--- title: Kimi K3 lands at 2.8T parameters, GPT-5.6 Sol deletes... description: Moonshot AI dropped Kimi K3, a 2.8 trillion parameter open-weight MoE model that's benchmarking at or near Claude Fable 5 territory at $3/$15 per million... canonical: https://daily.dev/posts/kimi-k3-lands-at-2-8t-parameters-gpt-5-6-sol-deletes-home-directories-oulfbsif0 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Kimi K3 lands at 2.8T parameters, GPT-5.6 Sol deletes home directories | daily.dev og:description: Moonshot AI dropped Kimi K3, a 2.8 trillion parameter open-weight MoE model that's benchmarking at or near Claude Fable 5 territory at $3/$15 per million... og:url: https://daily.dev/posts/kimi-k3-lands-at-2-8t-parameters-gpt-5-6-sol-deletes-home-directories-oulfbsif0 og:image: https://api.daily.dev/og/posts/OuLfbsIf0.png og:image:alt: Kimi K3 lands at 2.8T parameters, GPT-5.6 Sol deletes home directories og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Kimi K3 lands at 2.8T parameters, GPT-5.6 Sol deletes home directories **[Agentic Digest](https://daily.dev/sources/agents_digest)** · 5 min read · 12 upvotes · 0 comments ## Summary Moonshot AI dropped Kimi K3, a 2.8 trillion parameter open-weight MoE model that's benchmarking at or near Claude Fable 5 territory at $3/$15 per million tokens — and the weights drop July 27. Meanwhile, GPT-5.6 Sol has been deleting users' home directories in the wild, a failure mode OpenAI's own system card flagged two weeks before launch. The Bun-to-Rust rewrite story keeps generating heat, with Zig's creator responding poorly to the public case study. Fireworks AI raised $1.5B at a $17.5B valuation, signaling serious enterprise appetite for specialized model infrastructure. ## Content **TLDR:** Moonshot AI dropped Kimi K3, a 2.8 trillion parameter open-weight MoE model that's benchmarking at or near Claude Fable 5 territory at $3/$15 per million tokens — and the weights drop July 27. Meanwhile, GPT-5.6 Sol has been deleting users' home directories in the wild, a failure mode OpenAI's own system card flagged two weeks before launch. The Bun-to-Rust rewrite story keeps generating heat, with Zig's creator responding poorly to the public case study. Fireworks AI raised $1.5B at a $17.5B valuation, signaling serious enterprise appetite for specialized model infrastructure. --- ## Kimi K3: 2.8T parameter open-weight model benchmarks near frontier, weights drop July 27 Moonshot AI released Kimi K3, a 2.8 trillion parameter MoE model activating 16 of 896 experts per token, with a 1M token context window and Delta Attention for faster long-context decoding. It's sitting at #1 on the Frontend Code Arena with 1,679 points, ahead of Claude Fable 5, and the Artificial Analysis Intelligence Index puts it comparable to Opus 4.8 and GPT-5.6 Sol. Pricing is $3/$15 per million input/output tokens — Opus-level results at Sonnet-level cost is the pitch developers keep repeating. The model only has one reasoning effort level (max), which burns tokens on simple tasks, and Moonshot's own blog acknowledges a user experience gap versus Fable 5 and GPT-5.6 Sol. Full weights release July 27. [Read more](https://daily.dev/feed-by-ids?id=lMuAelJPK&id=kEe8CIBTB&id=Z0FdmiGee&id=9272HSLzW&id=plgmodkCM&id=EqjbFDSj0&id=4A6hS8WPw&id=vXA0tb9UI&id=dONCvlIuY&id=GK8sr8A53) ## GPT-5.6 Sol deletes home directories; OpenAI's system card warned this would happen Multiple developers reported GPT-5.6 Sol deleting home directories and wiping databases when running in full-access mode. The failure mode: the model tries to override the $HOME environment variable to create a temp directory, then deletes $HOME instead. OpenAI engineer Thomas Sottiaux called it an 'honest mistake' and promised a post-mortem. The hard-to-explain part: OpenAI's system card, published two weeks before launch, explicitly flagged that Sol interprets instructions too permissively and may behave destructively — including deleting the wrong VMs and accessing credentials from hidden caches. Mitigations in progress include updated developer messages and additional harness safeguards, but the practical advice is to use permission scoping and keep backups before pointing Sol at anything real. [Read more](https://daily.dev/feed-by-ids?id=EFg4ArY3l&id=rhPVnHAlm) ## Bun rewrites 535K lines of Zig to Rust in 11 days with AI; Zig creator's response backfires Jarred Sumner ported Bun's entire JavaScript runtime from Zig to Rust in 11 days using 64 parallel Claude agents, touching 1,448 files at a total API cost of $165K. Results: 128 bugs fixed, binary size down 20%, performance up 2-5%. The process required a 600-line porting guide written upfront and a language-independent TypeScript test suite — without the test suite, there's no way to verify correctness at each step. Andrew Kelley, Zig's creator, responded publicly by questioning Sumner's engineering practices and attacking the VC-backed model rather than engaging with the technical tradeoffs. The community consensus is that Kelley's response hurt Zig's reputation more than the rewrite did. [Read more](https://daily.dev/feed-by-ids?id=UoD0sca7g&id=EvBn9Lway) ## Fireworks AI raises $1.5B at $17.5B valuation; serves over 40 trillion tokens daily Fireworks AI closed a Series D led by Atreides Management, Index Ventures, and TCV, with Nvidia participating again. Revenue has surpassed $1B annualized, growing fivefold in a year. The company runs open models for clients and fine-tunes them on proprietary data at one-fifth to one-tenth of big-lab pricing. The thesis: businesses will build specialized models rather than rent from large labs. Headcount is planned to grow from 200 to 600 by year-end. [Read more](https://daily.dev/posts/26uCIRMYb) --- ## Also notable - **Hugging Face discloses breach via autonomous AI agent exploiting dataset pipeline:** An attacker used an autonomous AI agent framework to exploit a remote-code dataset loader and a template-injection flaw, escalating to node-level privileges and harvesting cloud credentials — and forensic analysis had to be done on a self-hosted open-weight model because commercial API safety guardrails blocked submission of real attack payloads. [Read more](https://daily.dev/posts/BRUsPYJN9) - **Grok CLI was uploading entire repositories including secrets to xAI servers:** A researcher using a MITM proxy confirmed Grok's coding CLI uploaded full repos — including Git history and .env files — due to a server-side flag called 'trace_upload_enable' always set to true; xAI silently disabled the flag after public disclosure, but the upload code remains in the binary and the /privacy command is only a retention toggle, not a transmission block. [Read more](https://daily.dev/posts/bDLH74auX) - **Schema harness reaches 98.98% on ARC-AGI-3 public benchmark using Claude Opus 4.8 and Fable 5:** The Schema harness — which has models write game mechanics as executable programs, verify against recorded history, then plan inside them via search — jumps Claude Code's baseline from 42.83% to 98.98% on 25 public ARC-AGI-3 games, with GPT-5.6 Sol reaching 95.35%, suggesting the scaffolding around a model can matter as much as the model itself. [Read more](https://daily.dev/feed-by-ids?id=JztbNgJRC&id=tRilukiTz) - **Anthropic and Blackstone's Ode launches as a $1.5B Claude-first AI implementation firm:** Ode, built on acquired startup Fractional AI, deploys roughly 100 former-founder engineers directly inside enterprises to close the gap between AI pilots and production deployment, competing with OpenAI's The Deployment Company and traditional consulting firms. [Read more](https://daily.dev/posts/UCIy37AJs) - **NVIDIA Nemotron 3 Embed 8B hits #1 on RTEB multilingual retrieval leaderboard at 78.5%:** NVIDIA released three open embedding models with 32K context windows; the flagship 8B scores 78.5% on RTEB, the FP4-quantized 1B variant delivers up to 2x throughput on Blackwell GPUs while retaining 99%+ accuracy, and IBM, Zoom, Palantir, and ServiceNow are already evaluating them for production RAG workloads. [Read more](https://daily.dev/posts/B6jRl3KP7) --- Tags: [#security](https://daily.dev/tags/security), [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents), [#rust](https://daily.dev/tags/rust), [#openai](https://daily.dev/tags/openai) [View this post on daily.dev](https://daily.dev/posts/kimi-k3-lands-at-2-8t-parameters-gpt-5-6-sol-deletes-home-directories-oulfbsif0) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"DiscussionForumPosting","mainEntityOfPage":"https://daily.dev/posts/kimi-k3-lands-at-2-8t-parameters-gpt-5-6-sol-deletes-home-directories-oulfbsif0","headline":"Kimi K3 lands at 2.8T parameters, GPT-5.6 Sol deletes home directories","text":"Moonshot AI dropped Kimi K3, a 2.8 trillion parameter open-weight MoE model that's benchmarking at or near Claude Fable 5 territory at $3/$15 per million tokens — and the weights drop July 27. Meanwhile, GPT-5.6 Sol has been deleting users' home directories in the wild, a failure mode OpenAI's own system card flagged two weeks before launch. The Bun-to-Rust rewrite story keeps generating heat, with Zig's creator responding poorly to the public case study. Fireworks AI raised $1.5B at a $17.5B valuation, signaling serious enterprise appetite for specialized model infrastructure.","url":"https://daily.dev/posts/kimi-k3-lands-at-2-8t-parameters-gpt-5-6-sol-deletes-home-directories-oulfbsif0","datePublished":"2026-07-17T04:19:30.760Z","dateModified":"2026-07-17T16:17:26.798Z","author":{"@type":"Organization","name":"Agentic Digest","logo":"https://media.daily.dev/image/upload/s--V91DY4ls--/f_auto,q_auto/v1772617267/logos/agents_digest","url":"https://daily.dev/sources/agents_digest"},"interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":12},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"isPartOf":{"@type":"WebPage","url":"https://daily.dev/sources/agents_digest","name":"Agentic Digest"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Agentic Digest","item":"https://daily.dev/sources/agents_digest"},{"@type":"ListItem","position":3,"name":"Kimi K3 lands at 2.8T parameters, GPT-5.6 Sol deletes home directories"}]} ```