The benchmark gap between open and closed models has effectively closed on knowledge tasks. That is the headline, and it is true. It is also close to useless as buying advice, because the models that closed the gap are mostly not runnable by an individual, and several of the published checkpoints are not the models that produced the benchmark scores.
This is the field guide version: what actually shipped this summer, what it takes to run, and what license you get.
August’s releases, in order
Qwen3.8-Max — weights published 12 August 2026. Alibaba’s flagship, a mixture-of-experts model at roughly 2.4 trillion total parameters with about 95 billion active, published in BF16 safetensors and fine-grained FP8, documented as compatible with vLLM and SGLang. It is the first open Max-tier Qwen.
Three caveats reported on the checkpoint, and they matter more than the parameter count: it is text-only, without the vision capability of the hosted model; its usable context is well below the 1M tokens advertised for the API; and it carries a bespoke license including revenue-share terms. A smaller Qwen3.8-27B — the size class most individuals would actually deploy — was reported as delayed.
DeepSeek V4 Pro — general availability, build 0813, on 13 August 2026, with revised pricing effective 16 August. V4 has posted the strongest raw benchmark claims of the open cohort, including figures in the low-to-mid 80s on SWE-bench Verified. Treat vendor-reported coding benchmarks with the usual scepticism; the direction is not in doubt even if the decimal places are.
GLM-5.3 — 14 August 2026. Zhipu AI’s Z.ai shipped it as the same base model as GLM-5.2 with scaled-up post-training. GLM-5.2 launched 13 June 2026 under an MIT license with a one-million-token context window, and that MIT licensing makes the GLM line the most genuinely permissive member of the frontier-adjacent open cohort.
Kimi K3 — API 16 July, open weights late July 2026. Moonshot AI’s 2.8-trillion-parameter MoE with 896 experts, 16 active per token, 1M context, distributed natively in MXFP4 with expert weights in 4-bit and attention layers in BF16. Download size: about 1.56 TB.
What it takes to run them
This is the section the release coverage skips.
Kimi K3 is the clearest illustration. Official vLLM metadata estimates a minimum VRAM footprint around 1,680 GB, with a deployment recipe calling for at least eight GB300-class GPUs. Even at Q4 the estimate is roughly 350 GB, and BF16 is around 594 GB. The MoE architecture does not help here: only 16 of 896 experts are active per token, but all 896 must be resident in memory.
Kimi K3 is not a local model in any sense an individual would recognise. It is a cluster model. “Running it locally” means owning an enterprise GPU cluster.
The same logic applies, with smaller numbers, to Qwen3.8-Max at 2.4T total parameters and to DeepSeek’s flagship. Sparse activation reduces compute per token; it does not reduce the memory needed to hold the weights.
So the honest tiering for a personal assistant is:
Tier 1 — genuinely local, consumer hardware. 7B to 14B class models, 4-bit quantized, running in 8–16 GB of RAM or VRAM. Qwen, Llama, Mistral, Gemma, and Phi all have credible entries. Good enough for summarization, drafting, extraction, classification, translation, and light code assistance. Not good enough for long-horizon agentic work.
Tier 2 — local on a serious workstation. 27B to 70B class, 4-bit, needing roughly 24–48 GB of VRAM or a high-memory Apple Silicon machine. This is where local becomes genuinely comparable to a hosted mid-tier model for most non-agentic tasks, and it is the sweet spot for a private assistant if you are willing to spend on hardware once.
Tier 3 — self-hosted on rented GPUs. The 100B–700B dense-equivalent range. You control the machine but you do not own it, which is a real privacy improvement over a consumer API and not the same thing as local.
Tier 4 — cluster only. Kimi K3, Qwen3.8-Max, DeepSeek V4 flagship. Open weights, closed practical access. Their release is strategically significant and personally irrelevant.
Most coverage conflates Tier 1 and Tier 4 under “open source AI.” They have almost nothing in common from a user’s standpoint.
Licenses, ranked by what they actually permit
MIT (GLM-5.2/5.3). Genuine open source. Use, modify, redistribute, commercialise, no revenue share, no acceptable-use rider. This is the real thing and it is rarer than the discourse implies.
Apache 2.0 (much of Mistral, some Qwen sizes). Equivalent freedom plus a patent grant. Also the real thing.
Bespoke community licenses (Qwen3.8-Max, Llama’s family, others). Free to use within conditions — user-count thresholds, revenue-share terms, acceptable-use policies, naming requirements, and restrictions on using outputs to train competing models. These are source-available with commercial strings, not open source, and the strings are the part nobody reads until legal asks.
For a personal assistant on your own machine, almost all of these permit what you want to do. For anything you build for other people, the license is the binding constraint and it varies enormously across models that get described identically in headlines.
What actually matters for a private assistant
Capability is the least interesting variable at this point. Four others matter more.
Does the checkpoint have the capability you need? The Qwen3.8-Max situation — text-only weights for a multimodal product — is now a recurring pattern. If your use case involves images, documents-as-images, or audio, verify modality on the actual checkpoint rather than the model card’s marketing section.
Real context length, measured. Advertised context and usable context diverge, sometimes dramatically, and the divergence is worse in released checkpoints than in hosted versions. If you intend to feed the assistant long documents, test with your actual documents.
Quantization degradation on your tasks. Q4 is roughly free for chat and noticeably lossy for structured extraction, code generation, and instruction-following under constraints. Test Q8 versus Q4 on your workload before committing to a hardware budget.
And tool-calling reliability. For an assistant that does anything — reads files, sends messages, runs commands — the binding constraint is not knowledge, it is whether the model emits valid structured tool calls consistently. This is the capability where the open cohort is furthest behind, and it is the one that determines whether a local assistant is useful or merely present. Agent frameworks covered separately here.
The privacy calculus, stated plainly
A local model is not automatically private and a hosted model is not automatically exposed, but the difference in failure modes is categorical.
With a local model, your prompts exist on your disk. There is no third-party record to subpoena, breach, retain past a policy change, or transfer in an acquisition. The failure mode is your own device security, which is a problem you can actually control.
With a hosted model, your prompts are records in someone else’s system. Enterprise terms may forbid training on them, and that is a different question from whether they exist, how long they are retained, who can access them internally, and what happens to them in litigation or insolvency. This site has covered therapy transcripts surfacing in court and customer data auctioned out of a bankruptcy estate. Neither required anyone to break a promise.
The practical rule is not “local for everything.” It is local for anything you would not want in a discoverable record — health, finances, legal exposure, employment, and above all other people’s personal information that you are handling on their behalf — and hosted for the rest.
What it means in practice
The open-weights ecosystem in August 2026 is two ecosystems wearing one name. There is a cluster-scale tier — Kimi K3 at 1.68 TB of VRAM, Qwen3.8-Max at 2.4 trillion parameters — whose releases are geopolitically significant and personally unusable. And there is a consumer tier of 7B-to-70B models that has quietly become good enough to be the default for private work, and that almost nobody markets because there is no announcement value in a 14B model that is merely adequate.
The second one is the one that changes your life. It is available today, it is free, it runs on hardware you already own or can buy once, and it produces no record anywhere.
What you can do
- Start at Tier 1 and move up only if it fails. Install Ollama or LM Studio, pull a 7B–14B model, and run it against a week of your real tasks. Most people discover the small model is sufficient for 80% of what they were sending to a cloud API.
- Check modality and real context on the checkpoint, not the product page. Assume the released weights are less capable than the hosted service until you have verified otherwise, because in August 2026 that assumption was correct more often than not.
- Prefer MIT and Apache models if you will build anything on top. GLM’s MIT licensing and Mistral’s Apache releases give you freedoms that bespoke community licenses do not, and the difference only becomes visible when it is expensive.
- Test quantization against your own workload. Q4 is fine for prose and unreliable for structure. The right quantization is task-dependent and takes twenty minutes to determine empirically.
- Route by sensitivity, not by capability. Write down which categories of your work go local — health, finance, legal, employment, other people’s data — and make that a rule rather than a judgment call you make while tired.
