stephane

Qwen3.8-Flash-Next runs 6B active and weighs 173 GiB

Official Qwen architecture diagram for Qwen3.8-Flash-Next, showing input tokens feeding a vocabulary embedding and a separate n-gram embedding layer at layer 2, a stack of Gated DeltaNet layers punctuated by Qwen Sparse Attention layers in a three to one hybrid block, Gated Residual read and write gates around each MoE block, and MTP modules beside the prediction head.

by stephane

3 September 2026

0

Somebody on the team read "6B activated" and asked which of our boxes could run it. Fair question, wrong number:...

Tencent’s 770B Hy4 tops one benchmark row in 46

The Agentic Coding section of the official Hy4 preview benchmark appendix published by Tencent, a table comparing Hy3 and Hy4 preview against DeepSeek V4 Pro 0813, Qwen 3.8 Max, GLM 5.3, Kimi K3, GPT 5.6 Sol and Claude Opus 5 across SWE-bench Multilingual, SWE-bench Pro, DeepSWE, three SWE Atlas tasks, SWE-Marathon, Terminal-Bench 2.1, NL2Repo-Bench, CyberGym, ProgramBench, PostTrainBench and Harbor-Index.

by stephane

3 September 2026

0

Most labs publish the launch chart and bury the table. Tencent did the opposite on 28 August, and the table...

Trainium4 takes Nvidia memory, which matters more than 2M GPUs

Nvidia render of an AI accelerator package on a black background, showing a silver metal lid frame around a central grid of dark compute dies flanked on both sides by six gold high bandwidth memory stacks, with fine interposer routing visible across the substrate.

by stephane

3 September 2026

0

Two million GPUs is the number in every headline, and honestly it's the least interesting thing in the announcement. AWS...

Microduck is $399, with open software and closed hardware

Four Microduck biped robots photographed side by side on a black background in the Sky, Graphite, Cream and Lavender colourways, each about 25 centimetres tall with a large duck-bill head housing a camera lens, exposed black servo-driven legs and coloured plastic feet.

by stephane

3 September 2026

0

Four robot ducks on a black backdrop, 25 centimetres tall, in cream and graphite and lavender and sky. That launch...

Nvidia pauses its $36B AI Compute Partnership

Answer card stating that the Wall Street Journal reported on 27 August 2026 that Nvidia stepped back from parts of the AI Compute Partnership financing programme launched on 1 July, after internal warnings about antitrust scrutiny and provider pushback over Nvidia approving their end customers, that Nvidia says the business model is still in place, and that its quarterly filing books 36 billion dollars of AI cloud agreements with none of it before fiscal 2028.

by stephane

29 August 2026

0

Two months. That is how long Nvidia's neocloud financing programme ran before the Wall Street Journal reported, on 27 August,...

AWS bought DuckLabs, and DuckDB stays MIT

by stephane

3 September 2026

0

You ran `pip install duckdb` this morning and nothing was different. Good, because that is the honest answer here: AWS...

Gemini Omni 1.1 Flash stitches its 40 seconds and upscales its 4K

Google announcement card for Gemini Omni 1.1 Flash, the model name in black on white with the line Available via APIs beneath it, ringed by eight rounded video thumbnails including a chipmunk beside a flower, a woman in a blue lit room and a man writing at a desk by a window.

by stephane

3 September 2026

0

Four dollars. That is what forty seconds of Gemini Omni 1.1 Flash costs at 720p list price, and for what...

MTurk closes on 30 September, and approvals run a month later

Answer card stating that Amazon Mechanical Turk closes permanently on 30 September 2026, that HIT submission stops that day while approvals, rejections and bonuses run until 30 October 2026, and that transaction history stays available until 28 January 2027.

by stephane

3 September 2026

0

Five weeks. That is how long a working `boto3` MTurk client has left, because Amazon is closing Mechanical Turk permanently...

Anthropic’s Nscale deal buys 460 MW for late 2027

Answer card stating that Bloomberg reported on 26 August 2026 that Anthropic agreed to pay Nscale roughly 45 billion dollars over six years for about 460 megawatts at the Monarch campus in Mason County, West Virginia, on NVIDIA Vera Rubin systems, with CNBC matching the report, Anthropic declining to comment, and the facility expected online at the end of 2027.

by stephane

3 September 2026

0

Two people who would not be named, a British infrastructure company founded in 2024, and a field in Mason County,...

Gemini 3.5 Transcribe costs half a cent a minute, ranks fifth

Google official announcement card for Gemini 3.5 Transcribe, the model name set in black beside the four point Gemini spark on a pale blue gradient, with a blurred white sound wave motif in the upper right corner.

by stephane

3 September 2026

0

Google published a chart on 26 August showing its new transcription model beating everyone else, and it is a perfectly...

Gemini 3.8 Flash keeps the price and the 1 January cliff

Anthropic moves retention into your own cloud, for 30 days

Android 17 hides the SNI, not your DNS or destination

Claude Fable 5.1 breaks forced tool use, cuts cache 75%

The EU now calls ChatGPT a very large search engine

ChatGPT Mil and Grok reached IL5 on GenAI.mil

Anthropic’s Model Hardware Standard is gated, and sits under MCP

Cohere Parse 5 is $1.50 per 1,000 pages, on three of five dimensions

WebMCP in ChatGPT needs GPT-5.6 Sol or Terra

Tencent’s 770B Hy4 tops one benchmark row in 46

Qwen3.8-Flash-Next runs 6B active and weighs 173 GiB

A judge vacated the Pentagon’s risk label on Anthropic

Qwen3.8-Flash-Next runs 6B active and weighs 173 GiB

Tencent’s 770B Hy4 tops one benchmark row in 46

Trainium4 takes Nvidia memory, which matters more than 2M GPUs

Microduck is $399, with open software and closed hardware

Nvidia pauses its $36B AI Compute Partnership

AWS bought DuckLabs, and DuckDB stays MIT

Gemini Omni 1.1 Flash stitches its 40 seconds and upscales its 4K

MTurk closes on 30 September, and approvals run a month later

Anthropic’s Nscale deal buys 460 MW for late 2027

Gemini 3.5 Transcribe costs half a cent a minute, ranks fifth

Follow Me

Instagram Photos

Welcome Back!

Retrieve your password