Inkling is not really that good.. (U.S. Version of GLM 5.2)

TL;DR

• Inkling still produced broken results after six commits across Ron’s coding benchmarks; the default setup is not ready for dependable work.
• Its headline size—975 billion total parameters and 41 billion active parameters—did not translate into better results than smaller alternatives in this test.
• Helm’s Deep, Hogwarts, Jabberwock, low-poly world, and Office Life all failed in different, useful-to-diagnose ways.
• The test was a one-shot, sub-agent workflow. Direct coding, temperature changes, role prompting, and fine-tuning were not tested.
• The reason to keep watching is open weight, meaning the model weights can be adapted: Inkling can be fine-tuned and paired with a stronger supervisor.

Ron’s verdict

Inkling is a no for default-settings coding today. Six commits without a working result is not a minor polish issue; it is a reliability failure. The open weights stop this from being a permanent write-off, but that is potential, not proof. If you need work completed now, use a model that already passes your task. If you are evaluating Inkling, treat it as a tuning project and put a second agent in the reviewer seat. (source video RMEcJXp80XM, 00:17)

Key moments

• 00:00 — The result up front: six commits, continued breakage, and a disappointed first verdict.
• 01:44 — Inkling’s architecture: 975B total parameters, 41B active, and the GLM 5.2 comparison.
• 02:36 — Helm’s Deep after six iterations: from a placeholder to a dark, still-broken scene.
• 03:49 — The GPT Luna comparison: a cheaper model completes the same task on its first attempt, despite inverted movement.
• 04:58 — The supervisor pattern: GLM 5.2 acts as the judge, then sends its feedback into Inkling’s shared workspace through tmux.
• 05:28 — The rest of the benchmark suite: Hogwarts, Jabberwock, low-poly world, and Office Life expose different failure modes.
• 07:28 — How Ron connected Inkling: free playground, API credits, and a custom OpenAI-compatible provider in OpenCode.

Useful quotes

“the point is we have six commits, and it’s still broken.” — Ron, source video RMEcJXp80XM, 00:30

“for now, on the default setting, it’s not good.” — Ron, source video RMEcJXp80XM, 01:35

“It just wants to get the job done ASAP. It’s minimum wage effort.” — Ron, source video RMEcJXp80XM, 04:10

“open weight models are certainly limitless in terms of potential, but so far the default setting is not impressive, but at least there’s the fine-tuning option.” — Ron, source video RMEcJXp80XM, 08:13

What the benchmark actually showed

This was not one vague “it feels bad” judgment. Each task exposed a specific failure that matters when you are deciding whether a model can operate inside a coding loop.

Test	Observed result	Why it matters
Helm’s Deep	The first run was a placeholder URL, the second was almost completely dark, and commit six was still broken.	Repeated iteration did not recover the core deliverable.
Hogwarts	The sixth commit remained broken.	The problem was not isolated to one prompt.
Jabberwock	The poem-to-game output was dizzying and unplayable.	A generated scene is not a successful game if the player cannot use it.
Low-poly world	It rendered as a 3D drawing, but movement did not work.	Visual completion hid a missing interaction system.
Office Life	Workers piled up at reception; meetings teleported them across unrelated rooms; lunch repeated the same movement logic.	The model appeared to fake movement rather than implement the simulation rules.

Those observations come from Ron’s sixth-commit review of the suite, not a claim about every possible Inkling configuration. (source video RMEcJXp80XM, 05:28)

The comparison with GPT Luna is useful because it separates finish rate from sophistication. Luna’s movement was inverted, so its output was not clean. It still produced a complete scene on the first attempt, while Inkling had not completed the task after six commits. For an operator, a flawed result that exists can be repaired; a loop that never reaches a usable artifact is harder to trust. (source video RMEcJXp80XM, 03:49)

Big model, weak default result

Inkling is a mixture-of-experts (MoE) model: it has a large pool of parameters but activates only part of that pool for each token. Ron reports 975B total and 41B active parameters. His comparison point, GLM 5.2, has 753B total and 40B active, giving GLM the higher active-to-total ratio despite Inkling’s larger headline number. Inkling was trained from scratch on a reported 45 trillion tokens across text, images, audio, and video, with an MoE design that largely follows DeepSeek V3. (source video RMEcJXp80XM, 01:44)

The practical lesson is simple: parameter totals do not rescue a weak task result. Specs tell you what might be possible. Your benchmark tells you whether the current model, configuration, and harness can do your job.

The test boundary you should preserve

Ron used one-shot prompts that delegated implementation to sub-agents. He did not test Inkling doing all coding directly, and he did not complete a fine-tuning pass. Direct execution might behave differently, but it would also take longer and consume more credits. Ron topped up $10 and had about $6 remaining at this point in the tests. (source video RMEcJXp80XM, 06:39)

That boundary matters. The honest conclusion is “default Inkling failed this harness,” not “Inkling can never code.” If you retest it, change one variable at a time:

1. Run the same prompt with Inkling coding directly instead of delegating.
2. Keep the acceptance checks identical so the comparison remains measurable.
3. Adjust role and temperature only after recording the default result.
4. Fine-tune on a narrow task, then rerun the same benchmark.
5. Use a separate reviewer agent to identify defects and force another iteration.

These are the untested directions Ron proposes, not confirmed improvements. (source video RMEcJXp80XM, 06:56)

The supervisor pattern is the useful part

The most reusable workflow in the video is not an Inkling feature. GLM reviewed Inkling’s work and acted as the judge (05:00). Ron kept GLM and Inkling in the same workspace and used tmux, a tool for keeping multiple terminal sessions alive, to pass feedback into Inkling’s session without copying the long review prompt himself (05:16).

That gives you a clean division of labor: the cheaper or experimental model builds; the stronger model inspects against explicit requirements; the builder receives concrete defects and tries again. It is only useful if the loop has a stop condition. In this test, six failed commits were enough evidence to stop spending credits on the default setup.

Should you try Inkling?

• Need dependable coding output now? Skip the default setup. This test did not establish production reliability.
• Want a free first look? The video shows a free playground on the lab’s site. (source video RMEcJXp80XM, 07:28)
• Want it inside an IDE? Ron used OpenCode with a custom OpenAI-compatible provider, put the API key in his shell configuration, and verified the model through /models. The transcript does not expose the full base URL or config, so this companion does not invent a pasteable snippet. (source video RMEcJXp80XM, 07:40)
• Prepared to tune and measure? This is the credible use case. Fine-tune narrowly, preserve the benchmark, and compare before-and-after results.

What changed since this video

This video was published on July 16, 2026, and this companion was source-checked on July 17, 2026. The supplied source packet contains no completed fine-tuning follow-up or later benchmark result. Treat the default-settings failure as the observed result and the proposed tuning upside as unverified. Ron said a follow-up would cover fine-tuning if the video reached 500 likes; this companion does not assume that follow-up happened. (source video RMEcJXp80XM, 08:35)

Related

• Chinese open-weight models: what “open” changes
• Cost-efficient model tier: pick by completed work, not headline size
• Frontier vs open-weight: a practical decision framework