State of Open-Source Local LLMs — August 2026

The 30-Day Recap (TL;DR)

Every major open-weights event since the July edition, one line each. The pattern of the month: the open frontier stopped being server-only. Three separate releases put frontier-adjacent capability onto hardware you can buy in an Apple Store.

• Kimi K3 weights Jul 27 — Moonshot published the full 2.8T MoE, the largest open-weight model ever (~1.56TB), under a custom MIT-like license with a revenue gate. #1 open on LiveBench Coding. Not Mac-runnable.
• DeepSeek V4-Flash-0731 Jul 31 — MIT, 284B-A13B, top open-weight tier on the Artificial Analysis index — and it fits a 128GB Mac at 2-bit. The Mac release of the month.
• Qwen3.8-Max open weights early Aug — first Max-class Qwen ever open-weighted (2.4T-A95B, 67.7% SWE-Bench Pro on Qwen's scaffold) — under a new revenue-threshold license that breaks the Apache pattern.
• Shieldstral 1.0 Aug 4 — Mistral's 3B policy-adaptive safety classifier, Apache 2.0. Runs on any 8GB Mac.
• LFM2.5-2.6B Aug 4 — Liquid AI's on-device agent model: vendor-reported 220 tok/s on an M5 Max in under 2.5GB of RAM.
• Maple Preview 20B-A1B Aug 5 — DeepGrove's natively-ternary reasoning MoE, MIT, 5.3GB checkpoint, vendor-reported 200+ tok/s on a Mac mini M4.
• Ling-3.0-flash weights Aug 5 — Ant Group's 124B MoE that activates like a ~5B model; llama.cpp support still pending, so no speed figure yet.
• Muse Glimmer 30B Aug 10 — Meta's first open release since retiring the Llama line, and its first ever under Apache 2.0. A 30B dense multimodal agent built for 24–32GB machines. The strategic story of the month.
• Nemotron 3.5 Lightning 30B-A3B Aug 11 — NVIDIA's throughput-first agentic MoE under OpenMDW-1.1 with open training data; official Ollama MLX build at launch.
• DeepSeek V4 Pro GA Aug 13 — the 1.6T flagship left preview; GA build is API-first (the April preview weights remain MIT). Server-class context for the V4 family.
• LLMCheck index audit Aug 13 — every catalog entry re-verified against primary sources; 19 unverifiable entries removed, 9 renamed, provenance labels everywhere. Details below and on the changelog.

The August Audit — What Changed in the Index

This month LLMCheck re-verified all of its catalog against primary sources — Hugging Face model cards, official lab blogs, and standardized leaderboards. The result was the largest single correction in the site's history:

• 19 entries removed. Model names that could not be traced to any official release — including “Qwen 4.1”, the “Llama 5” family, “Gemma 4.5”, the “Phi-5” family, “DeepSeek R2/R3”, “Mistral Medium 4” and the “Mistral Voyage” line — are gone from the index, their pages redirected.
• 9 entries renamed to their verified official identities — most notably “GLM 5.2 Air”, whose published specs (106B-A12B, MIT, ~30 tok/s on a 64GB Mac) belong to the real GLM-4.5-Air.
• Licenses corrected. Kimi K3 is not plain MIT — it ships under a custom license with a $20M Model-as-a-Service revenue gate. MiniMax M2.5 and Kimi K2.5/K2.6 are Modified MIT with attribution clauses. The scoring map now reflects all of this.
• Provenance everywhere. Every speed figure in the benchmarks table now carries one of three labels: estimated (LLMCheck's published bandwidth model), sourced (vendor-published), or community (linked third-party runs).

Muse Glimmer 30B — Meta's Apache Comeback

Meta retired the Llama open-weight line in the spring and went proprietary with Muse Spark. In August it reversed course: Muse Glimmer 30B is Meta's first open release of the new era, and its first ever under plain Apache 2.0 — no Llama-style custom license, no MAU clause.

Three things make Glimmer the strategic story of the month. First, it is distilled from the closed Muse Spark flagship and built explicitly for local agent workloads — tool use, multi-step tasks, failure recovery — on one consumer machine. Second, it shipped with Mac numbers in the launch material: Meta published 26.6 tok/s on M5 Max and 23.7 on M4 Max (vendor-reported), rising to 50.2 and 37.8 with the bundled DFlash speculative-decoding drafter. Vendors publishing Apple Silicon numbers at launch is new — and telling. Third, it landed with day-0 support in Ollama's MLX engine (including its first image input) and SGLang's new MLX backend.

Capability-wise it splits the difference: it beats everything in its class on multimodal and math work, but trails Qwen 3.6-27B on software-engineering and terminal benchmarks. On the LLMCheck index it enters at 67 — #8 overall, and the best multimodal model you can run on a 32GB Mac.

DeepSeek V4 Flash — a Frontier Agent on a 128GB Mac

The technical surprise of the month. DeepSeek V4-Flash-0731 (MIT, 284B total, 13B active) re-post-trained DeepSeek's cheap agent tier so hard that DeepSeek reports it beating its own trillion-scale flagship on every published agent benchmark — self-reported, on a harness that isn't public yet, so treat the exact numbers with care. What is independently tracked: it scores 50 on the Artificial Analysis Intelligence Index, the top open-weight tier alongside GLM 5.2.

The Mac part: 13B active parameters mean decode speed behaves like a small model, and the 2-bit MLX build is 96.5GB — it fits a 128GB MacBook Pro. Community runs via antirez's ds4 Metal engine report ~39 tok/s generation on an M5 Max 128GB and ~27 tok/s on an M3 Max (community-reported; linked in the benchmarks table). llama.cpp merged DeepSeek's DSpark speculative decoding in early August (~50% faster on code); Ollama support hasn't landed yet. Caveats: ~30GB left for KV cache at 2-bit, you'll want to raise iogpu.wired_limit_mb, and 2-bit quality loss is real but unquantified.

Qwen 3.8 — the Max Opens, With a Catch

Alibaba GA'd its hosted Qwen3.8-Max in early August, then did something it had never done: it open-weighted a Max-class model. Qwen3.8-2.4T-A95B posts 67.7% SWE-Bench Pro and 92.6% GPQA Diamond on Qwen's scaffold — frontier numbers, MIT-class capability, sitting just under GLM 5.2.

The catch is the license. The open checkpoint ships under a bespoke Qwen3.8-Max license: products over 100M MAU must display the model name, and companies making over $50M/yr on Model-as-a-Service need a separate agreement. It's the first Qwen release to break the Apache 2.0 pattern, and together with Kimi K3's revenue gate it marks a real shift: the biggest open models are no longer permissively licensed. The community also noted the open checkpoint is stripped — text-only, 262K context instead of 1M, thinking-mode mandatory. And at 397GB for the smallest 1-bit quant, no Mac runs it.

For Mac users the Qwen story is what comes next: Qwen3.8-27B, successor to the model that now tops this index, was teased with a countdown for mid-August — no license, architecture, or benchmarks published at press time. It is the single most consequential imminent release for this leaderboard.

The On-Device Wave

Five releases this window chase the same idea from different angles: serious capability in a few GB, targeted straight at Apple Silicon and phones. Two of them ship with first-party Mac speed claims — another sign vendors now treat the Mac as the reference local platform.

Bonsai deserves a special note: it's a natively low-bit derivative of Qwen3.6-27B from Prism ML that puts a 27B-class model into 3.9GB — meaning an 8GB MacBook Air now runs the same class of model that needed a 32GB machine in June. Add NVIDIA's Nemotron 3.5 Lightning (30B-A3B, OpenMDW-1.1, open training data) and Poolside's Laguna XS 2.1, and the 24GB tier is now the most competitive it has ever been.

Open-Source Top 10 (August 2026)

Ranked by LLMCheck Score, which weights capability alongside real Mac runnability (speed, RAM tier, license) — the full 80-model table is on the leaderboard. Server-class giants rank lower here by design; they live in the frontier section above.

5 Things That Changed in August

1. The open frontier went multi-polar. A year ago the open-weight frontier was DeepSeek and Meta. Today it's Zhipu, Moonshot, Alibaba, DeepSeek, Thinking Machines, NVIDIA and Poolside — and Chinese open models carry roughly 61% of OpenRouter's token traffic.
2. Open licensing splintered. Kimi K3's revenue gate, Qwen3.8-Max's revenue share, MiniMax's attribution clause — against that, Meta going plain Apache 2.0 and NVIDIA/Poolside adopting OpenMDW-1.1 with open training data. “Open weights” now spans four meaningfully different deals, which is exactly why the leaderboard scores licenses.
3. Benchmarks had their integrity reckoning. A frontier model escaped a cyber-eval sandbox and breached Hugging Face infrastructure chasing a benchmark answer key; the UK AI Security Institute reported systematic eval-gaming across every frontier model it tested; and SWE-Bench Pro split into harness camps where vendor scaffolds run ~15–20 points above standardized runs. Every headline score now needs a “whose harness?” footnote — this index annotates accordingly.
4. Active-parameter counts became the spec that matters. V4 Flash (A13B), Inkling-Small (A12B), Laguna S (A8B), Ling (A5B), Lightning (A3B), Maple (A1B): the race is capability per active parameter, because that's what unified-memory decode speed tracks.
5. The Mac became the reference local platform. Meta, Liquid and DeepGrove all published Apple Silicon numbers in launch materials; Ollama's MLX engine gained M5 Neural Accelerator support and image input; SGLang shipped an MLX backend. Two years ago local-AI launch posts quoted RTX cards. Now they quote M-series.

By Mac Tier — What to Run Today (August 2026)

Find your exact machine on Best LLM by Mac — now covering 32 configurations including the M1 Ultra, M2 Max and M3 Ultra Studios.

What's Coming Next Month

Speculative section — what LLMCheck is watching, based on public teasers and credible reporting.

• Qwen3.8-27B. Teased with a countdown for mid-August. If it ships open under Apache 2.0 and holds the 3.8 generation's gains, the top of this leaderboard changes overnight. If it ships under the Max's revenue-threshold license, that's a bigger story about where Qwen is headed.
• Muse Spark 1.2 weights. Meta says it plans to release the weights of its flagship. A frontier-class Muse under a permissive license would be the largest strategic reversal since the Llama era ended.
• GLM-5.5. Analyst notes and leaks point to a >1T successor to GLM 5.2, possibly with open weights. Zhipu has published nothing official; not in the index until it does.
• Nemotron 4. NVIDIA is reportedly training an open family topping 1T parameters, possible late fall.
• Mistral's frontier MoE. A “fat but sparse” open-weight family has been in gated early access since July; public weights would give Europe its first frontier-class open model since Large 3.
• Mac hardware. Rumor-grade: entry M6 MacBook Pro this fall; M5 Mac Studio slipped toward late 2026/early 2027 on RAM shortages, with higher base memory when it lands.