Quick Verdict
If you want the short answer first, here it is. The published numbers — and the caveats around them — follow below.
The frontier pick: Inkling
Apache 2.0, 975B-A41B, multimodal, and 77.6 on SWE-bench Verified — the cleanest license and the strongest verified software-engineering score of the three. If you consume frontier open models via hosted APIs or your own cluster, Inkling is the release with the fewest strings attached.
The Mac pick: run the siblings
None of the three fits any Mac ever made. Run Inkling-Small on a 128GB+ Mac Studio (80.2% SWE-bench Verified — the open-weight record), Qwen3.6-27B on a mainstream Mac (the LLMCheck Mac #1), or GLM-4.5-Air on a 64GB machine.
The Three Contenders
Qwen3.8 — Alibaba's hosted Max and its 2.4T open weights
Qwen3.8-Max went GA as a hosted model in August 2026, followed within days by the open-weight Qwen3.8-2.4T-A95B: 2.4 trillion total parameters, 95B active per token. It posts 67.7 on SWE-Bench Pro and 92.6 on GPQA — top-tier open numbers — but it is also the first Qwen ever released under a non-Apache license, a custom scheme with a revenue threshold. The smallest published quantization is roughly 397GB. Full background in our Qwen 3.8: what actually shipped writeup.
Kimi K3 — Moonshot's 2.8T record-setter
Kimi K3 launched in July 2026, with weights published later that month. At 2.8 trillion parameters (MoE) and a checkpoint of roughly 1.56TB, it is the largest open-weight model ever released. It ranks #1 among open models on LiveBench Coding and #3 overall on the Artificial Analysis Intelligence Index. The license is a custom "Kimi K3 License" — MIT-like, with a gate that adds obligations above $20M in annual model-as-a-service revenue.
Inkling — Thinking Machines Lab's Apache 2.0 debut
Inkling arrived in July 2026 from Thinking Machines Lab: a 975B-parameter multimodal MoE with 41B active, scoring 77.6 on SWE-bench Verified and 41 on the Artificial Analysis Intelligence Index — under plain Apache 2.0, the cleanest license at this scale. It shipped alongside a smaller sibling, Inkling-Small, which turns out to be the one Mac users should care about (more below).
Benchmark Head-to-Head
A caveat before the table: at launch, each vendor published a different benchmark suite, so coverage is patchy and the three models rarely meet on the same test. Treat cross-suite comparisons as provisional until third-party harnesses catch up.
| Metric | Qwen3.8-2.4T-A95B | Kimi K3 | Inkling |
|---|---|---|---|
| Developer | Alibaba | Moonshot AI | Thinking Machines Lab |
| Weights released | August 2026 | July 2026 | July 2026 |
| Architecture | MoE, 2.4T total / 95B active | MoE, 2.8T total | MoE, 975B total / 41B active |
| Multimodal | — | — | Yes |
| License | Custom, revenue-threshold | Kimi K3 License ($20M MaaS gate) | Apache 2.0 |
| SWE-Bench Pro | 67.7 | — | — |
| SWE-bench Verified | — | — | 77.6 |
| GPQA | 92.6 | — | — |
| LiveBench Coding (open models) | — | #1 | — |
| Artificial Analysis Index | — | #3 overall | 41 |
| Smallest published download | ~397GB (quantized) | ~1.56TB | — |
| Runs on a Mac | No | No | No |
Reading across the gaps: Kimi K3 owns coding leaderboards, Qwen3.8 owns knowledge-heavy suites like GPQA, and Inkling posts the best verified software-engineering score while being the only multimodal entry and the only Apache 2.0 entry. There is no single winner on capability — which is exactly why the license column ends up being the most decisive row in the table.
License Showdown
Two of the three arrived with custom licenses, and that is the real story of this generation of open releases.
Inkling: Apache 2.0. No revenue gate, no naming clause, no field-of-use restriction. At 975B parameters, it is the largest model ever released under a fully permissive license, and it sets the bar the other two are measured against.
Kimi K3: MIT-like with a $20M gate. For individuals, researchers, and almost every company, the Kimi K3 License reads like MIT. The exception is model-as-a-service providers: above $20M in annual revenue built on the weights, additional obligations kick in. It is a narrow gate aimed at hyperscalers, but it is not MIT.
Qwen3.8: the surprise. Qwen built its reputation on clean Apache 2.0 releases — Qwen 3.5 and 3.6 carry no strings at all. Qwen3.8-2.4T-A95B breaking that pattern with a revenue-threshold license drew the loudest community reaction of the summer. Anyone building a commercial product on Qwen weights should note that the Qwen3.6 line remains fully Apache 2.0.
Why None of These Runs on a Mac
The rule that governs local inference has not changed: active parameters drive speed, total parameters drive RAM. A sparse MoE only computes a fraction of its weights per token, but every expert must still be resident in memory. That is fatal at this scale:
- Kimi K3 — ~1.56TB of weights. Triple the memory of a maxed-out 512GB Mac Studio before you even load a context window.
- Qwen3.8-2.4T-A95B — smallest published quantization ~397GB. Technically within sight of a 512GB machine, practically not: context, KV cache, and the OS need room, and throughput at that size would not be usable.
- Inkling — 975B parameters with no Mac-targeted quantizations published. Thinking Machines pointed local users at Inkling-Small instead.
This is the widest the gap between "open weights" and "runnable weights" has ever been — and it is why the sibling models below matter more to Mac users than the flagships themselves.
The Mac-Class Siblings
Inkling-Small — the frontier-closest Mac model
Inkling-Small (276B-A12B, Apache 2.0) is the sleeper hit of the summer. Its 80.2% on SWE-bench Verified is the open-weight record — ahead of its own 975B parent — with GPQA at 89.5. The 2-bit MLX build is 88.4GB, fitting a 128GB Mac Studio; a GGUF IQ4 at ~130GB fits 192GB machines. The catch: it is MLX-only for now, with llama.cpp support pending. Details on the Inkling-Small on M5 Max page.
Qwen3.6-27B — the mainstream Mac #1
While Qwen3.8 stays server-class, Qwen3.6-27B (dense, Apache 2.0) remains the LLMCheck Mac #1 with a Score of 72 — 77.2% on SWE-bench Verified and an estimated ~40 tok/s on an M5 Max. It is the strongest model you can run on a mainstream 24–32GB Mac. Full breakdown on the Qwen3.6-27B on M5 Max page.
GLM-4.5-Air — the 64GB workhorse
GLM-4.5-Air (106B-A12B, MIT) brings big-model breadth to 64GB machines: according to the LLMCheck index, an estimated ~30 tok/s on a 64GB Mac at 4-bit. Its 12B-active MoE layout is the same design philosophy as Inkling-Small at a friendlier memory footprint. See GLM-4.5-Air on M5 Max.
Honorable mention: the other frontier-adjacent release that actually reaches Macs is DeepSeek V4-Flash (284B-A13B, MIT) — its 2-bit MLX build is 96.5GB, with community-reported ~39 tok/s on an M5 Max. It gets its own coverage on the DeepSeek V4-Flash on M5 Max page.
The Verdict
On the frontier itself, Inkling is the release that matters most. It does not top every leaderboard — Kimi K3 owns LiveBench Coding and Qwen3.8 owns GPQA — but 77.6 on SWE-bench Verified under a genuine Apache 2.0 license, with multimodality included, makes it the open flagship with the fewest asterisks. Kimi K3 is a landmark of scale whose 1.56TB footprint confines it to serious server clusters, and Qwen3.8 pairs excellent numbers with the license controversy of the summer.
On a Mac, the frontier arrives through the side door. Inkling-Small holds the open-weight SWE-bench Verified record and fits a 128GB Mac Studio; Qwen3.6-27B remains the best model for mainstream machines; GLM-4.5-Air covers the 64GB tier. According to the LLMCheck index, that trio — not the trillion-parameter flagships — is what August 2026 actually changed for local AI on Apple Silicon.