The Week in AI

Microsoft reports up to seven times the rack-scale token throughput of Maia 200, which it announced in January. Meta launched Muse Wonder and Muse Wonder Full on the same day, and Digital Trends reports that the Full model ranks second on the Artificial Analysis Intelligence Index, behind Claude Opus 5.5.

Anthropic's Frontier Red Team published research on self-replicating small models, and Zhipu started a limited test of a high-speed GLM-5.3. Only the first day of the week is covered so far, and this page will be revised as more days arrive.

What happened

Maia 300 succeeds Maia 200. It uses a chiplet design, which puts compute, the memory controller and input/output on separate dies. It carries 288GB of HBM4 memory in eight stacks, according to a chiplet breakdown by Andreas Schilling quoted on Techmeme.

A full rack holds 96 accelerators and draws about 167 kW, according to Reuters. Microsoft reports up to seven times the rack-scale token throughput of Maia 200. Scott Guthrie, who runs Microsoft's cloud and AI group, expects 40% more tokens per megawatt than leading GPU-based systems. Both figures are Microsoft's own.

Why it matters

Microsoft says Maia 300 will be the primary inference system for its own products. A chip built for serving models, and not for training them, changes how much Microsoft needs to buy from other suppliers for its own services. The size of that change depends on how fast the 2027 rollout goes.

What we don't know

Microsoft has not published independent benchmarks, and the comparison to GPU systems is an expectation and not a measurement. The records do not say how many chips will ship in 2027 or which products move first.

What happened

Both models add Active Memory, which keeps user preferences across chats, and Active Focus, a loop that tries other approaches when a task fails. Meta aims them at personal and professional agents.

Muse Wonder is in the Meta Model API and Muse Code. Muse Wonder Full is for Muse Code subscribers. Muse app plans cost $20 or $100 per month and Muse Code plans cost $50 or $200 per month. One report puts API input at $4 per million tokens, the same as Claude Opus 5.5.

What the ranking says

Digital Trends reports that Muse Wonder Full is second on the Artificial Analysis Intelligence Index, behind Claude Opus 5.5 and ahead of Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon. One third-party article puts it six points behind Opus 5.5. The Atlas could only partly confirm the ranking.

What we don't know

Meta's own benchmark numbers are not in the records. The model sizes are unknown, and so is whether the API price report is accurate.

What happened

The team tested models of 30B parameters and under. They could complete many of the individual replication steps but made mistakes when chaining them together. The targets discussed are devices such as smart ovens and off-brand Android TV boxes.

Anthropic says a replicator like this would have no central point to shut off and could work out unfamiliar devices by itself, which ordinary malware cannot. It expects the spread to be slow, with a doubling time of days.

Why it matters

The result is a measurement of component skills, not a working replicator. It gives a concrete hardware count for how far small models have come on one specific kind of task.

What we don't know

The Atlas only partly confirmed the source text, which cuts off mid-sentence. The records do not give the full test method or how Anthropic arrived at "months away".

What happened

GLM-5.3 HighSpeed keeps the 1 million token context window and 131,072 token maximum output of GLM-5.3, so the difference is speed. AI Ping measured 70.84 tokens per second for the standard GLM-5.3 API and 121.34 for DeepSeek V4.1 Flash.

The 490 figure comes from one tester, reported by ZheDongXi, and may not hold at scale. An earlier GLM-5.1 high-speed API reached 400 tokens per second with the TileRT team, which compiles the inference graph ahead of time so it does not change at runtime.

What we don't know

Zhipu has not said how the speedup is achieved or when a public API will exist. The price is unknown.

Models and agents

• Z.ai announced GLM-Legal, built on GLM-5.3-Flash, and reports a 19.2% all-pass rate on Harvey LAB, 12.5 points above its base model.
• Huawei Noah's Ark Lab reported that its Memento 3 agent cleared all 25 ARC-AGI-3 public games with a frozen model, according to Pandaily.
• Mistral AI unveiled ML4, an open-weight model, according to Digitimes.
• OpenAI launched GPT 5.6 Cyber and expanded its Daybreak security service, according to Digit.
• StepFun released Step 5 Preview on 8 October, a 600B-total, 27B-active mixture-of-experts model at $1 per million input tokens and $2.70 per million output tokens.

Hardware and interfaces

• NVIDIA contributed Vera Rubin rack and compute tray designs to the Open Compute Project through MGX 1.2, which supports racks up to 250kW.
• Runway showed Project Continuum, a demo operating system where parts of the screen are real-time generated video, and only a demo video exists.

Money

• DeepSeek reportedly led a seed round for a robotics lab founded by a former Tencent scientist, and reports differ on whether $400 million is the round size or the valuation.

No moves are recorded yet for this week.

The top models from Anthropic, OpenAI and Google are still closed. Meta's Muse Wonder models are closed too, and Microsoft's Maia 300 will not reach production scale until 2027.

• Open weights for Mistral Large 4 by the end of October, which Mistral announced on 6 October.
• The weights and technical report for Reflection AI's Beam, due later in October.
• Independent scores for Muse Wonder Full, to test the reported second place on the Artificial Analysis Intelligence Index.
• 16 October, when the Surface Laptop Ultra on NVIDIA and Microsoft's RTX Spark platform ships.
• 23 October, when ByteDance's TraeWork stops service.