Daily LLM news, Wednesday, 4 November 2026

Good afternoon. Tencent leads today with Hy4, a full release of its biggest model that adds vision and computer use, with open weights promised in the coming days. Anthropic has open-sourced a way to prove which model produced a reply, Stanford finds that nearly half of well-known findings about LLMs no longer hold, OpenAI says GPT-6 Astra designed a working gadget and hints at a much stronger internal model, and a laptop-sized open model has repeated part of a recent weight-reading result.

LEAD: Tencent releases Hy4, adding vision and computer use, with open weights to follow

Background: Tencent open-sourced Hy4 Preview in August: 770 billion parameters in total, 49 billion active per token, with a 1-million-token context window.

What happened: On Wednesday Tencent launched the full Hy4, which it says scales up both pre-training and post-training. It can now understand images, including charts and visual maths, and it can operate command-line tools and graphical apps, with targeted training on software such as Blender and CAD. Tencent says hallucinations in its internal tests fell from 12.95% to 7.2% compared with the preview, and that in a blind test by 213 finance professionals on 579 real tasks it scored 3.04, ahead of DeepSeek V4.1 Flash (2.97), Kimi K3.1 Pro (2.68) and GLM-5.3 (2.68). It's live now in Tencent's own apps and through OpenRouter. Pricing is unchanged from the preview at roughly $0.83 to $0.90 per million input tokens and $2.50 to $2.70 per million output tokens (Tencent's release and TechNode quote slightly different figures), with 50% off on Tencent Cloud until 31 December. Weights and a system card are promised "in the coming days", and a cheaper Hy4 Flash is coming. Tencent says outside security labs found no significant offensive hacking concerns, and that offensive exploit material was left out of training.

Why it matters: Tencent now has a serious contender in the open-weight race against DeepSeek, Moonshot, Zhipu and Alibaba, and the South China Morning Post notes that Tencent itself admits it still trails the best closed models on hard maths, long coding tasks and advanced visual questions. All the benchmark numbers are Tencent's own, and it isn't open until the weights actually appear.

Sources: Tencent, South China Morning Post, TechNode, Unite.AI

Anthropic open-sources FlashProve, a way to prove which model wrote a reply

Background: Right now, nobody using a model through an API can check that the reply really came from the model they asked for, rather than a cheaper substitute swapped in by a reseller, or a tampered stream.

What happened: Anthropic's Frontier Red Team, working with researchers from EPFL, MIT and elsewhere, released FlashProve, a research prototype that uses zero-knowledge proofs. The provider publishes a short "fingerprint" of the model weights once, then attaches a proof to replies showing that this exact model produced them, without revealing the weights. Tested on Meta's 400-billion-parameter Llama 4 Maverick, one H100 GPU produces proofs at about 110 tokens per second, each proof is about 19 MB per 1,024-token request, and checking one takes about 17 seconds on an ordinary CPU. Anthropic says Claude helped improve the system, and the key security claims are machine-checked in the Lean proof language. The limits are real: proving costs about 70 to 95 times the GPU time of normal serving, it only covers 1,024-token contexts and greedy decoding, and it hides the weights but not the model's architecture. Code is on GitHub under Apache 2.0, and TechCrunch reports there are no plans to deploy it at Anthropic or any other lab.

Why it matters: Verifiable inference could let evaluators, regulators and paying customers confirm they're getting the model they were promised, and Anthropic suggests it could help with things like checking compliance with future AI rules. It's far from production-ready, but it's dramatically faster than earlier open systems.

Sources: Anthropic research post, FlashProve on GitHub, TechCrunch

Stanford: 44% of well-known findings about LLMs have already expired

Background: Many widely cited claims about what language models get wrong were measured on models that are now years old.

What happened: Owen Queen and James Zou retested 50 often-cited findings on 20 models, using results as of 30 September 2026. They classed 44% as expired, 34% as significantly weakened, 16% as still holding and 6% as stronger. Among the findings with an estimated expiry date, the average lifespan was 15.7 months. Reasoning findings fared badly, with five of the nine already expired. Tendencies proved stickier: models favouring their own answers when acting as judges, sycophancy toward users, and a left-leaning political lean all persist. Made-up citations fell from 55% for GPT-3.5 in the original 2023 study to 0% for GPT-6.1 Sol and 1% for Claude Sonnet 5.5, but Qwen3.7-Flash still invented 69%.

Why it matters: It's a useful warning for anyone quoting old research about LLM limits, and a reminder that findings should be read as true of specific models at a specific time. Behaviour problems such as sycophancy look more durable than capability gaps.

Sources: Stanford project page, The Decoder

OpenAI says GPT-6 Astra designed a working gadget, and an unreleased model scores far higher on CAD tests

Background: OpenAI has been pushing ChatGPT and Codex beyond software into physical engineering work such as 3D CAD models, circuit boards and simulations.

What happened: In a post on 3 November, OpenAI said two staff used GPT-6 Astra for under a week to design a small "Codex pet", a desk device with a screen and push-to-talk button for talking to coding agents. OpenAI says Astra designed the case, chose the components, laid out the circuit board, and wrote the firmware and the software linking it to Codex. A contract manufacturer built a small batch in under a month, but it took several rounds because Astra didn't always account for manufacturing limits, tolerances and assembly. OpenAI also cited BenchCAD 2.0, which tests rebuilding 3D parts and assemblies and tracing circuit-board connections: GPT-6 Astra at maximum reasoning averaged 0.594, while an unnamed newer internal model scored 0.875. OpenAI says its hardware teams' token use has grown 23 times since February.

Why it matters: It's another sign that coding-style agents are moving into hardware design, though still with humans fixing what the model misses. The internal model's jump is a fresh hint, after recent sightings of a "gpt-6.2-astra" model name in Codex usage data, that OpenAI has a noticeably stronger model in the wings; OpenAI hasn't said anything more about it.

Sources: OpenAI, R&D World, The Decoder

Follow-up: a 27B open model on a laptop can read network weights too

Background: Yesterday's briefing covered MIT researchers showing GPT-6 Astra could read and directly edit another model's weights, including working out from raw weights what a small network had been trained to do.

What happened: Simon Willison ran Qwen 3.8 27B locally on his laptop, with no code execution or calculator, and it correctly identified Pac-Man, Pong and Breakout from raw policy-network weights. Each successful run took roughly 46 to 65 minutes, with the Pac-Man run using about 43,500 reasoning tokens. It got one wrong, mistaking Freeway for a Pac-Man-style game. This tested reading weights only, not editing them.

Why it matters: Reading a network's purpose from its weights isn't limited to the biggest closed models, which makes the MIT result easier for others to check and build on.

Sources: Simon Willison

Also worth a look

• Kimi K3.1 Pro weights due today, still missing: Moonshot promised the full weights "by Nov 4". As of about 13:05 UK time today, the most recently updated model on its Hugging Face page was still Kimi K3 (last updated 2 September), with no K3.1 repository. The promise named no time zone, so there's still time today. Kimi blog, Moonshot on Hugging Face