The Week in AI

The page covers Monday so far. MiniMax says the model is nearly three times faster than M3 and costs 70% less per agentic task, and the model is available only through its API. Artificial Analysis also measured a hallucination rate of 93% for it.

On the same day Inception Labs released Mercury 3, which it says matches GPT-6 Luna and Claude Haiku 5.5 on benchmarks at 4 to 8 times the speed. Amazon Web Services released TorchNeuron, and VAST open-sourced EdenLive, a real-time renderer for game footage. Reports also say Disney is set to hire most of Character.AI's staff and that the US Department of Homeland Security plans to ban DeepSeek from federal civilian agencies.

What happened

M3.1 Flash is the next Flash model after M3, built for agentic work and coding. MiniMax says it uses more aggressive sparsity, which means fewer parameters are active for each token, and credits that for the speed. The company reports output above 200 tokens per second, with a median of more than 170.

Pricing is $0.15 per million input tokens and $0.60 per million output tokens. Cached input costs $0.015 per million tokens. MiniMax says the model is nearly three times faster than M3 and costs 70% less per agentic task.

Before launch the model was tested anonymously under the name Space Bunny. MiniMax says that version processed more than 200 trillion tokens.

MiniMax reports these benchmark results, all company-measured. It gives 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 76.2 on BankerToolBench and 82% on CyberGym. On Agents' Last Exam it reports 29.6.

The independent numbers come from Artificial Analysis, as reported by The Decoder. It scores the model 42 on its Intelligence Index and puts the cost at about $0.22 per task. It also measures a hallucination rate of 93%.

Why it matters

The jump from 29 to 42 on one independent index is large for a model that the company describes as a cheaper, faster tier. At $0.15 per million input tokens it is priced for agents that make many calls per task.

The 93% hallucination rate is high. A reader who plans to use the model for factual answers should look at that figure before the intelligence score.

What we don't know

MiniMax has not published a parameter count in the records. The speed and the 70% cost reduction are company figures, and no outside party has measured them. The records also do not say how Artificial Analysis defines its hallucination measure, so the 93% cannot be compared with other models here.

What happened

Mercury 3 is a diffusion language model. It starts from a rough draft of the whole output and refines many tokens in parallel over a few steps, which lets it generate text faster than models that write one token at a time.

Inception reports 1,068 tokens per second on NVIDIA GPUs. It says the model matches GPT-6 Luna and Claude Haiku 5.5 on benchmarks while running 4 to 8 times faster than both. The context window is 1 million tokens, up from 260,000 for Mercury 2.5.

The model has more than 40 times the parameters of Mercury 1, and Inception says benchmark intelligence rose threefold from Mercury 1 mini. The company presents this as evidence that scaling laws hold for text diffusion models.

One agent result is quoted on Inception's blog. In a test with early-access partner TinyFish AI, Mercury 3 completed 50% of 50 live Mind2Web navigation tasks in TinyFish's production web-agent harness. The result is reported as significantly more than GPT-6.1-Sol. It is a company-quoted figure from a small sample.

The list price is $0.20 per million input tokens and $0.70 per million output tokens. At launch both are halved, to $0.10 and $0.35. Cached tokens cost 90% less than the input price. Access is through a closed API.

Why it matters

Diffusion language models have so far been smaller and had lower capability than the models they are compared with. Inception now claims parity with two named models at a fraction of the latency, and a larger model is the evidence it gives for the scaling claim.

Speed matters most for agents that chain many calls, where each step waits on the one before.

What we don't know

None of the speed or quality figures have been independently measured. The records do not say which benchmarks back the claim of parity, or how many parameters Mercury 3 has in total. The comparison models are small tiers from other labs, so the result says nothing about larger models.

What happened

TorchNeuron is adapted from PyTorch itself. Developers can move existing PyTorch apps onto Amazon's Trainium chips without rewriting code. Amazon lists the release as a research preview.

Amazon says 85% of AI apps are written with PyTorch, which is the company's own claim. It also says it is sending changes upstream to PyTorch to remove code that ties the framework to specific chips.

Amazon cites Deepgram as a customer. Deepgram moved to Trainium and processes 70% more audio per dollar at production-grade accuracy, a result that Amazon cites and that the records do not show an outside party measuring. TechCrunch framed the release as Amazon's biggest attempt yet to challenge Nvidia's software position.

Why it matters

Moving between chips has meant rewriting code for each vendor's software. A backend that accepts unmodified PyTorch removes much of that work for Trainium, and the upstream changes would help other chips too if PyTorch accepts them.

What we don't know

The records do not say which PyTorch features TorchNeuron supports or how performance compares with Nvidia GPUs on the same model. Whether PyTorch maintainers will accept the upstream changes is open. The Deepgram figure is a single customer result.

Models and agents

• VAST open-sourced EdenLive on 19 October, a real-time model that re-renders coarse game or video footage into detailed styled frames. It does not simulate the world. A game engine or code written by a coding agent runs the rules and outputs a coarse reference video, and EdenLive renders it guided by a target first frame and a text prompt.
• VAST says this keeps game logic explicit and testable. The 5B model builds on Wan2.2-TI2V-5B and the 14B model on Wan2.1-I2V-14B, under a restricted license, with streaming inference code, a technical report and demos.
• VAST reports that the 5B model runs above 24 frames per second at 720p on one RTX 5090, about 24.7 at 1280x704 using LightX2V acceleration. The technical report gives 29.8 frames per second for the 5B model on one H100, and 26.8 for the 14B model on eight H100s at 832x480.

Hardware and compute

• DeepSeek is working with Huawei on future chip designs and plans to pretrain on Ascend 950DT as soon as the end of 2026, according to The Information.
• Nscale is reportedly in talks on a second deal with Anthropic worth about $45 billion for 460 megawatts in West Virginia on Vera Rubin chips, according to Bloomberg.
• Meta described Prometheus, its gigawatt-scale Ohio supercluster that trained the Muse models. Meta says it will have about 1 million GPUs at full scale.

Policy and deals

• US Department of Homeland Security plans to ban DeepSeek from federal civilian agencies, according to CyberScoop.
• Disney is in a deal to hire most of Character.AI's staff after the company filed for bankruptcy, according to Ars Technica citing court filings. A judge must approve it.

• Karandeep Anand, Character.AI's former chief executive, is among the more than 60 employees set to join Disney, according to Ars Technica and Business Insider. The deal needs a judge's approval.

M3.1 Flash and Mercury 3 are both closed API models, and neither has public weights. EdenLive is the only open-weight release in the records so far this week, and it is under a restricted license. Independent measurements of Mercury 3's speed and quality are still missing, and so is any MiniMax parameter count.

• Independent speed and quality results for Mercury 3, since all current figures come from Inception.
• Whether Artificial Analysis or others reproduce MiniMax's output speed for M3.1 Flash.
• Whether PyTorch accepts the upstream changes that Amazon says it is contributing.
• 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.
• The Apple Intelligence beta in China, due in late October.
• The bankruptcy judge's decision on the Disney and Character.AI hiring deal.
• 15 November, when Sachin Katti leaves OpenAI.