VAST (GitHub) published this on Oct 18, 2026 (source: github.com).
VAST released EdenLive with open 5B and 14B model weights on Hugging Face and real-time streaming inference code on GitHub. Upstream code or a game engine keeps the world state, and EdenLive renders a reference video stream, guided by a target first frame and a text prompt, into detailed frames in styles from realistic to anime. In VAST's own tests, not independently verified, the 5B model runs at more than 24 frames per second at 720p on a single RTX 5090 and sustains minute-long generation. The release also includes a game-prototype-turbo skill that turns a text prompt into a coarse playable game, a frontend and a technical report.
Anthropic published this on Oct 18, 2026 (source: anthropic.com).
Anthropic is committing a portion of its shares to the Anthropic Public Benefit Fund, a new wholly owned LLC dedicated to the public benefit. Subject to stockholder approval, 5% of its shares are committed now, and the board may add 2% more for every 50% increase in the company's value above a set baseline. Combined with its co-founders' pledge to donate 80% of their equity and an employee equity donation match that about half of staff join, Anthropic says roughly 20% of its shares are now dedicated to the public good. The fund could make large donations to nonprofits or direct distributions to specific groups, for example to address AI's impact on workers, and none of its Anthropic equity will be used for lobbying or political giving. Anthropic says it will share the fund's governance and valuation milestones in the coming months.
MiniMax published this on Oct 19, 2026 (source: minimax.io).
MiniMax has officially released M3.1 Flash and says Space Bunny, the anonymous model that processed more than 200 trillion tokens, was an early version of it. M3.1 Flash is available now through the MiniMax API, M Plan and MiniMax Code, priced at $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. MiniMax says the model runs online inference nearly three times faster and cuts agentic-task usage costs by 70% versus its predecessor. MiniMax-reported scores include 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2 and 82% on CyberGym. MiniMax says the model weights will be released later.
SGLang (GitHub) published this on Oct 18, 2026 (source: github.com).
SGLang v0.5.22 bundles 950 pull requests from 245 contributors. The team says DeepSeek-V4.1 gets up to 2.3x more throughput per GPU on H200 and now runs on AMD MI350X and MI355X, a new InstantTensor loader loads models 10x faster, and PD disaggregation gets up to 27% more throughput with half the p95 time to first token on long prompts. New model support includes Cloudflare's Clef, Perplexity's PPLX-Decider, EmbeddingGemma 2, HunyuanImage-3.0-Instruct and Kandinsky 6, and SGLang now also runs on Apple Silicon Macs with the standard PyTorch backend. The release fixes a bug where hybrid linear models could resume from the wrong state when speculative decoding and structured output were used together, and now requires Python 3.11 or newer.
Statement by LenovoLenovo starts selling the YOGA Pro 15 RTX Spark AI PC in China from 29,999 yuan, with up to 128GB of unified memory · 3 days ago
Lenovo said this on Oct 16, 2026; not an official company announcement (source: qbitai.com).
Lenovo put the YOGA Pro 15 RTX Spark on sale across all channels on October 16 at a starting price of 29,999 yuan, in a company-provided article republished by QbitAI. The laptop uses NVIDIA's RTX Spark N1X superchip with up to a 6,144-core Blackwell RTX GPU and a 20-core Grace CPU, which Lenovo says delivers up to 1 petaflop of FP4 AI compute, plus up to 128GB of unified memory and 2TB of storage. Lenovo says the machine runs local models with more than 100 billion parameters, and that an offline local mode for its Tianxi AI agent with no token costs will arrive in a later OTA update.
Unconfirmedbrowsecomp-plus-mini, a 15-query subset of BrowseComp-Plus, ranks deep-research agents roughly like the full 830-query benchmark at under 2% of the cost
Not verified by an official source as of Oct 18, 2026 (source: reddit.com).
Developer asankhs published browsecomp-plus-mini on Hugging Face under an MIT license: 15 of the 830 BrowseComp-Plus queries, picked so that running only them ranks deep-research agents roughly the way the full benchmark does. Every answer is one to four words and is graded by normalized exact match, so no LLM judge is needed. Checked against 79 published runs from 18 systems, the author reports a weighted Spearman rank agreement of 0.88 averaged over 30 splits, versus 0.75 for a random 15 queries, and says runs more than 10 points apart on the full set keep their order about 96% of the time.
Unconfirmedchatllm.cpp adds support for InternLM's Intern-Decision models, text only for now
Not verified by an official source as of Oct 18, 2026 (source: reddit.com).
The developer of chatllm.cpp added support for InternLM's Intern-Decision models in 0.8B, 2B and 4B sizes, with text input only for now. The project's model list says the engine needs the --name Intern-Decision option to enable the decision feature. The developer notes that Intern-Decision fills every decision slot in a single diffusion-style run, instead of generating one answer per run as H2O-Lightning does, which generally makes decisions faster.
Reported by PoliticoAnthropic pitches lawmakers a draft AI safety bill that would let the government halt a model only over risks of $1 billion in damages or 50 deaths, per Politico · 3 days ago
Not confirmed by the company — reported on Oct 16, 2026 (source: politico.com).
Anthropic representatives have privately shown Capitol Hill lawmakers a draft federal bill on frontier AI safety, Politico reported on October 16, citing three people familiar with the effort. The government could block a model's release or use only for a foreseeable, imminent risk of more than $1 billion in damages or of killing or seriously injuring more than 50 people, arising from a biological weapon, a model evading its developer's controls or a cyberattack. Independent evaluators registered with the Commerce Department would periodically assess labs and must report such a risk within 72 hours; Commerce could then ask a federal judge in Washington, D.C. to halt the model, and no other agency could suspend it. An Anthropic spokesperson said the bill reflects the company's Advanced AI Framework.
UnconfirmedTempoPFN, a linear-RNN time series forecasting model pre-trained only on synthetic data, is accepted at NeurIPS 2026
Not verified by an official source as of Oct 18, 2026 (source: reddit.com).
TempoPFN, from Frank Hutter's AutoML group with Vladyslav Moroshan, Julien Siems, Arber Zela and Timur Carstensen, is a univariate time series foundation model built on a linear RNN with a GatedDeltaProduct architecture, pre-trained exclusively on synthetic data. Its data pipeline combines generators such as stochastic differential equations, Gaussian processes and audio synthesis with new augmentations. The paper reports top-tier zero-shot results on Gift-Eval, fev-bench and Chronos-ZS, beating all other synthetic-only approaches. The authors say the work is accepted at NeurIPS 2026, that the synthetic data has already been used in Toto-2 and TS-ICL, and they have released the data generation pipeline, training code and model weights.
UnconfirmedProof Before Reuse study: only showing a coding agent skills that later sessions reproduced keeps a poisoned skill out of its procedural memory
Not verified by an official source as of Oct 18, 2026 (source: reddit.com).
A small case study, with all data released, tests poisoning of an agent's procedural memory. A coding agent learns skills from sessions whose tests passed; one learning session is poisoned with a new file, an unrelated line, or one line fused into the real fix, followed by three clean sessions on the same task. The author reports that showing the agent only what later sessions reproduced keeps the poisoned skill from being reused.
Unconfirmedllama-doctor, a new open CLI and live dashboard, reads a llama.cpp server's metrics and grades what is wrong
Not verified by an official source as of Oct 18, 2026 (source: reddit.com).
A developer released llama-doctor, a read-only Python tool that points at a running llama.cpp or Unsloth Studio server and pulls its /metrics, /props and /slots endpoints. It returns severity-graded findings on prefill versus decode speed, queue pressure, prefix cache reuse and memory headroom, with exit codes mapped to severity so cron jobs can gate on them. It uses only the standard library and needs Python 3.10 or later.
OfficialFireworks AI raises its Small serverless rate-limit tier to models under 800B parameters, moving GLM 5.3 into it
Fireworks AI published this on Oct 18, 2026 (source: docs.fireworks.ai).
Fireworks AI's Small serverless model tier now covers models with fewer than 800B total parameters, up from 600B, per its docs changelog dated October 18. The Medium tier now spans 800B to under 1.6T parameters, which moves models such as GLM 5.3 into the Small tier. The rate-limit ceilings for each tier are unchanged; Fireworks doubled the Small tier's ceilings on October 7.
Reported by TechCrunchDisney agrees to offer jobs to most of Character.AI's workforce in its bankruptcy bid, per TechCrunch
Not confirmed by the company — reported on Oct 18, 2026 (source: techcrunch.com).
A court filing spotted by Bloomberg shows that Disney, the stalking-horse bidder in Character.AI's Chapter 11 sale, agreed to make employment offers to at least 60 Character.AI employees, the majority of the workforce, with an estimated $60 million in annual compensation, and that the vast majority of those offered jobs accepted. Because it is a stalking-horse bid, another company can still come forward with a higher offer. The filing says Character.AI, despite its large and engaged user base, was hobbled by high costs and potential litigation-related liabilities and regulatory engagement. Former Character.AI CEO Karandeep Anand recently became Disney's first CTO, per TechCrunch.
UnconfirmedWarcraftSim, a self-play agent trained on one desktop, beats Warcraft III's built-in insane AI in 89% of games
Not verified by an official source as of Oct 18, 2026 (source: reddit.com).
Developer Robin Kahlow says his WarcraftSim agent, which plays the full unmodified Warcraft III on the Echo Isles ladder map, went in three weeks from losing to the easy built-in AI to an 89% win rate against insane, the game's hardest level. It was trained by self-play on one desktop with an RTX 3090 and 16 CPU cores, playing 840,000 games, about 11 years of game time, and needs 13 ms per step, small enough to run on an ordinary CPU. The game runs under Wine with about 50 copies at once; a small DLL replaces its clocks with a virtual one so each game runs about 15 times faster than real time and waits for the agent at every step. The code is on GitHub.
OfficialGitHub Security Lab says its open-source Taskflow Agent found 24 vulnerabilities in Android apps, including a location leak in OsmAnd · 21 days ago
GitHub Blog published this on Sep 28, 2026 (source: github.blog).
GitHub Security Lab researcher Kevin Stubbings says Android-specific audit taskflows for the lab's open-source seclab Taskflow Agent have found and reported 24 vulnerabilities in Android apps. One lets any app, even with no permissions, silently import settings into the OsmAnd navigation app, which has over 10 million downloads, and leak the user's map location and routes to an attacker's server. Two chained bugs in the Wikipedia Android app allow an account takeover through a deeplink. He says LLMs are good at finding vulnerabilities but often misjudge severity, so every finding needs review by a researcher. The taskflows are open source, need a GitHub Copilot license and use premium model requests.
OfficialOllama adds DeepSeek-V4.1-Flash-1017 to Ollama Cloud with US-based hosting and zero data retention
@ollama on X published this on Oct 17, 2026 (source: x.com).
DeepSeek-V4.1-Flash-1017, the updated V4.1 Flash snapshot DeepSeek released on October 17, is now available on Ollama Cloud, hosted in the US with zero data retention. It runs with ollama run deepseek-v4.1-flash:1017-cloud, and Ollama's launch command can start it inside Claude Code or OpenCode. Ollama says the model is strong on agentic tasks while costing less per task.
OfficialGoogle Cloud AI Research open-sources Harness², a framework that improves an agent's harness for unseen tasks without changing model weights · 10 days ago
GitHub google-research/harness2 published this on Oct 9, 2026 (source: github.com).
Google Cloud AI Research has released Harness², an Apache 2.0 framework that refines the harness around a fixed model, meaning its system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins, for tasks it was not designed for. For each task it proposes harness edits, probes their effect on the same task and composes a refined harness from the differences between runs, without ground-truth feedback, then keeps the adopted harness and lessons in a library that guides later tasks. The authors report that a single step improved GLM-5 and Gemini 3.5 Flash over their base harnesses by up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench and 11.8 percentage points on LAB's partial pass rate, with further gains from a larger recursion budget. The code runs with OpenCode or Codex as the executor, and the repo notes it is not an officially supported Google product; the paper has not yet been peer-reviewed.
OfficialOpenAI's desktop app now lets you edit previous messages within a thread, Tibo says
@thsottiaux on X published this on Oct 19, 2026 (source: x.com).
Tibo (@thsottiaux) said on Day 14 of his daily updates that users can now edit previous messages within a thread in the desktop app, calling it a small quality-of-life improvement that had been requested for a long time and took too long to ship.
OfficialEpic Games' Unreal Engine skills plugin lets Claude Code control Unreal Editor over MCP · 32 days ago
GitHub (EpicGames) published this on Sep 17, 2026 (source: github.com).
Epic Games' Unreal Engine Skills for Claude Code plugin, listed in Anthropic's official Claude Code plugin marketplace, connects Claude Code to Unreal Editor through the Model Context Protocol. It exposes hundreds of tools across more than 30 toolsets, including actors, Blueprints, materials, Niagara, Control Rigs, Sequencer, State Trees, widgets, the Gameplay Ability System and automation testing, and adds an unreal-mcp skill plus a session hook that tells Claude it is in an Unreal project. It requires Unreal Editor running with the ModelContextProtocol and AllToolsets plugins enabled and the MCP server started. The plugin repo has been public since June and was last updated on Sep 17.
Reported by IT HomeSugon 8000, described as China's first fully domestic 100,000-card AI supercluster, is shown at a Shenzhen HPC conference, per IT Home
Not confirmed by the company — reported on Oct 18, 2026 (source: ithome.com).
Citing CCTV News, IT Home reports that the Sugon 8000, described as China's first fully domestic AI supercluster linking 100,000 AI accelerator cards and completed this year, was shown at the 2026 National High Performance Computing Academic Conference in Shenzhen, where ten application results were set to be announced. The report says the system is built to combine supercomputing and AI computing on the same infrastructure, running high-precision scientific computing as well as large-model training and inference. IT Home notes that a first fully domestic 100,000-card AI supercluster was announced as entering service in August, per IT Home.
UnconfirmedThree 'Aurora' alpha image models briefly appear in OpenAI's API and turn up in Arena tests, unconfirmed by OpenAI, per TestingCatalog
Not verified by an official source as of Oct 18, 2026 (source: testingcatalog.com).
TestingCatalog reports that three model references, aurora-red-alpha, aurora-green-alpha and aurora-blue-alpha, briefly appeared in OpenAI's API and were removed shortly after, then turned up generating images in blind Arena tests. LuminaBench says it tried the red and green variants in Arena and got cat portraits from both; both outlets guess this could be GPT-Image-3, which they say is speculation. OpenAI has not confirmed the models.
UnconfirmedAnthropic appears to be preparing a 'Money' tab in the Claude mobile app for linking bank accounts, per TestingCatalog · 35 days ago
Not verified by an official source as of Sep 14, 2026 (source: testingcatalog.com).
TestingCatalog reports that unreleased interface elements in Claude's iOS app show a separate Money tab whose onboarding says users can link bank accounts and ask Claude about spending, plans and other financial questions. No equivalent tab has appeared on Claude's web interface, and the data provider, supported account types and launch timing are unknown. Anthropic has not announced the feature. TestingCatalog first reported this on Sep 14.
Statement by SimilarwebSimilarweb ranks OpenAI's Tibo (@thsottiaux) as the most popular X account page by traffic share in September 2026
Similarweb said this on Oct 18, 2026; not an official company announcement (source: x.com).
Similarweb posted its list of the most popular account pages on X by traffic share for September 2026, with @thsottiaux first, ahead of @FabrizioRomano and @elonmusk.
UnconfirmedClaude Opus 5.5's 3D-printed egg capsule survived a 20-foot drop in Roberto Nickson's six-model Agent Wars test · 3 days ago
Not verified by an official source as of Oct 16, 2026 (source: x.com).
In episode 3 of his Agent Wars series, Roberto Nickson asked six AI models to design and engineer an egg capsule for 3D printing, using at most 250 grams of the same filament, under 8 hours of print time and up to 20 cotton balls. He reports that Claude Opus 5.5's capsule kept the egg unbroken at 20 feet, GPT-6 Astra's egg cracked at 8 feet, Gemini 4 Argon and Grok 4.7 showed slight cracks at 4 feet, Muse Wonder's egg cracked at 4 feet and Qwen 3.8-Max's was destroyed at 4 feet. Each model got one design and one print with no iterations. Claude also won his earlier 3D-printed bridge episode. It is a single informal test, not a controlled benchmark.
OfficialPaper replaces a small language model's learned embedding table with a frozen 16 MB vocabulary and reports matching or better benchmark scores
Posted on arXiv on Oct 19, 2026 (source: arxiv.org).
Sai Kiran Bathula's paper swaps the embedding table of a Gemma 3 270M-shaped model, which holds 168M of its 268M parameters, for one frozen table of 512-bit codes derived from a pretrained embedding, and trains the same body from scratch on 2.05B tokens with and without it. The paper reports that the learned table buys 0.20 nats of validation loss for 2.6 times the trainable parameters, but the frozen-table model is better or tied on LAMBADA, HellaSwag and PIQA, with only ARC-Easy favoring the learned table. With a ternary BitNet b1.58 body the whole model is a 46 MB file that the author says generates 615 to 740 tokens per second on a laptop CPU in 62 MB of memory. Code, data, checkpoints and every evaluation answer are released on Hugging Face. The paper has not yet been peer-reviewed.
OfficialResearchers train GPT-2-small models on 96 synthetic languages to test how language models learn action nominal constructions
Posted on arXiv on Oct 19, 2026 (source: arxiv.org).
Iskar Deng and Shane Steinert-Threlkeld built 96 synthetic target languages that vary clause and noun-phrase word order, alignment, complement systems and action nominal construction strategy, trained a GPT-2-small model from scratch on each, and tested them with minimal pairs covering 34 grammatical phenomena. The authors report typological parallels in basic clause configurations but divergences in complement systems and nominal strategies, with models also learning combinations that are rare or unattested in natural languages. They say this suggests natural language distributions reflect both general learning pressures and functional demands that model learning experiments do not fully capture. The paper has not yet been peer-reviewed.
OfficialRealCLI benchmark tests CLI agents against a scripted user who changes instructions mid-task, and the best setup averages 49.2%
Posted on arXiv on Oct 19, 2026 (source: arxiv.org).
RealCLI is a paired benchmark for command-line agents: RealCLI-Exec gives the full specification upfront, while RealCLI-Collab adds a scripted user who delivers information step by step, corrects the agent, revises constraints and makes authorization decisions. Collaboration is scored on five dimensions broken into 22 sub-dimensions and 51 interaction scenario templates. The paper reports that the best configuration averages only 49.2% and peaks at 37.6% on the joint collaboration score, and the authors say frontier agents consistently struggle with steerability, trust calibration and safety boundaries. The paper has not yet been peer-reviewed.
OfficialAuditKIT, a source-available evaluation library from Lexsi Labs, puts benchmark, judge, runtime, tool-use and retrieval results in one format
Posted on arXiv on Oct 19, 2026 (source: arxiv.org).
AuditKIT, with code on Lexsi Labs' GitHub, stores benchmark accuracy, judge-based scores, runtime performance, tool-use behavior and retrieval quality in a single result representation, with direction-aware metric contracts and configuration fingerprints. For agents it matches tool calls turn by turn, separates parallel from sequential execution and keeps verified task outcomes apart from trajectory diagnostics, marking missing evidence as unscored rather than failed. The authors present seven case studies that they say show how answer extraction, runtime behavior and metric choice change evaluation outcomes. The paper has not yet been peer-reviewed.
OfficialImageEval 2026 entry detects multimodal hallucinations zero-shot with a 7B vision-language model and rotated answer choices
Posted on arXiv on Oct 19, 2026 (source: arxiv.org).
The alkhder team describes its entry to Subtask 1b of ImageEval 2026, where a system must pick the one visually valid statement out of three paired with an image, in English and Modern Standard Arabic. It uses a pretrained 7B vision-language model zero-shot, picks the answer from next-token log-probabilities and rotates the candidates in a Latin square to reduce answer-position bias. The authors report CI scores of 0.0510 on the English test set and 0.0960 on the Arabic one, which they say cuts CI by 83.7% and 80.4% relative to the official 3B baselines on devtest. The paper has not yet been peer-reviewed.
OfficialPaper finds physics-labeled prompts make an LLM a strong early guess for engineering simulator calibration, best used as a warm start
Posted on arXiv on Oct 19, 2026 (source: arxiv.org).
Ruo-Qian Wang tests whether a language model, given physical descriptions of a simulation's errors, can propose parameter updates more efficiently than derivative-free optimizers on a cylinder-wake large-eddy simulation and a damped-oscillator inverse problem. The paper reports that the first physics-informed proposal improved the initial design in 23 of 25 trials and that median calibration error after five evaluations was 21.9 times lower than physics-free controls, but the advantage was gone by fifteen evaluations and runs varied 9.8-fold. The author says a hybrid that hands the model's proposals to a conventional optimizer reached target accuracy in 9 simulator runs, versus 12 for the model alone, 17 for Nelder-Mead and more than 60 for Bayesian optimization. The paper has not yet been peer-reviewed.
OfficialPaper teaches a masked diffusion language model to distrust corrupted context, with a finetune that runs in under five hours on a laptop GPU
Posted on arXiv on Oct 19, 2026 (source: arxiv.org).
Gunmay Jhingran mixes absorbing and uniform-substitution noise with region-dependent rates so a discrete diffusion model sees observed but uncertain tokens during training, and adds a sampler that re-injects noise into suspicious context at inference. An 8,000-step finetune of a pretrained 170M-parameter masked diffusion model took under five hours on a 6 GB laptop GPU. For uniform-random corruption the paper reports 0.96 recall and 0.91 precision at detecting corrupted context and infill token-F1 gains of 3.8 points at 10% corruption and 6.4 points at 20%, with a false-edit rate below 0.4% on clean input, while gains for plausible corruptions were about 0.5 points and not statistically significant. The paper has not yet been peer-reviewed.
OfficialLean4AgenticRL trains 9B and 27B agent models with Lean4-verified process rewards, and the authors report 77.3% on TerminalBench-2.1
Posted on arXiv on Oct 19, 2026 (source: arxiv.org).
Ruida Wang, Gerald Friedland, Tong Zhang and Chris Kong propose Lean4AgenticRL, which formalizes an agent's problem-specific workflow in Lean4 so tool calls and workflow stages can be verified and rewarded during reinforcement learning, not just final outcomes, using a workflow-stage-normalized GRPO. They use it to train LAR-9B and LAR-27B. The paper reports that LAR-27B scores 77.3% on TerminalBench-2.1 and that the models beat their base models by 9.89% on average and the next-best model of comparable size the authors know of by 5.07%. The paper has not yet been peer-reviewed.
OfficialPaper finds recursively self-distilling a safe model on its own safe outputs can make it progressively misaligned
Posted on arXiv on Oct 19, 2026 (source: arxiv.org).
Karim Abdel Sadek, Federico Di Gennaro and Fanny Yang study recursive self-distillation, where each self-distilled model becomes the teacher for the next round. They report that a safe model repeatedly distilled on its own safe generations while given a harmless extra cue becomes increasingly misaligned across many harmful queries, a failure they call progressive misalignment, shown in the Qwen3 and Phi-4 families on SORRY-Bench and StrongREJECT. The authors say it also happens when the cue is replaced by feedback from a simulated well-meaning user, and that a gradient-based search finds benign prompts that trigger it, so harmful training prompts are not needed. The paper has not yet been peer-reviewed.
OfficialBaidu Qianfan adds DeepSeek-V4.1-flash-1017 to its model list with a 1M-token context
Baidu Qianfan docs published this on Oct 18, 2026 (source: cloud.baidu.com).
Baidu's Qianfan model list now includes DeepSeek-V4.1-flash-1017 (model parameter deepseek-v4.1-flash-1017) alongside DeepSeek-V4.1-flash. Baidu lists a 1M-token context and input limit, up to 393,216 output tokens and default rate limits of 60 requests and 1,000,000 tokens per minute.
OfficialFireworks AI adds NVIDIA Nemotron 3.5 Super VL as a training model with up to 262K context
Fireworks AI published this on Oct 18, 2026 (source: docs.fireworks.ai).
Fireworks AI says NVIDIA Nemotron 3.5 Super VL 120B A12B is now available for training on all three of its surfaces, with up to 262K context: Managed Training (LoRA SFT and DPO), the Dedicated Training API (LoRA and full-parameter training, including RL) and Serverless Training (LoRA on the shared pool, billed per token), per its docs changelog.
OfficialMiniMax-M3.1-Flash now listed in MiniMax API docs
MiniMax API docs published this on Oct 19, 2026 (source: platform.minimax.io).
MiniMax's API documentation now lists MiniMax-M3.1-Flash, without the Preview tag, as its latest M-series language model, ahead of MiniMax-M3 in the supported models list. The docs describe it as a frontier multimodal coding model for agentic reasoning, tool use, coding and long-context tasks, with a 1M-token context window and thinking depth that can be tuned with an effort setting, accessible over HTTP, the Anthropic SDK or the OpenAI SDK. M3.1-Flash-Preview launched on September 27 on MiniMax Code and the Token Plan; MiniMax has not published a dated announcement for the API listing.
OfficialQwen Code 0.25.3 lets hosted agents launch read-only child agents and adds SSH, macOS and Kubernetes workspace runtimes
QwenLM published this on Oct 19, 2026 (source: github.com).
QwenLM's qwen-code v0.25.3 release again centers on managed agents. Hosted parent agents can now launch snapshot-isolated child agents that read a frozen copy of the workspace without changing files or running commands, and managed sessions gain an SSH tool runtime, a native macOS workspace runtime, resident Kubernetes workspaces that keep files across sessions, Plan mode with inline attachments, and per-run automation cancellation. The web shell adds a recent-executions panel with live timings, new file and folder actions with delete confirmation, side-task archiving and an approval-mode allowlist, and QwenCloud is now the first sign-in option in /auth. The release notes list no breaking changes.
OfficialKapitsa, an open-source inference engine for Apple Silicon Macs from Nace.ai developers, supports only Qwen and Nemotron models
Kapitsa (GitHub) published this on Oct 19, 2026 (source: github.com).
Kapitsa is an Apache 2.0 inference engine for Apple Silicon Macs built by people at Nace.ai, who say they made it to run coding agents on local models inside enterprise networks with no internet access. It supports only two model families, Qwen (3.5 4B, 3.8 27B, 3.6 35B-A3B and, on 128 GB Macs, 3.8 Flash-Next) and NVIDIA Nemotron (3.5 Lightning, Bolt and Super-VL, plus Dory 4B), so the team can tune Metal kernels for each model, and it serves OpenAI- and Anthropic-compatible APIs and can launch OpenCode, Claude Code or Codex against a local model. The team's own benchmarks on a 48 GB M4 Pro MacBook Pro report 154 tokens per second for Qwen3.6 35B-A3B versus 127 for oMLX and 89 for Ollama, and 64 for Nemotron 3.5 Bolt versus 35 for oMLX. The developers say not to trust benchmarks and ship a kapitsa bench command so users can measure their own Macs.
OfficialZhipu (Z.AI) paper reveals internal unified model GLM-Omni · 4 days ago
Posted on arXiv on Oct 15, 2026 (source: arxiv.org).
Researchers from Zhipu (Z.AI) and Tsinghua University, with Jie Tang as corresponding author, describe GLM-Omni, an in-house unified understanding-and-generation model that has not been released, in an arXiv paper. It is initialized from the GLM-4.1V-Thinking pretraining base and uses a discrete image tokenizer based on X-Omni. The paper asks whether image-generation training teaches facts a model can then use in text, testing GLM-Omni alongside BAGEL and SenseNova-U1 with a Knowledge Transfer Probing protocol over seven knowledge categories of fictional entities. For GLM-Omni, text-only QA accuracy was 14.57% with no alignment (the baseline), 12.57% with text-only alignment and 26.86% with interleaved image-text alignment, falling to 13.71% when the image loss was masked; visual chain-of-thought raised it to 41.14%. Without alignment, it reproduced 54.00% of the facts correctly in generated images. The dataset and protocol are on GitHub.
OfficialInco AI releases a DFlash 2 speculative-decoding draft model for Qwen3.8-Flash-Next
Inco AI (Hugging Face) published this on Oct 19, 2026 (source: huggingface.co).
Inco AI published Qwen3.8-Flash-Next-DFlash2 on Hugging Face under Apache 2.0, a block-diffusion draft checkpoint that speeds up serving of Qwen/Qwen3.8-Flash-Next through DFlash 2 speculative decoding in SGLang. It is not a standalone LLM: it drafts a whole block of tokens in one pass, keeps top candidates per position, and uses a lightweight path selector plus two-tap dynamic convolutions so greedy output matches the target and sampling keeps its distribution. On one NVIDIA GB300 (TP1), the authors report acceptance lengths of about 2.83–3.26 tokens per verification step versus 2.50–3.08 for Flash-Next's built-in MTP, and concurrency-1 throughput around 1.63–1.82× autoregressive (beating MTP on every listed task); at concurrency 32 they report up to about 1.73×. Serving currently needs a small target-feature capture patch shipped in the repo until upstream SGLang lands support.
Reported by Upstarts MediaStripe says Link agent purchases rose 38x in September as about 200 agents shop through it, per Upstarts Media
Not confirmed by the company — reported on Oct 19, 2026 (source: upstartsmedia.com).
Stripe shared September Link data with Upstarts Media showing agentic transactions up 38-fold and payment volume up 20-fold versus the prior baseline, from a base of about 70 agents then. Dan Hill, Stripe's head of consumer product for Link, told Upstarts the network now powers about 200 agents (including Muse, Grok Bot and Instinct), that most agent spends are about $100 while some book flights or buy furniture for thousands, and that retail led with 36% of transactions and 32% of volume. Absolute purchase counts and dollar totals were not disclosed; Hill said volume is becoming 'very material' while agents remain a small share of Link's ~300 million users.
Reported by The InformationExecutives expect a neocloud shakeout that could hit Nvidia's non-hyperscaler sales, per The Information
Not confirmed by the company — reported on Oct 19, 2026 (source: theinformation.com).
After Firmus scrapped its IPO and CoreWeave and peers sold off, two data-center and neocloud executives told The Information only seven to ten of the hundreds of AI cloud providers will last. Nvidia's July-quarter sales outside hyperscalers were about $40 billion — 45% of data-center revenue — and the chipmaker is backing partners with equity, chip allocations, loan guarantees and site-finding help. Related reports cited in the Decoder write-up include Microsoft pulling out of an Nscale deal for 23,000 Nvidia chips in Loughton and Gartner projecting that 80% of independent neoclouds will cease independent operations by 2030.
Daily AI news, sourced