The Week in AI

The page covers Monday 2 November and Tuesday 3 November. Meituan's model is built for long agentic coding tasks and reads images and text in one stream. All of its benchmark figures come from Meituan, and no independent evaluation has been published.

MIT CSAIL researchers reported that GPT-6 Astra can read a small model's raw weights and rewrite them without any training loop. Anthropic released FlashProve on 2 November, an open-source prototype that proves a response came from a specific set of committed weights, and put message threads into beta in the Claude Messages API. Company figures are marked as such, and records the Atlas could only partly confirm are written as reported.

What happened

LongCat-2.5 builds on LongCat-2.0 and adds image understanding. Images are turned into multi-level discrete tokens, which are sequences of integer codes at several levels of detail, and these sit in the same stream as text. One model therefore reads both without a separate vision component.

Training added more long-horizon trajectory data, meaning records of agents working through tasks with many steps. It also added large-scale reinforcement learning (RL), where the model is rewarded for finishing those tasks. Meituan takes failure trajectories from real agents and turns them into verifiable training tasks, so the model trains on the situations where agents actually broke.

The context window is 1 million tokens, with output up to 128,000 tokens according to Crypto Briefing. Meituan also optimized W4A4 inference on domestic Chinese chips. W4A4 means both the weights and the activations are stored in 4 bits, which cuts memory and compute at some risk to accuracy.

Meituan reports the following scores, all measured by the company.

• Terminal Bench v2.1, 88.4, against 70.8 for LongCat-2.0.
• AutomationBench, 51.8, against 15.8 for LongCat-2.0.
• BrowseComp, 91.9.
• PaperBench, 82.1.
• 3DCodeBench, 55.4.
• Vision2Web, 78.6.

Access is through a closed API.

Why it matters

The jump on Terminal Bench v2.1 is 17.6 points over the previous LongCat version, and the AutomationBench score more than tripled. Both are comparisons between Meituan's own models, run by Meituan, so they show the direction of progress within one product line and say nothing about rank against other labs.

The model is larger than Mistral Large 4, which has 1 trillion total parameters and whose weights Mistral released on 29 October. LongCat-2.5 is available only through an API, so outside groups cannot inspect it or run it themselves.

The W4A4 work on domestic chips tells a practitioner something about deployment. Meituan is serving a model of this size in 4-bit precision on Chinese hardware, and the company presents that as part of the release.

What we don't know

No independent evaluator has published results for LongCat-2.5. The records give no API price and no licence terms, and Meituan has not said how much accuracy the W4A4 setting costs against higher precision. It is also unknown which settings Meituan used for each benchmark.

What happened

The paper is titled "Sparks of Weight-Space Programming" and has the subtitle "LLMs can natively read & write neural network weights". Earlier methods set weights through optimizers such as Adam or Muon, through trained networks that write weights, or through compilers. Here GPT-6 Astra reads the raw weights as input, reasons about them in natural language and writes edits, with no gradients and no training loop.

The authors say the model found the reasoning loop problem that Qwen's model card warns about in Qwen3.5-0.8B. In that failure the small model keeps reasoning without reaching an answer. Astra fixed it by editing 0.016% of the weights.

According to the authors, MATH-500 rose from 7.9% to 48.1% under the same sampling settings. IFEval, an instruction-following benchmark, improved about threefold, which the authors describe as generalization from the maths edit.

A second test used the Kinetix car-on-ramp task, a reinforcement learning control problem. Changing 3.5% of a policy's weights took success from 0% to 100% over 1,024 episodes, on a single try. The access label for the work is paper only.

Why it matters

If the method holds up, a model can repair another model's specific behaviour by reading it, without collecting data or running training. The reported edit is small, and the fault it fixed has a published description in the model card.

Both results come from the authors, and no independent group has reproduced them. The Atlas checked the record against several write-ups and the paper's blog post.

What we don't know

It is unknown whether the method works on larger models, since the tests use a 0.8 billion parameter model and a small control policy. The records do not say how many attempts Astra needed per edit or what the editing costs in tokens.

What happened

A provider publishes a short commitment to an exact set of weights. Each prompt and response pair then gets a zkSNARK proof, which is a compact cryptographic proof that a computation was done correctly and that anyone can check without seeing the secret inputs, here the weights. Verification runs on a CPU.

The prover is built for Llama 4 Maverick, a mixture-of-experts model that sends each token through a small subset of its expert sub-networks. Anthropic says the work scales with about 17 billion active parameters, and reportedly the expert choices stay hidden. The design is software only and does not depend on trusted hardware.

Anthropic reports the following figures, all measured by the company.

• Proving speed of 110 tokens per second on Llama 4 Maverick.
• A batch of 24 requests proved in about 223 seconds on one H100 with 24 CPU cores, using a 4-bit conversion of the model.
• Proofs of 18.78 MB per request.
• Verification in 17 seconds with 24 threads and 21 seconds with one thread, against a commitment of about 6.5 MB.

The cost is high. Anthropic puts proving at about 72 times the per-GPU time of FP8 vLLM on eight H100s, and about 94 times at the highest vLLM load it tested. The code is reportedly under the Apache 2.0 licence and labelled an unmaintained research artifact, and the access label is research preview.

Why it matters

The aim is to let a customer of a model whose weights are hidden check that an answer came from the exact weights the provider committed to. Today that check relies on trusting the provider.

At 72 to 94 times the GPU time of ordinary serving, the method is far from something a provider would run on every request. The proof sizes and CPU-only verification do keep the checking side cheap for the customer.

What we don't know

Nobody outside Anthropic has published measurements of FlashProve. The records do not say whether Anthropic plans to apply it to its own models, since the demonstration uses Llama 4 Maverick. The Atlas could only partly confirm the claims about hidden expert choices and the licence terms.

What happened

Until now a client has sent the whole conversation with every request. With message threads the API stores the conversation on the server, so each later request carries only the new messages.

Anthropic says smaller requests give long conversations and agent loops a lower time to first token, which is the delay before the first word of the reply arrives. A thread expires 5 minutes after its last request and is deleted no more than 1 hour after its last use. The feature is in beta on the closed API.

Why it matters

Agent loops repeat the same long history on every step, so the request size grows with each turn. Moving that history to the server removes the repeated upload. The five minute expiry means a thread has to stay active to stay usable, which affects agents that pause while waiting on a tool or a person.

What we don't know

The records give no pricing for stored threads and no date for the end of the beta. They do not say how threads interact with existing prompt caching, or whether the stored conversation falls under any retention policy beyond the stated deletion window.

Models and products

• OpenAI is merging the Chat and Work tabs in the ChatGPT desktop app, TechCrunch reported on 2 November. Users can bring files and plugins and ask ChatGPT to build a website or act on the web, Codex stays separate, and ChatGPT and Codex now draw from one shared usage allowance, so some users will wait for resets more often.
• Alibaba Cloud launched Memory on Model Studio, a service that lets agents keep a user's preferences, plans and earlier commitments across sessions. It reports 85.15% on HaluMem-medium question answering against 81.71% for Zep, and 56.82 on PersonaMem-v2, first among eight systems.
• SenseTime released the official SenseNova 6.8 Flash on 2 November, a multimodal agent model for office and design work that also stresses memory across conversations.
• Baidu launched Apollo World, a simulator that produces camera and lidar data for training its Apollo Go robotaxis. Baidu says it generates up to 100,000 km of simulated driving a day, which it equates to about three years of real-world testing.
• DeepSeek added a creator mode to DeepSeek Harness, two days after releasing Super-agent. It lets users turn an idea into a working plugin, such as a 3D model renderer or a video editing workbench, without writing it by hand.
• Tencent Cloud launched WAND-Vega on 30 October, a multimodal model for images, video and speech. Tencent says its WAND platform serves more than 2,000 companies.
• ServiceNow released Apriel Voice 27B Preview on 29 October, a speech-language model built on Gemma 4 26B-A4B that handles speech and tool calls in one model.

Apple

• Apple reportedly opened Siri AI integration through App Intents to all developers in the iOS 27.3 betas, according to Techmeme's summary of the announcement. The API details and limits are unconfirmed.
• Apple began beta testing Audio Intelligence for Apple Watch Series 12 in iOS 27.3, with four opt-in features for sounds, songs, conversation catch-up and notes.
• Apple says Apple Intelligence in China, which uses Alibaba's Qwen, has regulatory approval. A wider launch is reported for December in iOS 27.3.

Policy and money

• CrowdStrike agreed to buy Fiddler AI for about $400 million and Dreadnode for about $95 million, according to the Wall Street Journal.
• Anthropic warned that it is very close to recursive self-improvement capability, meaning AI that builds AI, TechCrunch reported.
• OpenAI consumer product lead Adam Fry told Big Technology that ChatGPT has 1.2 billion weekly users and search volume about 20% of Google's.
• Google took its Dreambeans AI stories app to India on 2 November with 13 Indian languages, after Sensor Tower counted about 5,500 US downloads in October.

• Jill Hazelbaker, president and chief corporate affairs officer at Uber, is reportedly joining OpenAI on 3 November to lead global communications as Chief Corporate Affairs Officer, a role The Information says has been empty since the start of the year.
• Dan Driscoll, the 26th Secretary of the Army, reportedly joined OpenAI on 2 November in a new role as Senior Counselor for National Security Modernization. The Information reports he left his Army post in September after reportedly clashing with Defense Secretary Hegseth, and that detail is only partly confirmed.

Moonshot AI has not released the weights of Kimi K3.1 Pro, which it promised by 4 November 2026. The model remains API only. GPT-6.1 Terra is also a closed API, like the rest of OpenAI's GPT-6 line.

Gemini 4 Carbon still has no date for access beyond selected cyber defenders. Siri AI is still unavailable in the EU and mainland China. OpenAI's GPT models are still due to reach Google Cloud's Model Garden, with general availability planned in the coming weeks.

The ETH proofs OpenAI announced on 26 October have no community confirmation yet, and the records say nothing new about the proof Anthropic reportedly claimed.

• Independent results for LongCat-2.5, since every figure in the records comes from Meituan.
• A replication of the weight-editing results on a larger model than Qwen3.5-0.8B.
• Any outside measurement of FlashProve, and any sign that Anthropic applies it to its own models.
• Pricing and an end date for the beta of Anthropic's message threads.
• 4 November, the date by which Moonshot promises the full Kimi K3.1 Pro weights.
• Confirmation or rejection of the ETH proofs by the community review.
• 15 November, when Sachin Katti reportedly leaves OpenAI.
• December, when Apple Intelligence with Qwen is reported to launch widely in China in iOS 27.3.