NPDev Daily AI Briefing - November 4, 2026
Tencent shipped the official version of Hy4 with vision and computer use, but held back the weights, so the week's biggest open-model claim is not yet downloadable. Mistral's 1-trillion-parameter Large 4 weights are downloadable, Moonshot's Kimi K3.1 Pro weights missed their own November 4 deadline, and the most useful engineering story is Revolut's self-hosted GLM migration.
Written by NPDev Labs with AI assistance and reviewed before publication. How we work.
Tencent launches the official Hy4: vision, computer use, and weights still held back
Tencent released the production version of Hy4, its 770B-parameter mixture-of-experts flagship, on November 4. It adds visual understanding and computer use (command line and graphical interfaces) for the first time, and Tencent reports gains over the August Hy4 preview in instruction following, long-document understanding and long-horizon agent work. The model is live in WorkBuddy, CodeBuddy, Yuanbao, WeCom's assistant, ima and Tencent Cloud TokenHub, and its API is available through third-party routers including OpenRouter. List pricing is unchanged from the preview at roughly $0.90 per million input tokens, $2.70 per million output tokens and $0.045 per million cached input tokens, with a 50 percent TokenHub discount through December 31.
Why it matters: The weights are not out. Tencent says they and the system cards arrive "in the coming days," which means today you can rent Hy4 but cannot self-host it; the only download is the August preview under Apache 2.0. Treat "top-tier open-source" as a claim about a model nobody can inspect yet, and treat the benchmark numbers as vendor-run: a 579-task financial evaluation scored by 213 professionals put Hy4 at 3.04 against DeepSeek-V4.1-Flash at 2.97, Kimi K3.1 Pro at 2.68 and GLM-5.3 at 2.68, and Tencent reports its hallucination rate falling from 12.95 to 7.2 percent versus the preview.
Sources:
Mistral opens a 1-trillion-parameter model, under a license with a hosting catch
Mistral published the weights of Mistral Large 4 ("Le Chonk"): about 1 trillion total parameters with roughly 52 billion active per token, 128 routed experts with 4 active plus one shared expert, a 1M-token context, native text-and-image input, and three reasoning efforts. It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters. The release notes an embedded DSpark speculative-decoding draft head and a separate NVFP4 checkpoint for Blackwell, served by vLLM and SGLang. The blog announcing the weights is dated October 29 and the repository was updated November 2.
Why it matters: This is the first roughly-trillion-parameter checkpoint most teams can realistically download, so the license matters more than the benchmark: it is the Mistral Ecosystem Open-Weight License, a modified MIT license that permits commercial and non-commercial use but requires a separate commercial agreement to run it as a model-as-a-service. If your product resells inference, the open weights do not cover you. If you self-host internally, the DSpark draft head means speculative decoding is already in the checkpoint, so serving configuration drives both latency and cost.
Sources:
Kimi K3.1 Pro's weight deadline passed with no weights published
Moonshot's own K3.1 Pro technical blog said the full model weights would be released by November 4. As of this check, the moonshotai organization on Hugging Face has no K3.1 repository: its newest artifact is still Kimi-K3, last updated September 2, and a Hugging Face model search for Kimi-K3.1 returns only third-party derivatives, not a Moonshot checkpoint. The model itself is live on the Kimi API, Kimi apps and Kimi Code at a 1M-token context, described by Moonshot as 2.8 trillion total parameters with 104 billion active.
Why it matters: A weights promise is a procurement plan, not a fact. If a self-hosted K3.1 Pro is in your roadmap, the blocking questions are still open: whether the flagship or a distilled variant ships, whether the tokenizer and configs come with it, what the license permits for hosted use, and whether the published benchmark numbers came from the shipped checkpoint. Yesterday's deadline is now a datapoint about release predictability, not a delivery date.
Sources:
Revolut moved 2,000 engineers to self-hosted GLM and cut coding inference cost about 90 percent
David Abitbol, who runs AI operations and infrastructure at Revolut, described the migration at the AI Engineer conference in New York: roughly 2,000 engineers now default to self-hosted GLM-5.3 from Z.ai instead of Claude Code and Codex, served from about 56 NVIDIA B200 GPUs at a blended cost near $0.30 per million tokens, against previous monthly API bills in the thousands per engineer. Revolut reports coding time down about 60 percent and lead time down about 70 percent. Context is capped at 256,000 tokens, image tasks route to Gemini, and Claude Code and Codex remain available as the exception for the hardest work.
Why it matters: The interesting part is not the model, it is the operating change: a flat self-hosted bill removed per-token rationing, and the lead-time gain exceeded the coding-time gain because engineers stopped hoarding AI for hard tasks. Two caveats to keep straight before copying it. The savings exclude hardware, staff and monitoring, and out-of-the-box performance was, in Abitbol's word, terrible until tuning cut time to first token from roughly 14 seconds to 1.7 on the same GPUs. The 90 percent figure is a company measurement that does not isolate the model change from the removal of token limits.
Sources:
Gemini 2.5 Pro and 2.5 Flash start auto-rerouting on December 1, with breaking API changes
Google's November 2 Gemini API release notes deprecate gemini-2.5-pro and gemini-2.5-flash. From December 1, 2026, requests to gemini-2.5-pro automatically route to gemini-3.8-flash and are billed at Gemini 3.8 Flash rates, and requests to gemini-2.5-flash route to gemini-3.5-flash-lite at its rates. The documented breaking changes include function-calling and thought_signature validation, thinking_level replacing thinking_budget, removed sampling parameters, and removed prefilled model turns.
Why it matters: This is a silent model substitution plus a silent price change on a date, which is the worst combination to discover in production. Pinning the old model string does not protect you after December 1. Inventory every call site now, move to the new strings deliberately, and fix the parameter-level breakage (thinking budget, removed sampling parameters, prefilled turns) rather than waiting for the requests to start failing.
Sources:
Model releases & progress
Anthropic's FlashProve makes "which model answered me" a provable claim
Anthropic published research, with EPFL and MIT collaborators, on cryptographic proofs of inference: a provider commits to a set of weights and proves that each response came from them. Their system, FlashProve, generates zkSNARK proofs for a 400B-parameter Llama 4 Maverick mixture-of-experts on a single H100 at about 110 tokens per second, producing roughly 19 MB of proof per 1,024-token request with 4-bit weights, at about 70 to 95 times the GPU time of running the model itself on vLLM. The claimed security properties ship with top-level Lean statements so the guarantees themselves can be audited.
Why it matters: Routing substitution, silent model swaps and tampered token streams are all things you currently cannot detect from output quality alone, so an audit trail that binds a response to weights is a real capability. The 70 to 95 times overhead rules it out as a serving path today; it is an attestation path for regulated or high-value requests, not a default.
Two Anthropic dates and one documented behavior change
Anthropic's October 30 release note deprecates fast mode on Claude Opus 4.8 with removal on November 30, 2026: after that date requests to claude-opus-4-8 with speed set to fast return an error, and the documented migration is Claude Opus 5.5, the current flagship. Separately, the November 3 release note confirms that text generated by Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6 carries Anthropic's text watermark, with the caveat that on Sonnet 4.6 a request setting temperature, top_p or top_k to a non-default value might not carry it.
Why it matters: Both belong in a changelog you actually read. The fast-mode deadline is a hard error date for anything relying on it. The watermark is a sampling-level change: it can move outputs even when your prompt and parameters are unchanged, and if a compliance process assumes generated text is marked, the non-default sampling caveat on Sonnet 4.6 breaks that assumption.
No new frontier checkpoint from OpenAI, Google or Meta in the window
OpenAI's platform changelog still tops out at GPT-6.1 Terra from October 26, at $1 per million input and $5 per million output tokens, plus GPT-6 Rosalind from October 30, which is restricted to organizations approved for life-sciences research. Its November 3 posts were customer and partnership stories, not model releases. Google's callable frontier model remains Gemini 4 Argon, generally available since October 13. Meta shipped no model in the window; reports describe a $25 billion bond sale to fund AI infrastructure. The only Anthropic model news is unconfirmed: outlets reported that Anthropic's site briefly listed a Claude Mythos 5.5 model on November 1 and 2, with no announcement from the company.
Why it matters: A quiet checkpoint window is worth stating plainly, because it means last week's evaluations are still valid and the risk is on the platform side, not the model side. The Mythos 5.5 story is a leak, not a release, and should not enter a roadmap until Anthropic announces it.
Bottom line
The open-weight race and the open-weight download are diverging: Tencent's best model is API-only for now and Moonshot missed its own weight deadline, while the downloadable frontier moved to Mistral's 1T checkpoint and its model-as-a-service license clause. For working engineers the actionable items this week are calendar entries, not model swaps: Gemini 2.5 strings reroute on December 1, Opus 4.8 fast mode errors on November 30, and the cheapest real lever on coding cost is self-hosting plus tuning, with the caveats Revolut's own numbers leave out.