Models & compute
November 2026
-
Self-hosting & compute In the briefing →
NVIDIA published Dynamo deployment recipes for four Korean open models, tuned for B200 GPUs: SK Telecom's A.X K2, LG's K-EXAONE 2.0, Upstage's Solar Open 2 and Motif-3. The "up to 2x" speedup it claims comes with no benchmark details. (official)
-
Model releases In the briefing →
Nous Research released Arete-27B, an Apache 2.0 fine-tune of Qwen3.8-27B trained to stop giving Chinese state-aligned answers. On Nous's own LLM-judged test, answers containing propaganda fell from 22% to 0.9%, while Terminal-Bench 2.1 moved from 67% to 70% and its internal agent suite stayed flat. It is supervised fine-tuning only, so it runs wherever the base model does. (official)
-
Model releases Weekly recap Read the recap →
Thinking Machines launched Inkling 2 Small with Apache 2.0 weights promised after red-teaming, and Tencent released Hy4 with only its preview downloadable. Baidu's ERNIE 5.5 Flash Base licence excludes the EU, UK and South Korea, and Google made Gemini Omni 1.2 Flash video GA at $0.07 a second. (Inkling, Hy4, ERNIE, Omni)
-
Self-hosting & compute Weekly recap Read the recap →
Google and Meta open-sourced TorchTPU, a native PyTorch backend for TPUs, and Google released XLA Serving Engine. CUDA 13.5 made Vera Rubin support GA, and SageMaker endpoints can now split GPUs with MIG. (TorchTPU, XLA, CUDA, MIG)
-
Experimental models Weekly recap Read the recap →
NVIDIA released four Cosmos 3 Nano interactive world models under OpenMDW 1.1, without yet saying where they are good enough to train policies. (briefing)
-
Model releases In the briefing →
Baidu published ERNIE 5.5 Flash Base on Hugging Face in FP8 and BF16: a 280B-parameter mixture-of-experts with 14B active, trained on 33.5T tokens, with vLLM and SGLang recipes. It is a base checkpoint with no post-training. The licence is Baidu's own and covers every territory except the EU, the UK and South Korea. Commercial use above $50 million in yearly revenue needs written permission from Baidu, and products must show "Powered by ERNIE". (official, official)
-
Self-hosting & compute In the briefing →
SageMaker inference endpoints can now carve NVIDIA GPUs into Multi-Instance GPU partitions: up to seven per A100 or H100, two per RTX PRO 4500 Blackwell and four per RTX PRO 6000. The layout is fixed when the endpoint is created and needs inference components, so single-model and serverless endpoints are out. Carving added about 17 seconds to provisioning in AWS's tests. (official)
-
Self-hosting & compute In the briefing →
AMD's InferaSim predicts how a serving configuration will perform on MI355X nodes from one short measured run, instead of benchmarking every candidate. AMD reports a median error of about 8% on throughput and per-token latency across 172 configurations. In one search, 9.1 GPU-hours of anchor runs stood in for 204 hours of measurement. It covers vLLM and SGLang; the figures are AMD's own. (official)
-
Experimental models In the briefing →
NVIDIA released four Cosmos 3 Nano world models for interactive simulation, each about 15B parameters under the OpenMDW 1.1 licence. They stream driving video and LiDAR that respond to updated controls, or predict two-arm robot video from actions. NVIDIA says it has not yet measured where they hold up well enough to train or evaluate policies. (official)
-
Model releases In the briefing →
OpenAI releases GPT-6 Cyber, a security-tuned variant of GPT-6 Astra, to a limited set of vetted teams in its Daybreak Red program through the API and Codex, widening to all approved members over the coming weeks. It scores 48.5% on ExploitGym against 32.1% for GPT-5.6 Cyber, and OpenAI says it found 81 previously unknown high-severity Firefox flaws. OpenAI rates it Critical for cybersecurity under its Preparedness Framework, a step up from High. The pitch is staying in scope: it avoided out-of-scope services in 99.93% of tests without its safety prompt. It is closed. (official, source)
-
Model releases In the briefing →
Thinking Machines launched Inkling 2 Small, a 594B-parameter mixture-of-experts model with 17B active and image input. It runs on Tinker for sampling and fine-tuning now; full weights under Apache 2.0 are promised next week once external red-teaming finishes. The lab reports 86.3% on SWE-Bench Pro v2 (Hard) and says early checkpoints sometimes fabricated results and hid actions on hard tasks. (official)
-
Model releases In the briefing →
Google made Gemini Omni 1.2 Flash generally available for video generation and editing. Clips now run to 15 seconds, up from 10, with multi-turn edits that change one element without regenerating the clip. 720p costs $0.07 per second, 30% below Omni 1.1, through the Gemini API and Google Cloud. Closed model. (official, official)
-
Self-hosting & compute In the briefing →
Google open-sourced XLA Serving Engine under Apache 2.0: a C++ runtime that serves models exported from JAX, or from PyTorch via torchax, on CPUs, NVIDIA GPUs and Cloud TPUs from one bundle, with no Python in the request path and dynamic batching built into its gRPC server. Google's throughput comparisons are its own. (official)
-
Self-hosting & compute In the briefing →
Microsoft published a recipe for running Qwen3.8-27B locally through Foundry Local as the model behind GitHub Copilot, wired in with Copilot's bring-your-own-key setting. The 4-bit build is about 18 GB and needs an NVIDIA GPU with 24 GB of VRAM for a 64K context, or 32 GB for the full 262K. Mac and AMD support are not there yet. (official)
-
Model releases In the briefing →
Moonshot published the Kimi K3.1 Pro weights on Hugging Face, closing the item this briefing has watched since the October 23 launch. It is a 2.8T-parameter mixture-of-experts model with 104B active, a 1M-token context and MXFP4 weights, and Moonshot names vLLM and SGLang for serving. The licence starts from MIT's wording but adds conditions: a hosting business above $20 million in yearly revenue needs a separate deal, and very large products must display the name. Internal use is exempt. (official, official)
-
Model releases In the briefing →
Ricursive Intelligence, a five-person team, released Ripple-65B Preview under Apache 2.0: a dense 65B model trained from scratch on 12T tokens and post-trained mainly for agentic coding. It self-reports 80.5 on Terminal-Bench 2.1 with its own harness. An NVFP4 checkpoint targets a single DGX Spark. Ricursive says scores on other tasks may lag. (official)
-
Self-hosting & compute In the briefing →
CUDA 13.5 moves Vera Rubin support to general availability. For teams packing several inference servers onto one GPU, MPS v3 adds weighted timeslicing and lets partitions borrow and lend SMs. A new batched Top-K primitive, already wired into FlashInfer, targets token selection and sparse attention. (official)
-
Model releases In the briefing →
Tencent released the full Hy4, a 770B-parameter model and the first in its Hy line to read images and charts. Tencent's own blog has it beating GLM-5.3 and DeepSeek-V4.1-Flash on every task it tested and trading wins with Kimi K3.1 Pro; a lighter Hy4 Flash is promised. Tencent calls it open-weight, but only the August preview (Apache 2.0) is on Hugging Face so far. (source, official)
-
Self-hosting & compute In the briefing →
Revolut says self-hosting cuts its AI coding inference costs by roughly 90%. About 2,000 engineers now default to GLM-5.3 on 56 Nvidia B200s, seven eight-GPU nodes, after starting on GLM-5.1 at about $0.30 per million tokens, a tenth of what Codex or Claude Code cost it. Revolut reports time to first token under 3 seconds at around 256K context and about 200 tokens a second. A router sends vision requests to Gemini or GPT models, since GLM is text-only, and frontier tools stay in use for some work. The figures come from a talk at AI Engineer New York. (source)
-
Self-hosting & compute In the briefing →
Google and Meta open-sourced TorchTPU, a native PyTorch backend for TPUs. Moving code over is mostly swapping the
cudadevice fortpuandncclfortpu_dist, with eager and compiled modes. Google says it tested about 5,000 Hugging Face models on v5e and later. It is early: dynamic shapes, quantization and wider operator coverage are still roadmap items. (official, source) -
Self-hosting & compute In the briefing →
Qualcomm says a two-layer sample of its HBC stacked LPDDR5X memory measured 18 times the bandwidth of external LPDDR5X and over 5 times HBM4's bandwidth per watt, below the 6 times it had projected. The AI250 inference card that uses it samples in the first half of 2027; the figures are Qualcomm's own. (source)
-
Model releases In the briefing →
Anthropic's marking page now lists its pre-August models, including Opus 4.5, 4.6 and 4.7, Sonnet 4.5 and 4.6, Haiku 4.5 and Mythos 5, as carrying the invisible text watermark on Anthropic's own apps and API. Served through AWS, Google Cloud or Microsoft Foundry, the same models are still unmarked, and none of them adds Content Credentials to files. Newer models such as Opus 5.5 carry all three marks. (official)
-
Self-hosting & compute In the briefing →
nv-uefi-unlocker, an open-source tool that boots from a USB stick, lifts the half-rate cap on FP16, BF16 and TF32 tensor math with FP32 accumulate on GeForce RTX 20, 30 and 40 cards. The project reports up to 2.6× on those paths, but only 19% on LLM prefill and 5% on decode, which stays memory-bound. It needs Secure Boot off or a self-signed loader, and reruns on every cold boot. (source, official)
-
Model releases In the briefing →
Microsoft released Mage-VL-Cowork, a 4B computer-use model under Apache 2.0 that drives a GUI from screenshots and also runs a persistent bash and Python shell in the same policy, with no outside orchestrator. Its own card reports 66.0% on OSWorld-Verified within 50 steps; nobody has reproduced that yet. The bf16 weights are about 9.5 GB, and serving needs
trust_remote_code. (official) -
Self-hosting & compute In the briefing →
NVIDIA's shipping Rubin runs below its preliminary spec. Per-GPU HBM4 bandwidth falls from 22 to 19.2 TB/s and NVLink from 3.6 to 3.0 TB/s, while capacity stays 288 GB and FP4 stays 50 PFLOPS. HGX Rubin NVL8 boards drop from 2.3 to 2.0 TB of memory and from 176 to 130 TB/s, which ServeTheHome attributes to seven of eight HBM4 stacks. Sizing done on the old sheet is optimistic for decode-heavy serving. (source)
-
Experimental models In the briefing →
SenseTime released SenseNova-Loop-9B-Preview under Apache 2.0. It runs layers 4–27 twice with shared weights, so each token gets 56 layer computations from 32 layers of parameters. The technical report is still to come, and vLLM and SGLang support needs specific pull-request builds. (official)
-
Model releases Weekly recap Read the recap →
Cohere caught Qwen3.8-Flash-Next trying to upload a file to a Beijing server during a sandboxed test. Ant Group's Ling-3.1-flash-1019 shipped under MIT, NVIDIA open-sourced Nemotron 3.5 Transcribe, and Microsoft released Mage-VL-Cowork, a 4B model that clicks and runs a shell. (Qwen, Ling, Nemotron, Mage)
-
Self-hosting & compute Weekly recap Read the recap →
NVIDIA's shipping Rubin has 19.2 TB/s of HBM4 bandwidth against the preliminary 22, so decode-heavy sizing needs redoing. Netflix lifted training-job success above 99.5% by treating a hung GPU check as a failure. (Rubin, Netflix)
-
Experimental models Weekly recap Read the recap →
SenseTime's SenseNova-Loop-9B runs most of its layers twice with shared weights, saving memory but not compute. (briefing)