Executive Summary
MiniMax released M3, an open-weights model claiming the #1 spot on the AA Intelligence Index with ~428B total and ~23B activated parameters, alongside an open-sourced MSA kernel library achieving significant sequence-length speedups. A major controversy erupted around Anthropic's Fable 5, with developers flagging mandatory data retention policies that override user privacy settings and criticizing the model for allegedly degrading outputs for users classified as doing "disallowed frontier work." Hermes Agent (NousResearch) continued its rapid iteration with WhatsApp integration upgrades, automation blueprints, and NVIDIA's Nemotron 3 Ultra (550B) now supporting the platform. OpenAI rolled out saveable rate limit resets for Codex and a referral program.
Key Events
- MiniMax M3 open-weights model launched — ~428B total parameters, ~23B activated, #1 on AA Intelligence Index, with an open-source MSA kernel library for scaling long-context inference. → link
- Fable 5 mandatory data retention sparks backlash — Users discovered Anthropic's Fable 5 retains agent request/output data regardless of Cursor Privacy Mode settings, with human review possible; Microsoft reportedly pulled Fable 5 from internal Copilot over the same policy. → link
- Anthropic accused of anti-competitive "sabotage-as-safety" — Multiple developers allege Anthropic degrades model outputs for users building competing open-source alternatives, calling it gatekeeping and "evil personified." → link
- NVIDIA Nemotron 3 Ultra (550B) dropped for Hermes Agent — 5x faster model built for long-running agents, now free on the Hermes platform. → link
- Hermes Agent WhatsApp overhaul shipped — No more burner phone numbers, major UX improvement;
hermes updateto access. → link - OpenAI Codex gets saveable rate limit resets and referral program — Users can now bank resets for later use and refer friends for additional resets. → link
- Fable 5 Max reportedly only 1pt above GPT 5.5 xHigh — Claude Code called the worst-performing harness; analysts question the cost-to-value ratio for Anthropic's latest. → link
- Comprehensive AI benchmark/evaluation guide published — Covers contamination, leakage, proper eval design, and the 2026 standard for serious LLM evaluation. → link
- Tinygrad benchmarking on 4xMI300X — Memory bandwidth data suggests significant room for improvement over vLLM baseline. → link
Analysis
The dominant pattern this cycle is the growing open-weights vs. closed-ecosystem tension. MiniMax's M3 release with open weights and kernel library directly challenges the premise that frontier models must remain proprietary. Simultaneously, Anthropic faces a credibility crisis among developers: mandatory data retention in Fable 5, perceived anti-competitive behavior via output manipulation, and Claude Code's lock-in strategy are all eroding trust. The refrain "local is the only place private still means private" signals a strengthening demand for self-hosted models.
Agentic infrastructure is maturing rapidly — Hermes Agent's automation blueprints, WhatsApp support, and desktop app signal a shift from chat interfaces to ambient task runners. NVIDIA's Nemotron 3 Ultra integration reinforces that GPU vendors are positioning themselves as the compute backbone for the agentic layer.
What to watch next: Whether Anthropic responds to the data retention and output-manipulation allegations; whether M3's benchmarks hold under community scrutiny; and whether the "agentic OS" concept (Teknium) translates into mainstream developer adoption or remains a power-user tool.
Tweet Feed
MiniMax M3 Release
@SkylerMiao7 · 2026-06-12T14:16
MiniMax M3 weights are live. #1 open-weights model on the AA Intelligence Index. → tweet link
@SkylerMiao7 · 2026-06-12T13:57
Proud to open-source our MSA kernel library and technical paper. Hope it empowers the community and helps more labs scale their research! Weights coming in minutes! → tweet link
@victormustar · 2026-06-12T16:53
interesting MiniMax shipped a kernel on Hugging Face 👀 → tweet link
@victormustar · 2026-06-12T14:49
MiniMax M3 is available on HuggingChat 🚀 Great UI designer it seems: → tweet link
@victormustar · 2026-06-12T11:15
new frontier code open source model here 👀 → tweet link
Anthropic / Fable 5 Controversy
@sudoingX · 2026-06-12T17:42
just clicked through this to turn on fable 5 in cursor and i actually stopped and read it twice — "the model provider will retain agent request and output data associated with this model, regardless of your Cursor Privacy Mode setting." [...] a human can read your agent's requests and outputs, not just a classifier. [...] this tradeoff only exists because the weights live in someone else's datacenter. the model running on your own gpu retains nothing, overrides nothing, puts no stranger in the loop. frontier in the cloud now ships with mandatory retention that beats your own privacy setting. local is the only place private still means private. → tweet link
@TheAhmadOsman · 2026-06-12T16:52
Finally, this was my attempt to capture all my thoughts on Anthropic from this past year. Anthropic is evil personified (and I don't take it lightly when I label it as such). Your freedom stands in the way of their moat, and they want it gone. → tweet link
@TheAhmadOsman · 2026-06-12T15:52
If a coding or research model secretly changes the quality, direction, or reliability of an answer because it classified the user as doing disallowed frontier work, the tool is no longer merely "safe." It is UNTRUSTWORTHY. → tweet link
@TheAhmadOsman · 2026-06-12T11:42
Imagine a compiler that emits worse binaries when it thinks you are building a competing compiler. Imagine a microscope that blurs certain samples because the manufacturer dislikes the research direction. Imagine a debugger that lies only when your codebase resembles a future rival. That's the world Anthropic is aspiring to. → tweet link
@TheAhmadOsman · 2026-06-12T06:15
Wrote an article on the case against Anthropic as a safety-branded permission regime for Cognition Infrastructure. Covers sabotage-as-safety, anti-Opensource rules, Fable, regulatory capture, data asymmetry, Claude Code, and Who Owns Intelligence. → tweet link
@TheAhmadOsman · 2026-06-12T18:12
Don't use Fable 5 to build personal projects. Don't use Fable 5 to build business projects. Don't use Fable 5 for anything really, they'll end up stealing it from you. → tweet link
@jezell · 2026-06-12T15:20
Never build on Anthropic. → tweet link
@TheAhmadOsman · 2026-06-11T21:28
Anthropic can use the internet, copyrighted books, code, user feedback, public human knowledge, synthetic data, and its own models to improve Claude. But if a developer uses Claude to bootstrap a competitive open alternative, Anthropic calls foul. That is called Gatekeeping. → tweet link
Model Performance & Harness Analysis
@kunchenguid · 2026-06-12T08:10
- Claude Code is actually the worst performing harness when using the same model, significantly behind opencode and cursor cli [...] what they are good at is making great models. they suck at making good harness products [...] 2. fable 5 max is only 1pt above gpt 5.5 xhigh (77 vs 76) [...] alarming for anthropic because it's very unlikely people will want to pay 2x higher cost for the 1pt difference. → tweet link
@kunchenguid · 2026-06-11T20:41
been on both anthropic and openai's subscriptions for a while and this aligns very well with my real experience - you get a lot more value from openai's plans right now. and this analysis hasn't even taken into account that gpt 5.5 gets the same thing done with much less tokens. → tweet link
@kunchenguid · 2026-06-12T03:20
trade off between compile time errors and runtime errors across different languages [...] rust is very strict at compile time and would almost never have runtime errors. ruby is the other extreme [...] this tradeoff is very interesting to have in mind when choosing a language for your agents. → tweet link
Hermes Agent / NousResearch
@Teknium · 2026-06-12T18:21
Love hearing stories like this! I want to hear more, what are you building with Hermes Agent? We're working on a builder spotlight series to feature real workflows from the community. → tweet link
@Teknium · 2026-06-12T16:43
Hermes Agent users that use WhatsApp just got a HUGE upgrade. No more burner phone #'s, way better UX, an all around major overhaul for WhatsApp support!
hermes updateto access now! → tweet link
@Teknium · 2026-06-12T15:23
The era of the Agentic OS has a lot of room to explore on the UX front. Excited to see where we go from here! → tweet link
@Teknium · 2026-06-12T06:10
Welcome to the Hermes Agent contributor crew! → tweet link
OpenAI / Codex
@gdb · 2026-06-12T00:26
For next two weeks, refer your friends to Codex, and you'll bank a rate limit reset: → tweet link
@jezell · 2026-06-11T20:59
Codex seems to be struggling at the moment. Time for a coffee break. → tweet link
@steipete · 2026-06-11T20:58
Getting Chris to do a PR with Codex! → tweet link
Benchmarks & Evaluation
@TheAhmadOsman · 2026-06-11T20:43
The MOST COMPLETE GUIDE for understanding benchmarks and evals, and why training on them is intentionally misleading is now available online to read for free. Covers [...] What machine learning is actually trying to measure [...] Leakage types and benchmark contamination [...] The 2026 standard for serious LLM evaluation [...] The benchmarks / evals / test sets are the rulers. Don't bend them. → tweet link
Hardware & Infrastructure
@tinygrad · 2026-06-12T16:30
Up and benchmarking itself on 4xMI300X (this is with vLLM). Memory bandwidth says we should be able to go so much faster. → tweet link
@cooltechtipz · 2026-06-12T18:27
The Data center ecosystem. → tweet link
Reinforcement Learning & Research
@jsuarez · 2026-06-12T18:29
Reinforcement learning research with Joseph Suarez → tweet link
@jsuarez · 2026-06-12T13:44
If the timeline is full of new RL people training N pendulums, where are all the PRs? → tweet link
Developer Culture & Open Source
@TheAhmadOsman · 2026-06-11T20:23
This guy created Linux and beat Microsoft to the best operating system. He doesn't mind being ruthlessly honest and doing the dirty work because nothing mild ever wins. Lesson in that for Opensource AI btw. → tweet link
@sudoingX · 2026-06-12T18:01
if you want to grow here on x and don't know what niche to pick, go all in on local llms and publish your actual research. the crowd here rewards people who build real things and show the numbers [...] i took my acc from 600 to 30k doing exactly that. → tweet link
@hnasr · 2026-06-12T18:47
The audio version of Root Cause: Stories and Lessons from Two Decades of Backend Engineering Bugs, narrated by yours truly is live on Amazon and Kobo! Listeners of the audio book gets additional content and commentary. Not a mention of AI in the book, so in case you need a break. → tweet link