TL;DR: Two open-weight families. Kimi K3 had the fewest failed steps of its test, 7.2%, and took the fewest steps overall (Jul 31, 2026). GLM-5.2 was the only model to refuse a login screen nobody asked for (Jul 30, 2026), and started saving your data properly on Aug 1, 2026. Route by task inside Taskade Genesis.
What TSK-1 Found
We have tested both families, and the findings split on two different kinds of reliability rather than producing a single winner. Kimi K3 led its test on execution hygiene: the fewest failed steps (7.2%), the fewest steps overall, and the most accurate account of what it had built (Jul 31, 2026). GLM-5.2 led on doing the right thing with a loose request, refusing the login screen nobody asked for that other models shipped (Jul 30, 2026), and it started saving your data properly on Aug 1, 2026.
- Kimi: Jul 31, 2026 — fewest failed steps of the test (7.2%), fewest steps overall, most accurate account of its own work.
- GLM: Jul 30, 2026 — refused the login screen nobody asked for, the only model to push back on quietly added work; Aug 1, 2026 — it saved your data properly, with a cosmetic theme issue as the only thing left open.
See the full evidence at /tsk/kimi, /tsk/glm, and the TSK-1 hub.
Kimi K3 vs GLM-5.2
Both are open-weight, both reach a million tokens of context, and we graded them on different strengths. Kimi K3's finding is about execution hygiene: on Jul 31, 2026 it had the fewest failed steps of its test at 7.2%, took the fewest steps to a finished build, and gave the most accurate account of what it had and had not built. In agentic builds, how often a model's steps fail is the closest thing to a reliability floor: every failed step is a wasted step and a muddled context, and Kimi led that test on exactly that.
GLM-5.2's finding is about judgment. On Jul 30, 2026 the test left room for scope-creep — a login screen nobody had asked for. GLM thought about it and declined, offering an "Add Login" suggestion instead. It was the only model in that test to push back. On Aug 1, 2026 it started saving your data properly — the family had managed none of that on Jul 30 — with a cosmetic theme issue as the only thing between it and a clean pass. A model that refuses work you did not ask for and still writes real data is the story this testing exists to find.
Choose Kimi If…
A comparison that never concedes anything is not worth reading. Kimi K3 is the better pick in several common cases.
- The agent runs long tool-using jobs. The fewest failed steps of its test (Jul 31, 2026, 7.2%) is the measure that most directly predicts a smooth agentic build.
- You want the largest open-weight model available. 2.8 trillion total parameters with 104 billion active is the top of the open-weight range.
- Your inputs include images. Kimi K3 is multimodal, taking image and text input.
- You are self-hosting. Downloadable weights under the Kimi K3 License — read it before redistributing. The native MXFP4 checkpoint is about 1.56 TB and fits one 8x B300 node, but not an 8x B200 or 8x H100 node.
Choose GLM If…
- The request is loose and scope-creep is a real risk. GLM is the model that asked before it added a login screen nobody wanted (Jul 30, 2026).
- You want a published managed rate card. Zhipu publishes per-token pricing, cached-input rates, and coding plans on z.ai.
- Long-horizon engineering is the job. GLM-5.2 is positioned for exactly that, with a 1M-token context and 128K max output.
- You want the more permissive license of the two. GLM-5.2's published weights are MIT, with no revenue threshold and no regional restriction; Kimi K3's bespoke license needs reading before you redistribute. The managed API on z.ai means you can start metered and self-host later without renegotiating anything.
The Taskade Angle: Route, Don't Standardize
Most comparison pages end with "pick one". The evidence for these two families points the other way: one model owns execution hygiene, the other owns doing the right thing with a loose request. Serious teams run both and route between them.
Taskade routes across frontier models from top AI labs inside one workspace, with the AI allowance included in the subscription rather than a separate API account per lab. Paid plans start at Pro $10/mo billed annually, with Business $25, Max $100 and Enterprise $250 per month billed annually. Kimi K3 runs on Taskade AI credits on Max and Enterprise (see the Kimi models page). On Business and up, a Together AI key lets custom agents and Ask Agent pick Kimi K3. On Pro and up, Taskade EVE can build and edit your Taskade Genesis apps on your key too: pick a model under Your keys in the Taskade EVE model menu, and each Taskade EVE step on your key uses one fixed step fee in credits. You set the model per agent or per automation step — Kimi-style tool-heavy steps and GLM-style ambiguity-sensitive steps can each get the model that leads there. Leave a step on TSK-1 Auto and it adapts the depth instead — fast when the step is quick, deeper reasoning when it is not.
Four patterns that hold up:
- Hygiene model for the loop, judgment model for the request. Long tool-using jobs on Kimi K3; unclear requirements and scope decisions on GLM-5.2.
- Open-weight in the loop, governed model on the output. Both run where the work happens; the paragraph a customer reads can still get a closed model.
- Every step lands in the same project graph. Whichever model runs a step, the result becomes shared workspace memory, so the next agent inherits context instead of re-deriving it.
- Scheduled automations read from the same place. Model choice becomes a per-step setting, not a platform decision.
See 10 Best Open-Source AI LLMs in 2026 for how both families sit in the wider open-weight field.
Frequently Asked Questions
Which is better, Kimi or GLM?
They win on different measures, and both are worth having in one workspace. Kimi K3 had the fewest failed steps of its test — only 7.2% went wrong, with the fewest steps overall and the most accurate account of what it had and had not built (Jul 31, 2026). GLM-5.2 produced the cleanest behavioral finding we have recorded: it thought about adding a login screen nobody asked for, declined, and offered an "Add Login" suggestion instead — the only model to push back on quietly added work (Jul 30, 2026). Pick Kimi for step-efficient tool use. Pick GLM for doing the right thing when the request is open to interpretation. Route both and you get both.
What did TSK-1 measure for Kimi K3?
On Jul 31, 2026, Kimi K3 had the fewest failed steps of its test at 7.2%, took the fewest steps to a working build, and gave the most accurate account of what it had and had not built. How often a model's steps fail is one of the strongest predictors of a smooth agentic build, because every failed step is a wasted step and a muddled context. Kimi led that test on exactly that.
What did TSK-1 measure for GLM-5.2?
GLM-5.2 produced the standout behavioral finding on Jul 30, 2026: when the test left room to add a login screen nobody had asked for, GLM thought about it, declined, and offered an "Add Login" suggestion instead. It was the only model in that test to push back on quietly added work. On Aug 1, 2026 it started saving your data properly — it had not managed that on Jul 30 — with a cosmetic theme issue as the only thing left open. That is what this testing exists to catch: a model that does the right thing when the request is open to interpretation.
Are both Kimi and GLM open-weight?
Yes, and GLM has the more permissive of the two licenses. GLM-5.2's published weights are MIT, with no revenue threshold and no regional restriction, alongside a managed API with a published rate card on z.ai. Kimi K3 ships downloadable weights under Moonshot's bespoke Kimi K3 License, which permits self-hosting but should be read before you redistribute anything built on it. In Taskade, Kimi K3 runs on Taskade AI credits on Max and Enterprise, and on Business and up a Together AI key lets custom agents and Ask Agent pick it.
How big are these models?
Kimi K3 is a mixture-of-experts model at 2.8 trillion total parameters with 104 billion active per token, routing across 896 experts with 16 routed plus 2 shared per token, across 93 layers split into 69 Kimi Delta Attention layers and 24 gated multi-head latent attention layers. GLM-5.2's parameter count is not published in comparable detail; Zhipu publishes capability claims and a rate card rather than an architecture sheet. In every case the active count, not the total, is what drives running cost.
What are the context windows?
Kimi K3 specifies 1,048,576 tokens of context. GLM-5.2 specifies a 1M-token context with a maximum output of 128K tokens, positioned by Zhipu for long-horizon engineering tasks. Both reach a million tokens, so context length is not the deciding factor between them.
How much do they cost?
On published rates as of August 2026, GLM-5.2 is $1.4 per million input tokens and $4.4 per million output, with cached input at $0.26, and GLM-4.7 at $0.6 and $2.2. Moonshot publishes per-token API pricing for Kimi K3 on its platform page at platform.kimi.ai, and this page does not copy that rate card. Kimi's weights are also downloadable, which adds a self-hosting cost floor a metered API does not describe. Verify both rate cards at the source before committing.
Is GLM-5.2 good for coding?
Zhipu positions GLM-5.2 as a long-horizon coding flagship, with a 1M-token context trained specifically for agentic engineering scenarios, and publishes its own benchmark claims including Terminal-Bench 2.1 at 81.0 and SWE-bench Pro at 62.1. Those are vendor-published figures — treat them as vendor claims and validate on your own work. On our side, GLM-5.2's graded strengths are behavioral: handling a loose request correctly (Jul 30, 2026) and saving your data properly since Aug 1, 2026.
Which one should I use for an AI agent?
Match the model to the step. Reach for Kimi K3 when the agent runs long tool-using jobs where every failed step costs you — it led its test on exactly that (Jul 31, 2026). Reach for GLM-5.2 when the agent works from a loose request and you want it to ask before it quietly adds more (Jul 30, 2026). In Taskade you can build a multi-agent team where different agents take different models, all sharing the same workspace memory.
Can I use Kimi and GLM together inside Taskade?
Yes, on the right plan. Taskade routes across frontier models from top AI labs inside one workspace, with the AI allowance included in the subscription rather than a separate API account with each lab. Kimi K3 runs on Taskade AI credits on Max and Enterprise. On Business and up, a Together AI key lets custom agents and Ask Agent pick Kimi K3. On Pro and up, Taskade EVE can build and edit your Taskade Genesis apps on your key too: pick a model under Your keys in the Taskade EVE model menu, and each Taskade EVE step on your key uses one fixed step fee in credits. A common pattern is running bulk steps on one model and reserving the other for the steps that need its specific strength, with every result landing in the same project graph so the next agent inherits the context.
Final Word: Hygiene vs Judgment
Kimi K3 is the step-efficient open-weight flagship — the cleanest tool use of its test, the largest open-weight model, and downloadable weights. GLM-5.2 is the judgment flagship — the model that refused work nobody asked for, with a 1M-token context, MIT weights, and a published rate card.
Neither is the winner. The winner is the setup that puts hygiene where steps matter and judgment where ambiguity lives.
▲ Memory feeds Intelligence. ■ Intelligence triggers Execution. ● Execution creates Memory. Two open-weight families. One workspace. No single point of vendor failure.
This is the origin of living software. 🌱
Build with Kimi and GLM in one workspace →
Related reading
- Moonshot Kimi history — How the Kimi line evolved.
- 10 Best Open-Source AI LLMs in 2026 — Where both families sit in the open-weight field.
- No-code builders that run open-weight models — Where Kimi K3 and GLM run without your own GPUs.
- History of AI benchmarks — Why every model claims to be best, and how TSK-1 differs.
- Kimi vs DeepSeek — Bespoke license versus MIT, head to head.
- Qwen vs GLM — Two open-weight families, public-benchmark view.
- Multi-Model AI Access — How Taskade routes across providers.
- Multi-Agent Teams — Specialists with different model picks.
- Taskade MCP Server — Connect any MCP-compatible IDE to your workspace.
- TSK-1 Kimi profile — Full evidence for the Kimi family.
- TSK-1 GLM profile — Full evidence for the GLM family.
- TSK-1 hub — The complete model test dataset.



