GLM-5.3 vs Kimi K3: Squeezing a 743B Base or Building a 2.8T One — Which Bet Pays?

The Chinese open-weights race just split into two answers to the same question. Moonshot AI built Kimi K3 from scratch — a 2.8-trillion-parameter mixture-of-experts base, the largest open-weight model ever released, announced July 16, 2026 with weights following on July 27. G​LM-5.3 is the opposite bet: Z.ai kept the exact 743-billion-parameter base that powered GLM-5.2 and spent the entire release cycle on post-training, arguing the base model's intelligence ceiling was never the problem. Both are 1M-context, coding-and-agent-focused flagships, and both claim to be the top open coding model on the market. They cannot both be the best way to spend the same budget, and the benchmarks genuinely split down the middle — which makes this less a comparison than a referendum on how to build the next frontier model.

Two bets on how to build a frontier model

Z.ai calls G​LM-5.3 a "deep dig" rather than a generational upgrade. The base is byte-for-byte the same 743B model as GLM-5.2; the delta is entirely post-training, run through the open-source slime framework on tasks that span whole engineering sessions — diagnosing, fixing, verifying, and shipping across real clusters, storage systems, and codebases, some taking several days of a senior engineer's work. The vendor's reported gains are large: a 50% jump on its own Code Bench, Terminal-Bench 3.0 from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, and a cyber capability that has forced a safety review before the weights ship. Moonshot went the other way. Kimi K3 is a brand-new base — 2.8T total parameters with about 104B active per token (16 of 896 experts), a Kimi Delta Attention architecture, native image and video input, always-on reasoning that cannot be switched off, and a 1M context window on both input and output. Where Z.ai is betting that more of the same model is enough, Moonshot is betting that the base itself was the ceiling.

The head-to-head, with provenance attached

The only place both models appear on the same page is Z.ai's own launch table, so that is where the side-by-side numbers come from — labeled as vendor-reported, not independently audited. Kimi K3's standalone figures line up with what Moonshot published in July and what Artificial Analysis measures, but no neutral lab has run both yet.

• Terminal-Bench 3.0 — G​LM-5.3 at 28.3 versus Kimi K3 at 17.4 (Z.ai's table). The single biggest coding gap in the matchup, on the newest and hardest version.

• Terminal-Bench 2.1 — G​LM-5.3 at 88.2 versus Kimi K3 at 88.3. A dead heat.

• DeepSWE v1.1 — G​LM-5.3 at 66.9 versus Kimi K3 at 67.5. Kimi by a hair.

• SWE-Marathon v1.1 — G​LM-5.3 at 42.5 versus Kimi K3 at 48.1. Kimi's best coding win.

• ExploitGym — G​LM-5.3 at 105/130 versus Kimi K3 at 36/70. A near-total sweep for G​LM.

• GDPval-AA v2 (knowledge work) — G​LM-5.3 at 1,769 versus Kimi K3 at 1,682.

• Access — G​LM-5.3: Coding Plan subscription, weights gated until roughly August 28. Kimi K3: API live at $3 / $15, weights downloadable since July 27.

The security moat is the real separation

Strip the coding benchmarks down and the decisive difference is cybersecurity. G​LM-5.3 reports 84.5 on CyberGym against Kimi K3's 80.0, and its ExploitBench score of 54.4 is nearly double Kimi K3's 32.2. On ExploitGym the gap is not a difference, it is a different category — 105 and 130 versus 36 and 70 across the 2-hour and 6-hour runs. That is why the two launches feel so different in practice: Kimi K3's weights are out in the open under a Modified MIT license, while G​LM-5.3's are being held back for safety hardening precisely because Z.ai says the model is good enough at finding and exploiting flaws that shipping the weights immediately felt irresponsible. If your work involves auditing someone else's code, or defending your own against automated vulnerability hunting, this is the single most relevant line in the entire comparison — and it is also the least independently verified one.

Where Kimi K3 fights back

Kimi K3's wins cluster where G​LM-5.3 is weakest: long-horizon repository work. It holds the edge on DeepSWE and SWE-Marathon, leads on Toolathlon Verified, and its 1M-token output window means it can write an entire codebase or a long agent trace in a single pass. Its always-on reasoning streams the full trace to the API, which developers have called far more useful for debugging agents than opaque summary explanations — with the operational catch that you must pass the reasoning fields back verbatim across tool calls, and there is no assistant prefill. On Artificial Analysis's Intelligence Index Kimi K3 sits at 57, fourth overall, with roughly $0.94 cost per task measured on its standard set. It is also the most expensive model a Chinese lab has shipped: $3 per million input tokens and $15 per million output, Sonnet-tier pricing, where G​LM-5.3 cannot yet be priced per token at all because it only exists inside a subscription plan.

The access and cost reality

Kimi K3 is on OrcaRouter right now at Moonshot's own list price — $3 / $15 per million tokens, $0.30 cached reads, 0% markup — so the 1M-context agent work is callable on the same key as the rest of your stack, and the open weights are the exit option if you want to self-host. G​LM-5.3 is the one that is hard to get: a $18-to-$168 monthly subscription with a points system that spends credits at 6.9x on input and 24x on output, off-peak pricing that halves the burn outside 14:00–18:00 China time, and no per-token API until the safety review clears. The routing layer actually flattens this asymmetry for the two-week window: when G​LM-5.3's API lands on the same key, a panel call can run both open-coding flagships against your own tasks and let the judge reconcile them — and if you cannot wait, Kimi K3's already-shipped weights make it the only one of the two you can take fully on-premises today.

Which bet is paying

The honest verdict after one week is that both bets are paying, but on different workloads. If your work is terminal-heavy, security-adjacent, and agent-orchestrated, G​LM-5.3 is the stronger direction on the evidence Z.ai has published — and the cybersecurity gap is big enough that it is worth waiting two weeks for the weights to check yourself. If your work is long repository sessions, multimodal input, or anything where you need the weights in hand today, Kimi K3 is the only one of the two that is actually available — and its deep-repo wins are the ones you can verify right now from its live API. The one thing to keep straight: every number in this comparison's head-to-head comes from Z.ai's own table, so treat the coding deltas as directional until an independent lab runs both. That is exactly the kind of verification the two-week wait will bring.