GLM-5.2 vs Kimi K3: Open Weights vs API-Only Frontier (2026)

TL;DR

Last updated August 21, 2026.

GLM-5.2 is Zhipu AI's 753B open-weight (MIT) coding model, released June 13, 2026, scoring 62.1 on SWE-bench Pro and 81.0 on Terminal-Bench 2.1 at about $1.40/M input and $4.40/M output. Kimi K3 is Moonshot's 2.8-trillion-parameter flagship, launched July 16, with open weights on Hugging Face since July 27 under Modified MIT, vendor-reported 81.2 on FrontierSWE and 88.3 on Terminal-Bench 2.0 at $3/$15 per M tokens, thinking always on, and native vision. K3 is the capability-and-vision pick; GLM-5.2 is the price-and-practical-self-host pick.

The 30-Second Verdict

These two models sit at opposite ends of nearly every axis, so the choice is rarely close once you know which axis you are on.

• Optimizing for cost per task? GLM-5.2. Roughly a quarter of the output price on its hosted API, and you can cap reasoning effort at high; K3 is locked to max at launch.
• Optimizing for peak capability? Kimi K3. It posts the stronger agentic numbers, and the independent Intelligence Index (60 vs 53) backs the direction, even if the vendor tables overstate the gap.
• Need vision? Kimi K3 only. GLM-5.2 is text-only.
• Need weights today? Both, since July 27. The practical difference is footprint: GLM-5.2 fits one 8-GPU node at 1.5TB BF16; K3 ships natively MXFP4 quantization-aware trained and still needs multi-node serving.
• Long-context-heavy workload? Both run 1M context; K3 prices it flat at $3/$15, GLM-5.2 is cheaper per token but you serve the context budget against a verbose model.

Spec-by-Spec

The parameter counts tell the story in one line: K3 is nearly four times the total size, which is why it costs more to serve and why self-hosting it is a different weight class. Active parameters, the number that drives per-token compute, are closer, roughly 50B for K3 (estimated) versus about 40B for GLM-5.2.

Both hit a 1M-token context, but they get there differently. GLM-5.2 uses IndexShare, a lightweight sparse-attention indexer reused across every four layers, which Z.ai reports cuts per-token FLOPs by 2.9x at 1M length. K3 uses Kimi Delta Attention, a hybrid linear-attention design that interleaves linear and full-attention layers in a 3:1 ratio and cuts KV-cache memory by up to 75%, which is what lets Moonshot price the 1M window flat. Full details on each are on the GLM-5.2 and Kimi K3 pages.

Benchmarks: Vendor-Reported vs Independent

This is the section where care matters most, because the two models publish different benchmarks on different harness versions. GLM-5.2 reports SWE-bench Pro and Terminal-Bench 2.1; K3 reports FrontierSWE and Terminal-Bench 2.0. Terminal-Bench 2.0 and 2.1 are different harnesses, so the 88.3-vs-81.0 gap is not a clean head-to-head. Both vendors self-report, so treat every vendor row as a claim.

The one apples-to-apples number is Terminal-Bench 2.1 run independently: Z.ai reports 81.0 for GLM-5.2, and Vals independently measured 80.9 for K3. On the same harness version, they are effectively tied on terminal-agent tasks, which reframes the vendor-table gap. Where K3 genuinely pulls ahead is the broad Artificial Analysis Intelligence Index, 60 to GLM-5.2's 53. See our SWE-bench Pro breakdown for how these harnesses behave and where scaffolds inflate scores.

The honest read: K3 is the stronger model, but by less than its vendor table suggests. Launch-day Hacker News skepticism centered on benchmark contamination in open-lineage models; independent replication has since accumulated. Together AI's DeepSWE head-to-head measured K3 at 68.5 pass@1, within 1.4 points of Claude Fable 5's 69.9, and Moonshot now publishes the Kimi-Vendor-Verifier conformance harness (OCRBench, MMMU Pro Vision, BEAM-1M, DeepSWE) so anyone can check whether a provider serves the model faithfully.

Price and Token Economics

On sticker price, K3 is about 4x GLM-5.2 on output and roughly 3x on input. But both models are heavy reasoners, so the effective cost per task is what matters, and both have a verbosity tax that the sticker hides.

The verbosity picture cuts against both, but K3 has less headroom. Artificial Analysis measured GLM-5.2 at about 43,000 output tokens per Index task at max effort, 37,000 of them pure reasoning, which pushes its effective cost near $0.46 per task despite the cheap per-token rate. The mitigation is real: GLM-5.2 "high" cuts token usage 2 to 2.5x with little quality drop, and effort is an opt-in you control. K3 gives you no such lever at launch. Thinking mode cannot be disabled and reasoning_effort supports only max, so every request pays a full reasoning trace at $15/M output. Higher rate plus no way to dial down effort means K3's effective cost-per-task gap over GLM-5.2 is wider than the 4x sticker. One more K3 billing detail: multi-turn conversations must replay reasoning_content and tool_calls verbatim, which grows the billed prompt (Moonshot's docs example: 599 tokens replayed versus 198 stripped); harnesses that strip thinking history save those tokens but silently degrade the model.

The one place K3's pricing genuinely wins is flat 1M context. Anthropic charges a premium above 200K input tokens and Google tiers Gemini by context length; Moonshot charges $3/M whether you send 4K or 900K tokens. For repo-scale analysis or long agent traces, that flatness can matter more than the headline rate. GLM-5.2 also prices its 1M window without a long-context tier, at a lower base rate, so for pure long-context economics GLM-5.2 is still cheaper per token; K3's flatness only pulls ahead if you need its specific capabilities in that window.

Verifiability and Open Weights

This is the axis that separates them most cleanly today, and it is the one that decides the "best open-source coding model" question. GLM-5.2 is open in the full sense right now: weights on Hugging Face (zai-org/GLM-5.2) under MIT with no regional restrictions, and its published scores have independent replication. You can download it, fine-tune it, serve it on your own hardware, and audit its behavior.

Kimi K3 launched API-only on July 16, with every number Moonshot-reported. The gap closed in eleven days: independent scores landed within 24 hours, and on July 27 Moonshot shipped the weights on Hugging Face (moonshotai/Kimi-K3, 96 shards plus a technical report) under Modified MIT, following the K2 family (K2, K2.5, K2.6, and K2-Thinking all shipped open). Moonshot is meanwhile consolidating its API lineup on K3: kimi-k2.5 and the moonshot-v1 series are closed to new users and sunset fully on August 31, 2026. Verifiability got its own tool too: Kimi-Vendor-Verifier, a public conformance harness (OCRBench, MMMU Pro Vision, BEAM-1M, DeepSWE) that scores whether a given provider serves K3 faithfully. Both models are now downloadable and auditable. What remains different is the weight class of actually running them.

Vision, Context, and Self-Host

Three capability differences decide many real choices before benchmarks enter the picture:

• Vision. K3 has native multimodal input for images and video; GLM-5.2 is text-only. FastAI's Jeremy Howard, who otherwise rated GLM-5.2 "at least as good as Opus 4.8 and GPT-5.5" for his coding work, flagged the missing vision as its main gap. If you feed the model screenshots, diagrams, or design mockups, K3 is the only option here.
• Context. Both run 1M tokens. The difference is what you fill it with. GLM-5.2 is verbose enough that 37K reasoning tokens per turn can crowd out your code in a long agent run, so context management pays off. K3's KDA design cuts KV-cache memory up to 75%, which is why its 1M window is priced flat, but its always-on max reasoning fills that window fast too.
• Self-host. Both are downloadable now, but at very different cost: GLM-5.2 fits a single 8-GPU node; K3, even with its shipped checkpoint natively MXFP4 quantization-aware trained, needs multi-node clusters. Community efforts already stream it from SSD on a MacBook at 13-16 tokens per minute, which proves the point about practicality.

Which One to Pick, by Workload

A common practical answer is to route: run GLM-5.2 as the high-volume default and reach for K3 on the hard, multimodal, or long-horizon tasks where its capability edge is worth 4x the token bill. That is exactly the pattern the Morph model lineup is built to serve, one OpenAI-compatible API across models, priced per token.

Running GLM-5.2 and Kimi K3 on Morph

Morph serves both models on custom codegen kernels with speculators trained on coding traffic. GLM-5.2 runs as morph-glm52-744b at $0.8415/M input and $3.1365/M output, below Z.ai's own list. Kimi K3 runs as morph-kimik3 at $2.80/M input, $0.29/M cached, and $14/M output, under Moonshot's $3/$15 list on every leg, with a latency-tuned morph-kimik3-fast SKU at $6.00/$22.50. Host choice matters because serving fidelity diverges: many serverless providers quantize activations to FP8 to cut cost, which moves output away from the reference weights. Morph serves GLM-5.2 at bf16.

K3 on Morph also gets program-aware scheduling: tag each agent run's requests with a run_id and every turn routes to the worker holding the run's KV cache, tool-call returns are admitted ahead of new arrivals, and under load whole runs pause instead of every run slowing down. In an agent loop where most input is replayed context, that is what keeps you on the $0.29 cached rate instead of the $2.80 fresh rate. See the Agent Runs docs.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.morphllm.com/v1",
    api_key="YOUR_MORPH_API_KEY",  # Authorization: Bearer <key>
)

resp = client.chat.completions.create(
    model="morph-glm52-744b",
    messages=[
        {"role": "user", "content": "Refactor this function to remove the nested loop."},
    ],
)
print(resp.choices[0].message.content)

The same key works for every model in the Open Source Models lineup, including morph-kimik3, MiniMax M3, Qwen 3.5 397B, and DeepSeek V4 Flash. Swap the model string and the rest of the code is unchanged. Both models also speak the Anthropic Messages API on Morph, so Claude Code runs either with two environment variables; the Kimi API guide covers Moonshot's first-party API if you want to compare hosts.

GLM-5.2 vs Kimi K3: The Trade-Offs

Frequently Asked Questions

Is Kimi K3 better than GLM-5.2 for coding?

On capability, yes, by a modest margin. K3 leads vendor-reported agentic benchmarks (81.2 FrontierSWE, 88.3 Terminal-Bench 2.0) and the independent Artificial Analysis Intelligence Index (60 vs 53). But on the one apples-to-apples independent number, Terminal-Bench 2.1, they are effectively tied (K3 80.9 via Vals, GLM-5.2 81.0). GLM-5.2 wins on price (about a quarter of the output cost), open weights you can hold today, and independent replication of its scores.

What is the price difference between GLM-5.2 and Kimi K3?

K3 is roughly 4x on output: Moonshot charges $3/M input ($0.30 cached) and $15/M output, and Morph serves it at $2.80/$0.29/$14; Z.ai lists GLM-5.2 at $1.40/$4.40, and Morph serves it at $0.8415/$3.1365. K3's always-on max reasoning widens the effective per-task gap beyond the 4x sticker, since you cannot cap its effort at launch.

Can I self-host GLM-5.2 and Kimi K3?

Both, since July 27. GLM-5.2: MIT weights on Hugging Face, ~1.5TB BF16, 8x H200/B200 for a single node. K3: Modified MIT weights on Hugging Face (moonshotai/Kimi-K3), shipped natively MXFP4 quantization-aware trained with day-zero vLLM support, but at 2.8T total, self-hosting still needs multi-node clusters.

Does GLM-5.2 or Kimi K3 have vision?

Only K3. It has native vision for images and video. GLM-5.2 is text-only, a gap Jeremy Howard flagged in his otherwise positive review.

Which is the best open-source coding model in 2026?

Both are open-weight now. K3 (Hugging Face since July 27, Modified MIT) is the higher-capability option; GLM-5.2 (MIT) is roughly a quarter of the price and a quarter of the self-host footprint. Capability pick: K3. Cost-and-practicality pick: GLM-5.2.GLM-5.3 ties K3's 60 on the Artificial Analysis index, but its weights are gated at launch, so it does not yet count as an open-weight option. See Best Open-Source Coding Models in 2026.

How many parameters do GLM-5.2 and Kimi K3 have?

GLM-5.2 is 753B total, ~40B active per token. K3 is 2.8T total, ~50B active (estimated; Moonshot has not officially disclosed the active count). Total drives memory and self-host cost; active drives per-token compute.