Kimi K3 vs GLM-5.2: Benchmarks, Pricing & Which Wins
July 25, 2026 · Updated September 25, 2026 · kimi k3 vs glm 5.2 · glm 5.2 vs kimi k3 · kimi k3 vs glm · open weight ai models · ai model comparison
Two weeks before Kimi K3 (July 16, 2026) reset expectations for open-weight AI, GLM-5.2 (June 13, 2026) held the crown: Zhipu AI’s 744B coding-first flagship was the highest-scoring open-weight model ever measured. Now the two Beijing labs go head-to-head — and this is the most important comparison in open-weight AI right now, because both models target the same buyer: teams building agents and coding pipelines who refuse to pay closed-model prices.
The independent verdict is already in, and it splits cleanly: K3 is the stronger model. GLM-5.2 is the more practical one. K3 wins every shared capability benchmark and adds native vision. GLM-5.2 costs a third as much per token, runs roughly twice as fast, and its MIT-licensed weights fit on a single node. Here’s every number, plus a clear verdict for each type of team.
Update (Sept 25, 2026): K3’s weights have been public since July 27 (see Kimi K3 open weights). Artificial Analysis has since re-versioned its Intelligence Index: on v4.3.2, K3 scores 44 (#3 of 115 large open-weight models) and GLM-5.2 scores 34. The 57 vs 51 figures below come from the July v4.1 version and aren’t comparable. Current measured speeds are ~35 tok/s for K3 and ~75 tok/s for GLM-5.2 on Z.ai’s own API. For Zhipu’s newer, cheaper model, see Kimi K3 vs GLM-5.3-Flash.
The two models at a glance
| Kimi K3 | GLM-5.2 | |
|---|---|---|
| Maker | Moonshot AI | Zhipu AI (Z.ai) |
| Released | July 16, 2026 | June 13, 2026 |
| Parameters | 2.8T MoE (16/896 experts) | 744B MoE (~40B active; 753B per Artificial Analysis) |
| Context window | 1M tokens (1,048,576) | 1M tokens (1,048,576) |
| Max output | Not publicly capped | 131K tokens |
| Modalities | Text + image + video | Text only |
| Reasoning | Always on; low / high / max | High / Max toggle (reasoning_effort) |
| Input price (per 1M) | $3.00 ($0.30 cached) | $1.40 ($0.26 cached) |
| Output price (per 1M) | $15.00 | $4.40 |
| License | Custom Kimi K3 License (weights out since July 27) | MIT (single-node friendly) |
| Self-hosting | 64+ accelerators recommended | Single 8×H200 node (FP8) |
| Output speed | ~35 tok/s | ~75 tok/s (200+ on fast third-party hosts) |
| AA Intelligence Index (v4.3.2) | 44 (#3 open-weight) | 34 (#12 open-weight) |
Pricing: GLM-5.2 is roughly 2–3× cheaper per task
| Billing tier (per 1M tokens) | Kimi K3 | GLM-5.2 | GLM cheaper by |
|---|---|---|---|
| Input (standard) | $3.00 | $1.40 | 2.1× |
| Input (cache hit/read) | $0.30 | $0.26 | 1.2× |
| Output | $15.00 | $4.40 | 3.4× |
On Artificial Analysis’s blended basis (7:2:1 cached:input:output), that’s $2.31 per million tokens for K3 vs $0.90 for GLM-5.2. The number that matters most — cost per completed task — landed at ~$0.94 vs ~$0.32 on AA’s July index (v4.1), and sits at ~$2.00 vs ~$0.96 on the current v4.3.2 suite. One dollar buys roughly 67K output tokens from K3 versus 227K from GLM-5.2. For high-volume agent loops, that 2–3× gap compounds into a serious budget difference.
K3’s counter: Moonshot’s Mooncake inference stack reports >90% cache-hit rates on coding workloads, pulling most real input bills to the $0.30 tier — and K3 needs fewer retries on hard tasks, which partially closes the per-task gap. GLM-5.2’s counter-counter: its High/Max reasoning toggle cuts token burn 2–2.5× on routine work, while K3’s low/high effort modes only shipped post-launch — though K3’s new k3-256k variant (~half quota for ≤256K workloads) now narrows the cost-control gap. Zhipu also sells GLM Coding Plan subscriptions (from roughly $10–18/month) for predictable-cost coding access, and third-party providers list GLM-5.2 as low as $0.77/$2.42 per million tokens.
Benchmarks: K3 wins every shared test
Moonshot’s launch suite ran both models through matched agentic harnesses — this is the cleanest capability comparison available:
| Benchmark | Kimi K3 | GLM-5.2 | K3 margin |
|---|---|---|---|
| DeepSWE | 67.5 | 46.2 | +21.3 |
| Program Bench | 77.8 | 63.7 | +14.1 |
| Terminal-Bench 2.1 | 88.3 | 82.7 | +5.6 |
| FrontierSWE | 81.2 | 67.3 | +13.9 |
| SWE Marathon | 42.0 | 13.0 | +29.0 |
| Automation Bench | 30.8 | 12.9 | +17.9 |
| GPQA-Diamond | 93.5 | 91.2 | +2.3 |
The shape of these numbers matters. On bounded coding tasks (Terminal-Bench, Program Bench) GLM-5.2 stays within striking distance. On long-horizon autonomous work — SWE Marathon, Automation Bench — the gap explodes to 3×+. That’s exactly the workload class K3 was built for, and it’s where GLM-5.2’s smaller active-parameter budget shows.
What independent testing says
Artificial Analysis ran both on its standardized Intelligence Index v4.1 (same harness, max effort) at launch in July. AA has since re-versioned the index to v4.3.2, where K3 scores 44 and GLM-5.2 scores 34; scores aren’t comparable across versions, but the ordering held. The July v4.1 breakdown:
| Evaluation | Kimi K3 | GLM-5.2 (Max) |
|---|---|---|
| Intelligence Index (overall) | 57 | 51 |
| Coding Index | 76 | 69 |
| Agentic Index | 50 | 43 |
| GDPval-AA v2 (Elo) | 1,668 | 1,514 |
| AutomationBench-AA | 53% | 28% |
| Terminal-Bench v2.1 | 85% | 78% |
| GPQA Diamond | 94% | 89% |
| Humanity’s Last Exam | 44% | 40% |
To GLM-5.2’s credit: its v4.1 score of 51 was the highest open-weight result ever before K3 shipped five weeks later. And there’s one metric where GLM-5.2 clearly wins — AA-Omniscience non-hallucination rate: 72% vs K3’s 49%. GLM-5.2 is far more likely to refuse or defer rather than fabricate an answer. If your application punishes confident wrong answers harder than unanswered questions, that 23-point gap deserves real weight.
Zhipu’s own published numbers (not in Moonshot’s suite) are also respectable: SWE-bench Pro 62.1 (ahead of GPT-5.5’s 58.6), AIME 2026 99.2%, and a real-world demo completing root-cause analysis across 740,000 server log lines in one session.
Speed: GLM-5.2 is roughly 2× faster
Artificial Analysis currently measures GLM-5.2 (max) on Z.ai’s own API at roughly 75 output tokens/sec vs K3’s ~35 on Moonshot’s endpoint (launch-week figures were ~168 vs ~62), and the fastest third-party GLM-5.2 hosts exceed 200 tok/s. That’s the largest practical difference between these two models after price. K3 is a deliberate, always-max-thinking engine built for long autonomous runs; GLM-5.2 at High effort is built for interactive loops where a developer is waiting. For chat-facing products and IDE autocomplete-adjacent flows, GLM-5.2 feels dramatically snappier.
Context and output
Both models advertise a 1M-token context window — a genuine tie, and a shared engineering flex (GLM-5.2 uses IndexShare attention to cut per-token FLOPs 2.9× at 1M; K3 uses Kimi Delta Attention for 6.3× faster long-context decoding and scored 90.4 on a full-1M eval with no context management). GLM-5.2 guarantees a 131K max output; K3’s output isn’t publicly capped. Filling the window costs $1.40 with GLM-5.2 versus $3.00 with K3.
Open weights and self-hosting: different leagues of “open”
This is where the comparison stops being about benchmarks:
- GLM-5.2 is open today. Full weights on Hugging Face and ModelScope under a plain MIT license, no regional restrictions. Community quantizations already span BF16 (~1.5TB), FP8 (~750–860GB — fits a single 8×H200 node), and 4-bit GGUF (~376GB, runs on 4×H100 via llama.cpp). Day-one support across vLLM, SGLang, Transformers, and KTransformers, plus an Anthropic-compatible endpoint so Claude Code, Cline, and OpenCode work out of the box.
- Kimi K3 has been open since July 27, 2026 — 2.8T weights on Hugging Face (
moonshotai/Kimi-K3) under the custom Kimi K3 License: commercial use and fine-tuning are allowed, Model-as-a-Service businesses with more than $20M revenue over any 12 months need a separate agreement with Moonshot, and a “Kimi K3” attribution clause applies only above 100M monthly active users or $20M monthly revenue (irrelevant for nearly everyone). But Moonshot recommends supernode deployments of 64+ accelerators. Even in native MXFP4, K3’s weights are ~1.56TB before KV cache (full breakdown in our VRAM requirements guide).
The translation: a single well-equipped team can self-host GLM-5.2 this afternoon. Self-hosting K3 is a decision for companies and labs with serious infrastructure. If “single Hopper-class node, fine-tunable” is your requirement, GLM-5.2 is the only answer in this comparison.
Multimodality: K3’s clean sweep
GLM-5.2 is text-only. K3 natively understands images and video — which powers use cases GLM-5.2 simply can’t touch: screenshot-to-code, UI mockup implementation, video analysis, and visual debugging. It’s a big part of why K3 holds the #1 Frontend Code Arena ranking while GLM-5.2 doesn’t appear in that arena at all. If your agents look at screens, this section is the whole comparison.
Kimi K3 vs GLM-5.2: which should you choose?
| Your situation | Pick |
|---|---|
| Long-horizon autonomous agent runs (SWE Marathon +29 pts) | Kimi K3 |
| Frontend / UI development (#1 Arena, native vision) | Kimi K3 |
| Hardest reasoning and knowledge work (44 vs 34 on AA’s current index) | Kimi K3 |
| Multimodal inputs (screenshots, mockups, video) | Kimi K3 |
| Bounded coding & terminal tasks at scale | GLM-5.2 (near-tied, 3× cheaper) |
| High-volume production workloads ($0.96 vs $2.00/task on AA v4.3.2) | GLM-5.2 |
| Latency-sensitive / interactive products (~75 vs ~35 tok/s) | GLM-5.2 |
| Self-hosting on a single node, fine-tuning | GLM-5.2 (MIT, fits one 8×H200 node) |
| Hallucination-sensitive applications (72% vs 49%) | GLM-5.2 |
| Air-gapped deployment on a single Hopper node | GLM-5.2 (K3 needs 16×H200 or 8×B300-class) |
| Best absolute capability, cost secondary | Kimi K3 |
The pattern many teams have converged on: run GLM-5.2 as the high-volume default, escalate to K3 for the hardest, multimodal, or longest-horizon tasks. You capture GLM-5.2’s speed and price on 80–90% of traffic and K3’s ceiling where it matters. The Kimi app’s free tier lets you try Kimi, but confirmed K3 access (and the full 1M context) needs a paid plan; the API unlocks after a $1 top-up, which is enough to benchmark the escalation tier on your real tasks today. Here’s how to get K3 on your PC.
FAQ
Is Kimi K3 better than GLM-5.2? On raw capability, yes. K3 leads 44 vs 34 on the current Artificial Analysis Intelligence Index (v4.3.2; it was 57 vs 51 on the July v4.1 version) and beats GLM-5.2 on every shared coding benchmark in Moonshot’s matched-harness suite (DeepSWE 67.5 vs 46.2, Terminal-Bench 2.1 88.3 vs 82.7, SWE Marathon 42.0 vs 13.0). K3 also adds native vision and video. But GLM-5.2 costs about a third of K3’s per-token list price, runs roughly 2x faster, and its MIT weights fit on a single 8x H200 node.
Is GLM-5.2 cheaper than Kimi K3? Yes, significantly. GLM-5.2 costs $1.40 per million input tokens and $4.40 per million output tokens ($0.26 cached read), versus K3’s $3.00/$15.00 ($0.30 cached). On a blended 7:2:1 agentic workload that’s roughly $0.90 vs $2.31 per million tokens. Artificial Analysis puts cost per Intelligence Index task at about $0.96 vs $2.00 on the current v4.3.2 index ($0.32 vs $0.94 on July’s v4.1), so GLM-5.2 is roughly 2-3x cheaper per completed task.
Which is faster, Kimi K3 or GLM-5.2? GLM-5.2, by a wide margin: Artificial Analysis currently measures GLM-5.2 (max) on Z.ai’s own API at roughly 75 output tokens per second versus about 35 for Kimi K3 on Moonshot’s endpoint (launch-week figures were ~168 vs ~62), and the fastest third-party GLM-5.2 hosts exceed 200 tok/s. K3 defaults to maximum thinking effort, which slows it further (low/high modes and a half-quota k3-256k variant have since shipped); GLM-5.2’s High/Max reasoning toggle lets you trade depth for speed on routine tasks.
Can I self-host Kimi K3 or GLM-5.2? Both are downloadable now, but they live in different hardware classes. GLM-5.2’s full 744B weights are on Hugging Face under the MIT license: FP8 fits a single 8x H200 node, and 4-bit GGUF runs on 4x H100. K3’s 2.8T weights were released on July 27, 2026 under the custom Kimi K3 License (commercial use allowed; very large Model-as-a-Service businesses need a separate agreement), but serving them takes roughly 1.7TB+ of VRAM and Moonshot recommends 64+ accelerators for production.
Kimi K3 vs GLM-5.2 for coding — which wins? K3 wins on capability: it leads every shared coding benchmark and ranks #1 on the Frontend Code Arena, with native vision for screenshot-to-code work. GLM-5.2 wins on economics for bounded coding tasks: on Terminal-Bench 2.1 the two are nearly tied (82.7 vs 88.3 by Moonshot’s numbers, 78 vs 85 by Artificial Analysis) while GLM-5.2 costs about a third as much per token and runs roughly 2x faster. Long-horizon autonomous coding goes to K3 decisively (SWE Marathon 42.0 vs 13.0).
Should I use both together? Yes — a common production pattern: GLM-5.2 as the high-volume default, K3 for the hardest reasoning, multimodal inputs, and long autonomous runs. Both expose OpenAI-compatible endpoints, so routing between them is a config change, not a rewrite.
The verdict
Kimi K3 vs GLM-5.2 isn’t really a fight — it’s a stack. K3 is one of the strongest open-weight models measured (#3 of 115 on AA’s current index): it sweeps every shared benchmark, sees images and video, and since July 27 has been the largest downloadable model in history. GLM-5.2 is the most practical open-weight model ever shipped: a third of the price, nearly 3× the speed, fewer hallucinations, plain-MIT weights that fit on one node, and a reasoning toggle for cost control — available right now.
If you can only pick one: GLM-5.2 for volume, K3 for ceiling. For Zhipu’s newer model, see Kimi K3 vs GLM-5.3-Flash. For the closed-model fights, see our Kimi K3 vs Claude Opus 5 and Kimi K3 vs Claude Fable 5 breakdowns — or start from the complete Kimi K3 guide.