Open Weights vs Open Source: 112 Days Later, Which Model Actually Fits Your Mac?

Three Key Takeaways

1. This is 112-day-old news. On July 24, 2026, 25 organizations signed the open letter "Open Weights and American AI Leadership" addressing whether Washington should restrict open-weight models. Huang Jensen's first X post amplified it (68.5M views), but OpenAI, Anthropic, and Google did not sign.

2. "Open weights" ≠ "open source." Most models only release weights — not training data, code, or data pipelines. Licenses range from Apache 2.0/MIT to custom commercial clauses that can block commercial use.

3. The most "open" model is the one you can't run. Kimi K3.1 Pro's weights are 1.56 TB (requiring ~12 accelerators), while the largest model that fits a 24 GB Mac is Qwen3.8-27B Q4 at 16.1 GB.

Timeline: The Letter and the Tweet

July 16: Moonshot releases Kimi K3, sparking Washington debate on national-security risks of open weights (distillation concerns).

July 24: 25 entities sign the letter arguing against blanket restrictions, noting that concentrating advanced AI in a few closed models creates a "single point of failure." Huang Jensen tweets the letter at 9:18 PM, generating 68.5M views, 45K reposts, 188K likes.

Metric clarification: 68.5M = views; 689K = follower count at the time.

Who Signed — and Who Didn't

Signatories: Chip vendors (NVIDIA, Dell), model makers (Meta, Mistral, Black Forest Labs), enterprise software/security (Microsoft, IBM, Palantir, CrowdStrike, ServiceNow), open-source orgs (Hugging Face, Mozilla, Linux Foundation), investors (a16z, Y Combinator, Emergence Capital).

Non-signatories: OpenAI, Anthropic, Google, Amazon — highlighting a business-model divide.

NVIDIA's incentive is transparent: more open-weight models → more training, fine-tuning, and inference demand for its hardware.

Open Weights vs. Open Source: The Licensing Trap

Open weights = trained parameters downloadable, runnable on own infra, usually allowing inspection, modification, fine-tuning, redistribution.

Open source (software sense) = training data, training code, data-processing pipelines reproducible, plus an OSI-approved license.

Many "open-source models" are strictly open-weight only. The license determines commercial viability:

Apache 2.0 / MIT: Qwen3.8-27B, DeepSeek-V4.1-Flash-1017, Ripple-65B, gpt-oss-120b — easiest commercial reuse.

Custom licenses: Mistral Large 4 (proprietary); Kimi K3.1 Pro (MIT text but adds clause: if MaaS revenue > $20M in any rolling 12 months, separate agreement required).

Promises without weights: Qwen4/Qwen5 — announced at 2026 Cloud Summit (460+ models, 3B+ downloads, 300K+ derivatives), but no license, no Max/Plus/Flash parameters, no timeline, no commitment that every tier will release weights. Max tier has been API-only since Sept 2025; self-hosted ceiling remains the 27B line.

Reminder: "Downloadable" ≠ "commercially usable." Check each Hugging Face repo's LICENSE; license: other demands reading every addendum.

Tim O'Reilly replied: open weights are good but insufficient — need "architecture of participation" (modular design, open protocols), a phrase he coined at OSCON 2003. Denny's Diner posted a Venn diagram: red circle Denny's (24/7), green circle NVIDIA (models stay open), intersection: "Knows the importance of staying open."

112 Days Later: Model Sizes Have Exploded

An observer noted: "For two straight weeks the top open-weight score belongs to a model ordinary people can't download." The table below combines size (determines if it fits), license (determines commercial use), and benchmark score (unrelated to the first two).

Kimi K3.1 Pro (Oct 24 scores, weights promised by Nov 4): 2.8T total / 104B active / 1M context, MXFP4 quantized to 1561 GB . At ~141 GB per accelerator, minimum 12 cards just to load weights (excl. KV cache). A 512 GB Mac Studio falls ~3× short.

Mistral Large 4 (Oct 29): ~1.05T total / 49B active / 1M context, custom license. Community 1-bit UD-IQ1_S ~ 233 GB — only tier reaching a 256 GB Mac, but 1-bit is lossy and requires manual iogpu.wired_limit_mb tuning.

Qwen3.8-Flash-Next 4-bit 106–111 GB; DeepSeek-V4.1-Flash-1017 (MIT) ~92 GB resident / 296 GB download (196B Engram offloaded to SSD). Both need 128 GB Mac.

Ripple-65B (Nov 4): 65B dense, Apache 2.0, 4-bit MLX ~ 37 GB — runs on 48 GB but requires --trust-remote-code.

Most practical: Qwen3.8-27B Q4 — 16.1 GB , Apache 2.0, runs on 24 GB Mac, richest tooling ecosystem.

Pitfall: Don't estimate size from model name. Qwen3.8-Flash-Next "125B" carries ~55B extra lookup params → 106–111 GB actual. Kimi K3.1 Pro active params 104B, but total params dictate download size, active params dictate compute .

Sizes sourced from HF repos & community aggregates (Oct–Nov 2026), non-uniform benchmarks; quant variants differ widely. "Runnable" means weights resident only — add ~20% headroom for KV cache/context. Kimi weight upload date varies (Nov 2/4/5); article uses "early Nov" per official "by Nov 4" promise.

Which Tier for Your Mac?

Unified memory → viable open-weight tier:

24 GB → Qwen3.8-27B Q4 (16.1 GB)

48 GB → Ripple-65B 4-bit (37 GB)

128 GB → Qwen3.8-Flash-Next / DeepSeek-V4.1-Flash (~100 GB)

256 GB → Mistral Large 4 1-bit (233 GB, quality loss)

512 GB+ → still short of Kimi K3.1 Pro (1.56 TB)

Bottom line: The gatekeepers aren't Washington or Huang's tweet — they're your unified memory and the repo's LICENSE file. Get the 16.1 GB tier running smoothly before chasing the 1.56 TB leaderboard topper; for "promised" flagships, wait until weights actually land in the repo before adjusting procurement.

Data Sources

Primary (tweet & letter): Huang @JensenHuang first post, 9:18 PM Jul 24 2026, 68.5M views (screenshot verified Nov 13 2026); open letter PDF dated July 24 2026 (images.nvidia.com/pdf/Open-Weigh...).

Signatories & context: 25-org list, non-signatories, July 16 Kimi K3 regulatory debate, distillation controversy, "single point of failure" argument, Nadella public support — CCTV (Jul 26), China Daily (Jul 25), The Paper/Tencent News (Jul 25), Open Source For U, YourPAI full list. Note: some English reports incorrectly list OpenAI as signatory; article follows the official list.

Tim O'Reilly: Reply screenshot Jul 25; "architecture of participation" from 2003 OSCON (oreilly.com). Denny's Venn diagram screenshot Jul 26.

Kimi K3.1 Pro: 2.8T/104B/1M/MXFP4/Nov 4 weights — kimi.ai blog; 1561 GB, 16-bit ~5.6 TB, 12-card estimate, custom commercial clause — 21st Century Business Herald (Nov 5), AILog, aimodeling, frontiernow.dev.

Mistral Large 4 / license tiers: Oct 29 HF release, 1.05T/49B/1M, custom license, Apache 2.0/MIT roster, UD-IQ1_S 233 GB + iogpu.wired_limit_mb — BaaO Intelligence "Did Open Models Regress?" (Nov 1), OrcaRouter CN compare, frontiernow.dev.

Qwen open pledge (third-party research): 460+/3B+/300K stats, Sept 22 Cloud Summit, Qwen4 27B first open, no license/timeline/"every tier" promise, Max API-only since Sept 2025, "self-host ceiling = 27B" — Shanghai NYU RITS "Qwen Restates Its Open-Weight Promise, Max Tier Still Closed"; Nov 4 reaffirmation — OrcaRouter; Qwen4-Max preview no weights — gpts24.

Model sizes (HF & community, Oct–Nov 2026): Qwen3.8-27B Q4 16.1 GB, Ripple-65B 4-bit MLX ~37 GB, Qwen3.8-Flash-Next 4-bit 106–111 GB (incl. ~55B lookup), DeepSeek-V4.1-Flash-1017 ~92 GB resident / 296 GB download (196B Engram on SSD). All non-uniform baselines; margins noted in text.