table above, because it is the one you can check yourself without a GPU. If you are deciding whether to build on either model's weights, the lab that gave you a date has given you something to hold it to.
Which to use today
If you need agent or coding work and can live on an API, Inkling 2 Small has the published numbers to justify a trial. If you need a model that can look at a screen and operate it, Hy4 is the one of the two that claims to do it at all, and the pricing is public at 6 yuan per million input tokens. If you specifically need weights you can run, neither is available yet and only one has told you when to check back.
Frequently asked questions
Which model is bigger?
Hy4, at 770B total against 594B. But Inkling publishes 17B active parameters and Hy4 does not publish an active count, so the figure that governs inference cost is only available for one of them.
Can I download either set of weights right now?
No. Inkling 2 Small is API-only through Tinker with closed weights at launch, and Hy4 was API-only at release. Both labs have said weights are coming.
Which is better at coding?
Inkling 2 Small is the only one with published coding benchmarks: 86.3% on SWE-Bench Pro v2 Hard and 72.1% on DeepSWE v1.1. Hy4 published no comparable figure, so the honest answer is that it has not been measured in public.
What does Apache 2.0 mean for a model?
A permissive licence allowing commercial use, modification and redistribution with attribution, and no copyleft obligation. It is among the least restrictive licences a lab can choose for open weights.
Background: Inkling 2 Small and Hunyuan Hy4.
Sources