Week in AI: Meta's 816B open MoE, GPT-4 API shutdown Jan 2027, Accenture caps Claude at $29
ksraj1001 · reddit · 2026-10-21
Reddit user ksraj1001 rounds up the week's AI news (Oct 12-16), mostly second-hand via aggregators citing Reuters/Bloomberg/The Information.
Open weights
- Meta Muse Spark Open: MoE, 816B total / 42B active, 1M context; reportedly runs on 2-4 M5 Ultra Mac Studios via MLX.
- Nvidia Nemotron 3.5 Super VL: 120B/12B active, 1M context, claimed up to 5x throughput (vendor claim).
- Black Forest Labs FLUX 3 Image Klein: 5.6B, free commercial use under $5M revenue.
- Microsoft's MAI-Base-2 (2.8T/80B active) is not open; trusted partners via Foundry.
Cost / platform
- Accenture reportedly capped centrally funded Claude at $29/month/employee after some spent $500-$1,000.
- OpenAI: GPT-4o, 4o mini, 4.1, 4.1 mini leave the API Jan 14, 2027.
Policy
- FTC civil investigative demands to OpenAI and Anthropic over safety incidents; EU AI Office's first AI Act probes.
Builder's take: deprecation timelines cost real engineering time — six months is fine with evals, painful without; skeptical of Mac Studio claims until sustained long-context throughput is shown; open weights suit data-sensitive clients but total cost including ops needs a hard look.
More from Models
- NatWest benchmarks British-accent ASR: Cohere Transcribe hits 15.3% WER — Nils_Reimers · 2026-10-21
- Building the same library twice: open-weight model cost $19.86 vs Opus's $346.70 — robeybeswick · 2026-10-21
- Pangram AI detector flags 2016 human-written papers as 100% AI-generated — GaryMarcus · 2026-10-21
- User's riddle stumps ChatGPT, Claude and friends — screenshot inside — DoItYourselfMate · 2026-10-21
- Grok voice mode shines but drains the $30 plan; ChatGPT talk mode lags — DevDminGod · 2026-10-21
- Independent security tests of Gemini 4 Argon diverge from Google's launch table — Thirumalaivasan_GJ · 2026-10-21