Video summary
Let's Run Inkling 2: USA's Local AI That FINALLY Beats Qwen & GLM 🤯
Main summary
Key takeaways
Inkling 2 Small Review
The video reviews Inkling 2 Small, a newly released open-weights model from Thinking Machines. The presenter runs it locally and tests its coding, visual-generation, game-building, reasoning, and knowledge abilities, comparing it with models including GLM, DeepSeek, Qwen, and a model referred to as “Ling.”
Model and Local Setup
- The presenter describes Inkling 2 Small as a multimodal, open-weights model. Its four-bit quantized version is said to occupy about 400 GB, while the stated deployment requirements call for several H200 GPUs.
- Rather than using a GPU cluster, the presenter runs the model on a Mac. The video reports a generation speed of roughly 16 tokens per second in general use, rising to around 30 tokens per second during a short Tetris test.
- The model offers several reasoning-effort settings, from off to maximum. The presenter finds that more reasoning does not consistently improve results: medium effort produced a good human-face demo, while higher settings sometimes used more tokens without improving the output.
- The presenter also tests different temperature settings and observes that the model completes some tasks reliably even at higher temperatures. They caution that quantization can affect token selection and potentially code quality.
Benchmarks and Comparisons
- The presenter cites benchmark results that appear to place Inkling ahead of GLM Flash and GLM 5.3 on one coding measure, but behind DeepSeek V4.1 Flash in another comparison. Some benchmark figures and model names are unclear in the auto-generated subtitles, so these should be treated as the presenter’s reported results rather than independently verified findings.
- The presenter questions benchmark comparisons in which models may use special tools, and notes that their own experience with GLM 5.3 can differ from benchmark rankings.
- In hands-on visual comparisons, the presenter says “Ling” produced a stronger volcano animation, while Inkling performed especially well on a complex planetary-rendering prompt.
Coding and Generation Tests
- 3D visual effects: Inkling generated a volcanic eruption with projectiles and produced a detailed interactive planet scene. The planet test worked with reasoning turned off and included alternate planet appearances, asteroid impacts, and a Death Star-like object. The presenter calls this one of the strongest results in the video.
- Human-face renderer: At medium reasoning effort, the model generated a runnable animated face with blinking eyes. Higher effort did not clearly improve it, and one version had a camera or axis issue.
- 3D helicopter: It produced an interactive model with X-ray and wireframe views, disassembly and reassembly, animated propellers, and takeoff. The presenter reports no execution errors.
- Website and infographic: The website received positive marks for its visual style, layout, and widgets. The infographic prompt produced something closer to a slideshow than the animated infographic the presenter had in mind.
- Music-generation app: The application interface worked and included playback and instrument controls, but the generated musical material was judged poor.
- Games: The model created a mouse-chasing game, a go-kart game, and an amusement-park shooter. The presenter particularly praised the park’s roller coaster and found the kart track and gameplay promising, though some graphics and polish could be improved.
- GTA-style game and agent loop: A single-file version was comparatively basic. When run in an agent loop with a multi-file project, the model generated much more code, tested and revised it, and produced a more elaborate game. The result had playable elements but also bugs, confusing behavior, and performance issues. The presenter emphasizes that tool calls and a project folder can improve larger builds, while excessive files and growing context can cause problems.
- Simple prompts: An animated weather-card task generated working code quickly, though the result was basic.
Reasoning, Knowledge, and Overall Assessment
- The model handled a basic arithmetic distribution question and a car-versus-walking question reasonably, including recognizing that a distant car wash would be more practical to reach by car.
- It made mistakes on several knowledge and recognition tests, including identifying a song, answering some questions about less-common APIs such as WinRT, and recalling certain technical details. The presenter found its basic Swift knowledge better than its answers to some C++ and UTF-16 questions.
- The presenter also observed safety-related refusals or limitations in response to sensitive prompts, including a request for a nuclear-explosion visualization.
- Overall, the review portrays Inkling 2 Small as a capable local model with notably strong visual coding and agent-assisted application generation. Its best demonstrations often worked without runtime errors, but results varied by task. More reasoning did not guarantee better output, and the model still struggled with some factual, niche-technical, and music-generation tasks. The presenter considers it a promising American open model but says further testing is needed.
Main Speaker and Sources
- Main speaker: The xCreate presenter, who narrates the local tests and comparisons.
- Sources discussed: Thinking Machines, the model’s published deployment and benchmark claims, and comparison models including GLM, DeepSeek, Qwen, and “Ling.”
Rate this summary
Your feedback will help improve summaries.
Improve this summary
Reprocess with a stronger model when the summary feels incomplete or inaccurate.
Translate summary in another language
Ask questions to this video
Chat for follow-up questions, clarifications, and source-backed answers.