Alibaba releases Qwen3.8-Max weights as DeepSeek and xAI ship flagships the same day

Three flagship models shipped inside 24 hours: Alibaba’s 2.4-trillion-parameter Qwen3.8-Max open weights, DeepSeek V4 Pro 0813 reaching general availability, and xAI’s Grok 4.6. Separately, CloudSEK published exposure data from the March compromise of LiteLLM’s PyPI packages, putting more than 2,500 companies and roughly 434,000 CI/CD pipelines in scope for credential rotation. Google Research reported that frontier models encode 95-98% of the facts they are tested on yet fail to recall 26-34% of them, which reframes a large share of factual error as a retrieval problem rather than a knowledge gap.

Model Releases #

Qwen3.8-2.4T-A95B #

Qwen / Hacker News (637 points)

Alibaba published open weights for Qwen3.8-2.4T-A95B: 2.4 trillion total parameters with 95 billion active, 92 layers, a 512-expert MoE with 11 experts routed per token, and a hybrid of Gated DeltaNet and gated attention. Native context is 262,144 tokens, extensible to just over one million, and the card reports 86.6 on Terminal Bench 2.1, 67.7 on SWE-bench Pro and 92.6 on GPQA Diamond, with reasoning depth selectable through a reasoning_effort parameter. This is the first time a Qwen-Max-class model has been released publicly rather than kept behind the API, and the weights run on stock vLLM and SGLang. The benchmark column is Alibaba’s own and has no third-party replication yet; the 2.4T/95B shape also means the file is far beyond single-node territory, so “open” here is a claim about inspectability and licensing more than about who can actually run it.

DeepSeek V4 Pro 0813 #

OpenRouter / Simon Willison / Hacker News

DeepSeek’s V4 Pro left preview and reached general availability at $0.435 per million input tokens and $0.87 per million output, with a one-million-token context window. Against the preview build the reported jumps are large — DeepSWE 12.8 to 62.7, CyberGym 52.7 to 83.3, Terminal Bench 2.1 72.1 to 87.9 — and the 0813 figures put it at 80.6 on SWE-bench Verified, level with Gemini 3.1 Pro and marginally behind Claude Opus 4.6. Two caveats travel with this one. None of the 0813 numbers has been reproduced by an independent evaluator, and as Simon Willison noted, DeepSeek published no announcement page at all, so the OpenRouter model listing is currently the closest thing to a primary source for a model in this weight class.

Grok 4.6 #

xAI / Hacker News

xAI shipped Grok 4.6 on the same 1.5-trillion-parameter foundation as 4.5, attributing the gains entirely to post-training: regenerated supervised fine-tuning trajectories, curated data and expanded reinforcement learning on agentic tasks. It scores 61 on the Artificial Analysis Intelligence Index against 56 for 4.5, with 69.9% on CursorBench v3.2, 65.9% on DeepSWE v1.1 and 61.3% on FrontierCode v1.1, priced at $2 per million input and $6 per million output. It is available today in the API, Cursor, Grok Build, OpenRouter, Vercel and Cloudflare. The interesting part is the control condition: holding the base model fixed and moving five points on a composite index is a fairly clean demonstration of how much headroom is still sitting in post-training rather than scale.

Putting sign language AI into users’ hands #

Google DeepMind

DeepMind shipped SL2T, a sign-language-to-text model that runs MediaPipe Holistic pose tracking on-device and sends landmark coordinates rather than video frames to a server for translation, skipping the intermediate gloss annotations most prior systems depend on. It reports a zero-shot 70 BLEURT on FLEURS-ASL, which the team says exceeds any previously published score, from training on over 100,000 hours across more than 50 sign languages with roughly a quarter of that in ASL. It launches as ASL-to-English in Gboard and Live Transcribe on Pixel 11 at no extra cost. The landmark-only transport is the architecturally interesting choice — it makes the privacy story defensible and cuts bandwidth, at the cost of discarding everything the pose extractor does not encode, which is where the acknowledged failures on rapid fingerspelling and rare signs likely live.

Security #

2,500+ companies and 434,000 CI/CD pipelines exposed in the LiteLLM supply chain breach #

CloudSEK / Ars Technica

CloudSEK published exposure data from the March 2026 compromise of LiteLLM, in which the group tracked as TeamPCP poisoned the Trivy security scanner using an automation token that had been rotated but never revoked, then rode Trivy’s unpinned presence in LiteLLM’s CI pipeline into PyPI releases 1.82.7 and 1.82.8. Those packages carried a .pth file executing at interpreter startup and a stealer that harvested SSH keys, AWS, GCP and Azure credentials, Kubernetes tokens, .env files and AI provider keys, exfiltrating to attacker infrastructure and public GitHub repositories. The packages were live for 40 minutes; the leaked corpus runs to 153GB from 118,829 pipeline runs across 2,488 domains, and the FBI issued a FLASH advisory on 2 July. The gap between a 40-minute exposure window and a five-month tail is the whole lesson: the artifact is trivially recalled and the credentials it copied are not, so the only closing action is rotation of everything the affected processes could reach.

Mass vulnerability scans are impersonating AI crawlers #

Known Agents / Hacker News (278 points)

Known Agents reports an active campaign spoofing AI crawler user agents — ClaudeBot, ChatGPT-User, Googlebot — to scan sites for paths associated with AI development tooling, including /.aws/credentials, /.env.local, /serviceaccountkey.json and /terraform.tfstate. The report gives baseline spoofing rates per agent (0.5% of Googlebot traffic, 0.1% each for ChatGPT-User and ClaudeBot) but publishes no campaign volume, observation window or verification methodology, directing readers to contact the company instead. Treat the rates as unverified. The operationally useful part needs no numbers: sites that allowlist AI crawlers by user-agent string alone are allowlisting anyone willing to type it, and the target paths say plainly that the attacker’s model of a modern repository includes an agent’s credential files.

Research & Papers #

Empty shelves or lost keys? Recall is the bottleneck for parametric factuality #

Google Research

Google Research evaluated 13 models on WikiProfile, 2,150 Wikipedia-derived facts each paired with 10 tasks separating encoding, recall and recognition, generating roughly 4.5 million responses. Gemini 3 Pro and GPT-5 encode 95-98% of the facts — they can reproduce them in pre-training-like contexts — yet fail to directly recall 26-34% of them when asked in varied phrasings, and thinking-optimised models recover 40-65% of that encoded-but-unrecalled set. The consequence for anyone shipping a RAG or grounding layer is that a measurable fraction of what looks like missing knowledge is retrieval failure inside the weights, which scaling addresses poorly because encoding is already near saturation. It also puts a ceiling on what “the model doesn’t know that” can be assumed to mean during error triage.

Agent Skills Can Be Harmful: An Empirical Study of Skill-Induced Failures in LLM Agents #

arXiv

The authors attribute agent failures to specific loaded skills by diffing a skill-guided run against a no-skill or semantically-matched reference run on the same task, and find 307 skill-induced regressions across SkillsBench and SWE-Skills-Bench: 125 functional failures and 182 efficiency regressions. The two findings that bite are that functional failures are rarely caused by obviously irrelevant skills — plausibly relevant ones lead the agent to implement task requirements incorrectly or omit them — and that efficiency regressions are not explained by prompt length, with excessive verification (67 cases) and heavy implementation pipelines (30 cases) the largest single sources. If you maintain a skills library, this says the audit that matters is not “is this skill relevant” but a paired A/B against no skill at all, because the harmful cases are exactly the ones that pass a relevance check.

Total Recall at What Cost? Benchmarking the Serving Cost of Agentic Memory Systems #

arXiv

The paper benchmarks Mem0, Hindsight and Mastra Observational Memory against a fixed-size rolling window and full-transcript resubmission across two backbones, conversations up to 400 turns, and 665 LoCoMo questions, pairing every cost measurement with accuracy. A regression that predicts the two reference strategies from conversation length and message size well misses the memory systems by 18-69%, because their cost is driven by internal memory behaviour rather than transcript size; break-even against just resending the transcript ranges from the first few dozen turns for the cheapest system to never within 400 turns for the most expensive. Accuracy spans 21-54% and no system leads on both axes. The practical reading is that a memory layer is not automatically a cost saving, and the backbone you pair it with moves the bill about as much as the memory system does.

The Wording Effect: Quantifying Two-Way Drift in LLM Benchmark Performance #

arXiv

BenchDrift generates meaning-preserving rephrasings of benchmark problems along linguistic, referential, pragmatic and structural axes and measures how often correctness flips, across eight models on GSM8K, MMLU and MATH-Hard. Two results stand out: phrasing sensitivity does not diminish as models improve but changes sign — weak models gain more from rephrasing than they lose while strong models lose far more than they gain, so the top scorers on a benchmark are the ones whose scores depend most on the exact wording they were handed — and models largely agree on which rephrasings are costly, meaning the fragility belongs to the phrasing rather than to any model. The abstract reports directions and mechanisms but no drift magnitudes, so the effect size has to be taken from the paper body. For leaderboard consumers the implication is direct: a single-phrasing score for a frontier model is a point estimate with an unquantified and asymmetric error bar.

Developer Tools #

Delta #

Zed / Hacker News (562 points)

Zed announced Delta, a multiplayer environment for running coding agents and reviewing what they produce, built on DeltaDB — a replicated store that syncs conversations and code edits in real time and captures them between git commits rather than at them. Comments anchor to code and follow it as it changes, conversations are editable documents rather than linear message logs, and third-party agents including Claude Code can sync terminal sessions into Delta threads. The same Rust application compiles to WebAssembly, so it runs natively or in a browser over WebGL. No benchmarks, no pricing, and it is private beta with invitations rolling out gradually. The bet worth noting is on the review surface rather than the generation surface: the claim is that the unit of collaboration in agent-heavy work is the pre-commit trajectory, which is precisely the artifact git throws away.

Funding & Business #

Cognition reportedly in talks to raise at a $40B valuation #

Bloomberg / TechCrunch

Bloomberg reports Cognition is in early discussions with investors at a $40 billion valuation, roughly three months after closing $1 billion at $26 billion in May. The reported basis is a $492 million annualised run rate as of that May round, 50% month-over-month usage growth among enterprise customers, and an expectation of reaching $1 billion annualised. CEO Scott Wu continues to position Devin against routine work — legacy updates, platform migrations — rather than as a headcount substitute, with Mercedes-Benz, NASA and Goldman Sachs named as customers. A 54% valuation step-up on a round that has not finished aging is worth reading as a statement about capital availability in coding agents as much as about Cognition.

Lovable raises $400M Series C at $13.3B #

Lovable / TechCrunch / Hacker News

Lovable closed a $400 million Series C at a $13.3 billion valuation, co-led by Menlo Ventures and EQT’s Scaleup Europe Fund, with Tencent, Balderton, Kaszek, Regent and returning investors Accel, CapitalG, DST Global and Salesforce Ventures participating. The company reports over 60 million projects created since its November 2024 launch, more than 900 million monthly visits to Lovable-built apps, presence in nearly two-thirds of the Fortune 500, and plans to reach roughly 450 employees by year end. TechCrunch’s June figure of $500 million annualised run rate is the number to anchor on, since the announcement supplies project and visit counts instead — engagement metrics that do not distinguish a shipped product from an abandoned prototype.

OpenAI-backed Thrive Holdings raises $2B #

TechCrunch

Thrive Holdings raised $2 billion at a $12 billion valuation from SoftBank, D1 Capital Partners and Altimeter Capital. The structure is the story: rather than selling software to service businesses, Thrive acquires them outright — accounting and professional-services firms — and applies AI internally, which converts model capability directly into operating margin instead of into seat licences. It is a different exposure to the same bet the application layer is making, and one that does not depend on convincing an incumbent to change how it works.

Open Source #

OpenWALDO aims to blow the doors off proprietary AI training models #

The Register

OpenWALDO — Open Weights, Artifacts, Licenses, Data, Origins — is an effort led by Gregory Kurtzer, founder of CentOS and Rocky Linux, and funded by his company CIQ, to assemble a training corpus whose provenance is fully inspectable. It currently holds 167.3 billion reference tokens drawn from government records, open-access academic papers, mailing lists and public domain literature, shipped with a bill of materials so a company can treat it as a verified baseline, add proprietary data and build. Against the trillions of tokens behind frontier models that is a rounding error, and CIQ did not answer The Register’s question about whether anyone has trained on it yet. The premise is nonetheless the right one for anyone who has tried to answer a procurement question about training data: the artifact that is missing from open-weight releases is not the weights, it is the manifest.

Other #

Twitch content has trained Amazon AI for years, but users can opt out now #

Ars Technica / TechCrunch

Twitch confirmed that streamers’ broadcasts, clips, VODs, highlights, chat and images default to being available for Amazon generative AI training, and introduced an opt-out. Twitch CPO Mike Minton’s explanation on a livestream was unusually direct: “If this was opt-in, nobody would opt in. That’s honestly the answer.” The platform also says the content has been used this way for years, making the announcement a disclosure and a control rather than a policy change. It is a clean statement of the economics every UGC platform faces — consent designed to be technically real and practically unexercised — and it lands in the same month EU AI Act transparency rules began applying to generated content.

Threads to Watch #

A three-flagship day, and only one of them came with a page you could cite. Alibaba shipped a full model card with architecture, benchmarks and a license; xAI shipped an announcement post with a benchmark table; DeepSeek shipped nothing, leaving an OpenRouter listing as the primary source for a model claiming 80.6 on SWE-bench Verified. None of the three has independent replication. The gap is widening between how fast frontier-class models arrive and how fast anyone outside the lab can check them, and the practical consequence for a buyer is that the evaluation you can actually trust is increasingly one you run yourself on your own tasks. Grok 4.6 is the cleanest data point in the batch precisely because it holds the base model fixed: five index points from post-training alone says the frontier is still partly a training-recipe problem rather than a compute one.

Credentials are the payload, and agent tooling is where they now live. The LiteLLM stealer went after SSH keys, cloud credentials, Kubernetes tokens and AI provider keys, and the scanning campaign impersonating ClaudeBot is probing for /.aws/credentials, /.env.local and /serviceaccountkey.json. Both treat the AI development stack as a credential store, which by now it is: an agent framework holds provider keys, cloud roles and repository access in one process, and each new integration widens the blast radius of a single compromised dependency. The LiteLLM chain is worth tracing in full because nothing in it was novel — an un-revoked token, an unpinned build dependency, a 40-minute publish window — and it still reached 2,488 domains. Rotation discipline, not detection, is what bounded the damage, and the five months between compromise and exposure data is how long an organisation had to be doing it without knowing why.

The measurements underneath agent engineering keep failing in the same direction: they flatter. A benchmark score is a single phrasing whose top performers are the models most dependent on that phrasing. A skills library produces 307 attributable regressions, and the harmful skills are the ones that pass a relevance check. A memory system’s serving cost cannot be predicted from conversation length and may never break even against just resending the transcript. In each case the default reading of the artifact — the leaderboard row, the relevant-looking skill, the memory layer marketed as an efficiency — is optimistic, and the correction requires a paired comparison the tooling does not run for you. The common remedy across all three papers is the same and it is cheap: measure against the null option. No skill, no memory system, a different phrasing. That is also the comparison vendors have the least reason to publish.