2026-08-13

1. The first Max-class Qwen opens — two days late, and nobody has named the licence

• Alibaba published open weights for Qwen3.8-2.4T-A95B, the open-weight form of the 2.4T MoE flagship, and for the 27B, at a time stated five days in advance: 10:00 UTC+8 on 2026-08-12 (source). It is the first time a Max-class Qwen has been opened.
• The announced date was 2026-08-10. This wiki recorded it as missed yesterday; that record stays on the page rather than being rewritten, because a release two days late is a different fact from one on time.
• The licence is still not named. The weights are downloadable and no source read states their terms. Earlier Qwen lines shipped Apache-2.0 — that is precedent, not a licence, and it is recorded as precedent. The 27B's architecture, context window and price stayed unknown through its own release.
• Scale, from Unsloth's published quantisations: 4.9 TB at full precision, 2.6 TB at Q8_0, and 397 GB for the smallest 1-bit build — an 89–91% reduction that still does not reach a single consumer GPU.
• Why it matters: "open" and "runnable" came apart here, and the vendor conceded it by shipping a 27B alongside. An open release at this scale now means inspectable and fine-tunable by well-resourced parties, not self-hostable by individuals — which keeps the security argument for open weights intact and largely removes the democratisation one. The sequence is also now timeable: closed preview 07-19 → paid API 08-03 → open weights 08-12, twenty-four days.
• → Qwen 3.8 Max, Qwen 3.8 27B, Open-Weights Policy Fight, Alibaba / Qwen AI Lab

2. DeepMind shipped sign language translation into a keyboard and published no accuracy number

• SL2T, 2026-08-12: a multilingual sign-language-to-text model, live the same day on the Pixel 11 in Gboard and Live Transcribe as sign-to-text dictation (source). Described as the first sign language model inside a real consumer product.
• Trained on more than 100,000 hours of multilingual sign data, of which about a quarter is ASL — the only pair the shipped feature supports. Sign languages are not mutually intelligible, so a multilingual training set does not make an ASL→English feature useful to a BSL signer, and no timeline for a second pair was published.
• The architecture is a split, and worth reading precisely: an on-device model reduces camera footage to "a sort of wireframe of geometric coordinates", and those coordinates go to Google's servers for translation. Raw video is stated not to leave the device. This is a hybrid, not an on-device model.
• No accuracy, word-error-rate or comparative figure appears in anything read — and no sign language benchmark appears in any eval snapshot this repo holds, so there is nothing local to check it against either.
• Why it matters: this is an accessibility product whose entire value claim is transcription quality, announced without a single quality number. For a user who cannot hear, a wrong transcription is not a degraded experience but a failed conversation — and the one figure that would let a deaf user, a reviewer or a procurement officer judge that risk is the one figure absent.
• → SL2T (new), Google DeepMind

3. Claude's Chrome side panel stops being a separate product

• The Claude in Chrome side panel now runs as a full Cowork session: conversations, skills and connectors carry between the browser and the desktop, web and mobile apps. Previously side-panel sessions were separate and "left their conversations and context behind" (source).
• Max and Team today, Pro "in the coming weeks"; Enterprise admins can enable it and restrict access to approved domains. Nothing read mentions the Free tier.
• The browser-agent capabilities named alongside are not new to this announcement — Claude in Chrome can already click links, type, fill forms and move between pages using existing logins.
• Why it matters: this is distribution, not capability, and it is the last surface Cowork had not absorbed. The unanswered part is the permission model: an agent with page-level action rights and your live sessions now shares state with a desktop session, and no change to how that is authorised was described.
• → Anthropic, Agents (LLM Agents)

Scoring note: the formula puts story 3 top at 2.24 if the Chrome item takes the agents/tool-use weight (1.5). It is scored here on product launches (0.7) instead — what changed is session plumbing and distribution, not what the agent can do — which puts it at 0.99. Stories 1 and 2 tie at 1.86; Qwen leads on the tiebreak because a frontier model release with weights is an Actively Tracked Signal in interests.md and an accessibility model is not.