Zero to MVP Weekly
Free · every Saturday · ~5 min
← All issues

Open weights, with a bill attached

Issue #10 · August 15, 2026 · ~5 min read

Welcome back to Zero to MVP Weekly. Same deal: one signal, one tool, one MVP idea — five minutes, every Saturday. No link dumps. This week "open weights" stopped being one thing: a 2.4-trillion-parameter model arrived with a license that bills at scale, and a coding SOTA arrived with its weights held back for safety. Let's go.

1. The signal of the week

On August 12 Alibaba released the weights of its biggest model ever — and the license, not the benchmarks, is the announcement.

Qwen3.8-2.4T-A95B landed on Hugging Face: the base model behind the Qwen3.8-Max API, and the first time a Max-class Qwen model has shipped open. The scale is real — a 2.4-trillion-parameter MoE with 95B active parameters, 262K native context (extensible to ~1M), and self-reported benchmarks at the frontier: SWE-bench Pro 67.7, Terminal Bench 2.1 86.6, GPQA Diamond 92.6.

Two prints matter more than the benchmarks. First, the license is not Apache 2.0. It's a bespoke "Qwen3.8-max" license — an MIT-style grant, plus mandatory attribution above 100M monthly active users or $20M monthly revenue, plus a separate paid commercial license for model-as-a-service or AI-work-assistant businesses above $50M trailing-twelve-month revenue, affiliates included. Internal use is carved out. Qwen 3.5 and 3.6 were plain Apache-2.0; this is the same vendor deciding that frontier-class weights ship under commercial thresholds.

Second, the open repo is not the API product. The released weights are text-only, thinking-mode-only, 262K native. The managed Qwen3.8-Max API adds vision input, non-thinking mode, 1M context by default, and built-in tools. Same family name, different things.

Then, two days later, the story got its twist. On August 14 the promised companion landed: Qwen3.8-27B — the size you can actually run — under plain Apache 2.0. A dense 27B with native image and video understanding, the same 262K context (extensible to ~1M), adjustable reasoning effort, and 61.7 on SWE-bench Pro, already on Hugging Face with hundreds of quantized variants for llama.cpp, Ollama and LM Studio. Note the inversion: the small open model is multimodal; the big open one is text-only.

Why this leads the issue: one vendor, one week, two licenses. The frontier-class weights — the ones inference providers and platforms will monetize — ship with revenue thresholds. The 27B — the one that lands on your GPU — ships genuinely open. That's the clearest picture yet of what "open weights" is becoming: a spectrum, priced by who you are. Meta's Muse Glimmer (Apache 2.0, August 10) and Z.ai's GLM-5.3 (weights in a two-week safety hold while its unexpectedly strong cyber capabilities get hardened) bracket the same week. The LICENSE file is now part of the model card you actually have to read: pick a model, and you're also picking a legal position.

One caveat to keep us honest: every number here is self-reported, and the 27B is hours old — no third-party benchmarks, no field reports. The headline scores (67.7 SWE-bench Pro) belong to the 2.4T you won't self-host; the model you will run scores 61.7 on the vendor's own table. The real answer arrives after a week of local testing, not from launch-day charts.

2. My take

I won't pretend the licensing homework makes me happy — picking a model now also means reading legal text, and that's one more thing on the checklist. But I think it's the unavoidable price of these models moving into more and more parts of real life and real work: the more value they carry, the more the terms around them matter.

And yes — I've already started testing Qwen3.8-27B. A video about it is coming to the channel soon.

3. Tool worth trying

Claude Code's auto mode — because on August 14 it became your default whether you tried it or not.

Anthropic flipped auto mode on by default in Claude Code for Pro, Max and Team plans (new sessions from August 14). Instead of asking permission per command, a classifier now evaluates each tool call and blocks what looks irreversible, destructive, or aimed outside your environment — the rest proceeds without interruption. The mechanics worth knowing: overly broad allow-rules like python:* are set aside in auto mode so they can't bypass the classifier; after 3 consecutive or 20 total blocks a session falls back to manual approval; Shift+Tab switches modes any time; admins can pin defaultMode or disable auto mode entirely. Enterprise and API deployments stay opt-in.

The scenario to try is a 30-minute audit, then a normal week. Check which mode you're actually in. Re-read your allow/deny rules knowing the broad ones are now ignored. Then work normally and keep a note of every block and every silent pass — the interesting list is the things it let through that you'd have declined, and vice versa.

In a poll I ran on my channel, more than half of the respondents were already running with permissions skipped — so I expect this mode to catch on fast. I've started trying it myself and I'm still testing, but on separate machines (a Raspberry Pi and my ZimaCube), where the blast radius is minimal if something goes wrong. What I like is that it's not just one checkbox: there's a dedicated settings file where you can tune the rules until they fit the way you work.

The limits, stated plainly: the classifier is a black box you're now trusting by default, "aimed outside your environment" is exactly the judgment call that's hard to audit, and a default changing under daily-driver users — however good the guardrails — is itself worth noticing. It fits this week's theme too: the vendor decided; your job is to read what changed.

4. MVP idea of the week

"Can I ship on this model?" — a license decoder for open weights.

  • Problem: "open weights" now spans Apache 2.0 (Glimmer), bespoke revenue-share (Qwen3.8-Max), delayed-for-safety (GLM-5.3), and "as licensing permits" (Nvidia's Nemotron data). A builder choosing a model can't answer "can my product legally use this?" without reading lawyer-grade text across a dozen repos.
  • Solution: a database plus a web quiz or CLI: pick a model, describe your use — SaaS? internal? reselling inference? revenue range? — and get a plain-language verdict with the exact license clauses quoted and linked. Show diffs when licenses change between releases.
  • For whom: indie hackers and small teams shipping on open models; platform and devrel teams doing model due diligence.
  • Minimal v1: hand-curated YAML for the top ~20 open models plus a static site with the quiz. No accounts, no LLM required.
  • Demand check: post the comparison table — "what 'open' means across 20 models, in one screen" — to HN or r/LocalLLaMA. Stars and "please add model X" issues are the vote.
  • Why now: the Qwen3.8-Max license landed August 12 as the first Max-class open release with commercial thresholds. The moment "read the license" became real advice is exactly the moment nobody wants to actually read them.

5. Useful links

I'm already testing the 27B; if GLM's weights land on the 28th, I'll read that license first — reply and tell me which one you're waiting for.
— Nick

P.S. Three of my own things, free and open on GitHub:

  • Texodus — a modern, fast, small and beautiful Markdown viewer and editor.
  • Prompt-Vault — coding prompts and detailed project specs for benchmarking and testing LLMs.
  • AI-Vault — my curated vault of practical AI-engineering resources: agentic coding, RAG, local models, evals.

Get the next one in your inbox.

One signal, one tool, one MVP idea — every Saturday. Free.