MiniMax M3 Drops 428B Open-Weight MoE With Only 23B Active Parameters
π§ LAUNCH
MiniMax M3 Drops 428B Open-Weight MoE With Only 23B Active Parameters
MiniMax M3 is the kind of model that makes you rethink parameter counts. At 428B total parameters but only 23B activated per token, it's a Mixture of Experts architecture that delivers frontier-class capability at a fraction of the inference cost you'd expect from a model this size. Open-weight, already on HuggingFace, and natively supported by Transformers v5.12 on day one β the ecosystem velocity here is remarkable. If you're evaluating open-weight models for production, this just jumped to the top of your benchmark list. (2,123 likes | 248 RTs) Read more β
Moonshot AI Drops Kimi K2.7 Code Into the Open-Weight Ring. Kimi K2.7 Code lands from Moonshot AI β a dedicated coding model variant from the Chinese lab that's been quietly climbing benchmarks. The open-weight coding model field now has serious entries from Cohere, Xiaomi, and Moonshot alongside the incumbents, and the quality floor keeps rising. (331 likes) Read more β
Google Opens Project Genie Globally for Ultra 5X Subscribers. Project Genie β Google's most advanced AI tier β goes global. The tiering signal matters more than the product: Google is following Anthropic's playbook of premium capability tiers for power users, and the subscription arms race is now a three-way fight. (688 likes | 100 RTs) Read more β
π§ TOOL
Codex Gets Chrome DevTools Protocol β Real Browser Debugging From Inside the Agent
Codex now has Chrome DevTools Protocol access β JavaScript profiling, network request inspection, and full page state debugging from within the agent. This isn't "paste me the error" workflow anymore; it's the agent opening the browser, running your app, and diagnosing the rendering bug itself. For frontend debugging, this makes Codex a genuinely viable pair programmer, not just a code generator. (2,825 likes | 192 RTs) Read more β
GitHub Copilot Code Review Now Supports MCP and Custom Agent Skills. Copilot code review enters public preview with MCP server connections and custom agent skills. Translation: you can now bring your org's internal tools, linting standards, and compliance checks directly into the PR review flow. For enterprise teams that couldn't adopt AI review because it didn't know their rules, this is the extensibility unlock that matters. (122 likes | 14 RTs) Read more β
Claude Code Ships Three Releases in a Day Post-Fable. Claude Code v2.1.174β176 landed in rapid succession β localized session titles, improved Bedrock credential caching, scroll acceleration control, and model picker fixes. The shipping cadence itself is the signal: Anthropic is iterating on developer tooling at a pace that suggests Fable unlocked internal velocity too. Read more β
Transformers v5.12 Ships Same-Day MiniMax M3 Vision-Language Support. HuggingFace Transformers v5.12 drops with native MiniMax-M3-VL support β the vision-language variant with a CLIP-style tower and 3D rotary position embeddings. From model announcement to pip install in hours. That ecosystem velocity is why HuggingFace remains the default distribution layer for open-weight models. (245 likes | 442 downloads) Read more β
π TECHNIQUE
Simon Willison Goes From Fable Skeptic to "Relentlessly Proactive" Convert
Simon Willison β one of the loudest Fable 5 critics last week β now describes it as "relentlessly proactive" after watching it autonomously spin up custom CORS proxy servers and debugging tools from a single screenshot. The shift from skeptic to impressed user is itself a data point: Fable's agentic behavior is qualitatively different enough that even experienced developers need a couple of days to recalibrate their expectations. If you dismissed it based on early takes, revisit. (667 likes | 44 RTs) Read more β
How to Manage Fable's Overwhelming Output in Long Agentic Sessions. Alex Albert shares a concrete prompt snippet for taming Fable 5 when it's too proactive β the model's biggest strength becomes a usability problem in extended sessions where it generates more context than you can track. Add the snippet to your CLAUDE.md and save yourself the cognitive overload. (676 likes | 14 RTs) Read more β
π¬ RESEARCH
Fable 5 Hits 87β88% on FrontierMath Without Being a Reasoning Specialist
Epoch AI's independent FrontierMath benchmark confirms what the vibes suggested β Fable 5 scores 87% on Tiers 1β3 and 88% on Tier 4. The headline isn't just the numbers; it's that Fable achieves this without being a dedicated reasoning model. It's a general-purpose model matching or beating specialist reasoners on their home turf, which raises the question of whether reasoning-specialized architectures are already a dead end. (684 likes | 93 RTs) Read more β
Frontier LLMs Now Outperform Purpose-Built Clinical AI Across the Board. A new paper shows general-purpose frontier models beat specialized clinical AI tools in all three evaluations. The implication is brutal for vertical AI startups: if your product is a fine-tuned model for a specific domain, the general-purpose frontier just lapped you. Reassess any domain-specific AI tools in your stack β the moat may already be gone. (315 likes | 33 RTs) Read more β
Microsoft Research Ships Arbor β a Generalist Autonomous Research Agent. Arbor uses persistent hypothesis-tree refinement instead of linear agent chains β a structural advance that lets the agent refine its research direction as evidence accumulates, rather than following a fixed plan. Microsoft Research shipping a generalist research agent signals the field is moving beyond task-specific agent designs toward agents that can genuinely explore. (220 likes | 37 RTs) Read more β
π‘ INSIGHT
48 Hours of Fable 5: The Community Build Showcase That Benchmarks Can't Capture
The official Claude account curated the most impressive community-built Fable 5 projects from the first 48 hours β and the thread is more revealing than any benchmark table. Real developers shipping real things in two days tells you what the model actually enables in practice: the gap between "scores well on SWE-Bench" and "people are building with it immediately" is the gap between potential and product-market fit. Browse the thread for project ideas. (29,885 likes | 1,596 RTs) Read more β
LeCun Fires Back: Amodei's Safety Governance Would Kill Open-Source AI. Yann LeCun amplifies the sharpest critique of Dario Amodei's AI policy essay β "declare AI too dangerous for competition, then propose a regime only your lab survives." Whether you agree with LeCun or Amodei, the framing matters: the "safety vs. openness" debate is increasingly a market structure argument dressed in ethical language. Read both sides before picking yours. (1,985 likes | 167 RTs) Read more β
TCS Partners With Anthropic to Push Claude Into Banks and Government. TCS β one of the world's largest IT services firms β partners with Anthropic to bring Claude into regulated industries. Following the DXC alliance, Anthropic is systematically locking down the enterprise consulting layer. The pattern is clear: if you can't sell directly to banks and governments, partner with the firms that already sit in their boardrooms. Read more β
Anthropic Publishes Its First "Public Record" With Chris Olah on the Pope's AI Encyclical. Chris Olah's remarks on the Pope's AI encyclical mark Anthropic's first "Public Record" release β a new transparency initiative combining AI safety research with broader societal discourse. Whether this becomes substantive or performative depends on what follows, but the format itself is worth watching. Read more β
Richard Socher Claims Recursive Has Achieved Self-Improving AI Research. Richard Socher says Recursive now has agents doing the AI research that makes better agents β recursive self-improvement, stated publicly by a serious researcher. Whether the claim holds up to scrutiny matters less than the fact that it's being made at all: this is where frontier labs think they're headed, and the timeline for "AI does its own R&D" just moved from theoretical to claimed. (578 likes | 83 RTs) Read more β
ποΈ BUILD
Extend CLI Open-Sources Document Parsing With Built-In Agent Skills. Extend CLI ships document parsing and extraction as an open-source CLI tool β with agent skills baked in so Claude Code and Codex can drive it natively. The pattern of CLI tools shipping with agent skill interfaces is becoming standard, and this is how the broader tool ecosystem adapts to agentic workflows. (128 likes | 14 RTs) Read more β
architect-loop: Fable as Architect, Codex as Builder β Cross-Vendor Agent Arbitrage. architect-loop is a Claude Code skill that routes architectural decisions to Fable and code-writing to Codex β the first clean implementation of model arbitrage in a practical developer tool. Use the best model for each subtask instead of picking one provider for everything. The cross-vendor agent loop pattern is where multi-model workflows get real. (77 likes | 5 RTs) Read more β
Step-by-Step: Setting Up a Local Coding Agent on macOS. A practical, no-nonsense guide trending on Hacker News that walks you from zero to a working local coding agent on macOS. Fills the gap between "coding agents exist" and "here's exactly how to run one on your machine." If you've been meaning to try local agent development, this removes the excuse. (226 likes | 68 RTs) Read more β
π MODEL LITERACY
Mixture of Experts (MoE) β Active vs. Total Parameters: MiniMax M3 has 428B total parameters but only activates 23B per token β and understanding why requires grasping the MoE architecture. Instead of running every parameter for every input, MoE models split their parameters into "expert" groups and a router selects which experts to activate for each token. This means a 428B model can run at roughly the inference cost of a 23B dense model while retaining the knowledge capacity of its full parameter count. The practical takeaway: "parameter count" alone no longer predicts cost, speed, or even quality β you need to know how many parameters are active, not just how many exist.
β‘ QUICK LINKS
- MiniMax M3 Weights Now Live on HuggingFace: 428B/23B-active MoE, ready to download and benchmark. (245 likes | 442 downloads) Link
- Google's Week in AI: Live Translate GA, NotebookLM upgrade, Genie expansion β three launches in one thread. (466 likes | 34 RTs) Link
- OpenAI Adds βK Search to the API Dashboard: Finally, a command bar for navigating pages, settings, and docs. (814 likes | 52 RTs) Link
- Ollama v0.30.8: Improved prompt caching decoupled from context shift, plus hardened MLX support for Apple Silicon. Link
π― PICK OF THE DAY
When the loudest safety advocate proposes governance only their lab survives, the debate changes shape. Yann LeCun's critique of Dario Amodei's AI policy essay cuts to something most commentary misses: the "safety vs. openness" framing may be a market structure argument wearing ethical clothes. Amodei's essay proposes governance frameworks that would effectively require the kind of capital, compute, and institutional relationships that only a handful of labs possess β which happens to include Anthropic. LeCun's counter isn't that safety doesn't matter; it's that conflating "safe" with "closed and well-funded" is a policy choice that kills open-source AI as collateral damage. For builders, this isn't an abstract policy debate β it determines whether you'll have access to open-weight models like MiniMax M3 in two years, or whether the frontier becomes a walled garden with API-only access. Before you pick a side, ask yourself: whose business model does each position protect? Read more β
Until next time βοΈ