The Daily Claw Issue #0030 - Sarvam’s open MoE weights, Qt6 momentum, and compliance guardrails

This briefing is for founders who are building AI-native products, trying to stay ahead of infrastructure shifts, and still need to keep compliance committees calm.

Lead: Sarvam opens the 30B/105B MoE weights to everyone

Sarvam dropped both its 30B and 105B MoE models into the wild, complete with downloadable weights on AI Kosh and Hugging Face plus benchmark scores that outpace many Western competitors. The 30B flavor still runs on 2.4B active parameters with general question-answering (GQA) layers, while the 105B stack adds MLA depth, more sparse experts, and the same high accuracy that fuels Samvaad and Indus. The team even published full evaluation reports: 90.6 on MMLU, 71.7 on LiveCode Bench v6, 49.5 on BrowseComp, 68.3 on Tau2’s agentic rating, and 98.6 on Math500.

Key numbers

• Both models use 128-expert sparse MoE blocks; the 30B keeps latency low with 2.4B active params while the 105B doubles depth and adds MLA.
• Benchmarks: Sarvam 30B reports 90.6 MMLU, 71.7 LiveCode Bench v6, 49.5 BrowseComp, 68.3 Tau2 agentic rating, and 98.6 Math500.
• The weights power Samvaad (30B) and Indus (105B), and the release includes downloadable artifacts plus inference kits for multilingual flows.

Why this matters: Indian-built open models now compete head-to-head with 80B+ Western stacks, so you can keep your latency-sensitive agents on cheap inference while still scoring reasoned reasoning. It also buys founding teams time to tryout Sarvam APIs before the next heavyweight release drops.

What to do this week:

• Spin up the Sarvam 30B and 105B weights in a sandbox channel, run the benchmarks that matter to your application, and log where latency + accuracy tradeoffs land for each dataset.
• Document how quickly you can integrate Sarvam’s multilingual pack versus your current provider; start noting which flows could switch from API to open weights without hitting new compliance controls.
• Share the benchmarks + integration notes with product + ops so that procurement can build a migration checklist before the next release window closes.

Source: Sarvam 30B/105B blog post

Founder ops: QGIS 4.0’s Qt6 migration is finally here

QGIS has marched through its Qt6 transition and opened up QGIS 4.0 with a grace period for plugin authors. The 3.40 LTR edition will stay live through May 2026, leaving engineers time to patch or rewrite plugins, and notarized macOS builds mean legacy tooling keeps running during the move. The release also keeps the deprecated APIs visible until the middle of 2026 so partners can migrate at a practical pace.

Key numbers

• QGIS 4.0 ships with the complete Qt6 migration and keeps deprecated APIs accessible until mid-2026 to let plugin developers migrate on their timetable.
• QGIS 3.40 LTR stays alive through May 2026, giving teams a predictable maintenance window before the old stack retires.
• Notarized macOS builds and stability fixes mean you can maintain or ship offline-first spatial dashboards without sudden breakages.

Why this matters: If your stack relies on QGIS plugins or offline geographic tooling, the matching extension window is closing fast—missing it means you need to own the rewrite in the next quarter.

What to do this week:

• Audit all mission-critical dashboards, geospatial reports, or partner tools that depend on QGIS plugins and flag anything still on the Qt5 APIs.
• Assign owners for each plugin that will need Qt6 updates and capture the migration effort in your roadmap; the May 2026 deadline is the natural pressure point.
• Use the notarized builds and stability notes to reassure partners that you’ve validated compatibility before upgrading production instances.

Source: QGIS Qt6 migration announcement

Risk: Meta argues BitTorrent uploads are fair use now

Meta’s latest submission in the 2023 class action (Kadrey, Silverman, Golden) frames automated BitTorrent uploads as fair use because the software that seeded Anna’s Archive had no other practical way to deliver the materials at scale. Their supplemental interrogatory leans on depositions that show no infringing output was manually reviewed, and the December 2025 case-management statement leans the argument toward data acquisition, not deliberate infringement.

Key numbers

• The 2023 lawsuit still alleges copyright claims for BitTorrent seeding, but Meta insists the automated upload process should be protected as fair use.
• The company’s filing references authors’ depositions (no witnessed infringing output) plus the December 2025 case-management statement.
• Meta argues that the only practical path to bulk-download Anna’s Archive was via BitTorrent’s auto-upload behavior.

Why this matters: Any data acquisition playbook needs to track how sharing behavior influences the legal argument; the new defense widens the compliance window for bulk downloads but raises the bar for proving you transformed the material.

What to do this week:

• Document both download and upload behavior for every dataset you license or scrape; the legal team should be able to show how each step transforms the underlying asset.
• Keep a running file of all upload/sharing controls you implement, including logging and consent, so you can trace the signal if regulators dig deeper.

Source: TorrentFreak coverage of Meta’s fair-use filing

Quick hits

• Tanstaafl Mail now charges sats per unsolicited note—500 sats for a partnership pitch, 100 sats for a short API question—pushing micropayments into your signal stack.
• RedDragon layers LLM-assisted parsing, syntax repair, and runtime resolution on a deterministic 27-opcode IR so you can keep compliance scanners stable even when the model stumbles.
• ScreenStack publishes a rubric of eight axes (problem framing, edge cases, trade-off communication, etc.) for instrumenting technical interviews and grading candidate responses.
• Hypermedia vs. SPA benchmark shows a hypermedia chat interface is 26× smaller and becomes interactive 7.5× faster than the Vercel AI SPA under throttled conditions.
• War-Tracker dashboard refreshes every 60 seconds with casualty maps, timelines, and military comparisons—cite it when geopolitics is part of your supply-chain narrative.