Today’s lead · Product Hunt
Microsoft ships Decision-1 for agent routing and verification
Microsoft is pitching Decision-1 as a cheap, single-pass scoring model for agent routing, classification, verification, and control. The useful bit is structured probabilities over fixed options, not chat. The dossier says it is post-trained from Qwen3.5-9B and available in Foundry; the 35x speed claim is Microsoft’s.
Top signals
5 moreMiniMax-H3-Turbo cuts audio-video generation to four steps
MiniMax-H3-Turbo targets the practical bottleneck in audio-video generation: iteration speed. Wuli-Art says variants can generate audio and video together in as few as four steps, with text-to-video plus first/last-frame conditioning. Treat the licensing and quality tradeoffs as model-card homework before building on it.
ANEMLL Forge v0.2 pushes 27B inference on Apple’s Neural Engine
ANEMLL Forge v0.2 is another serious datapoint for local, low-power inference on Apple’s Neural Engine. The update swaps in new 3-bit vector-lookup MLP weights, Hadamard rotations, GPTQ/QAT fitting, and reports better BF16 agreement at the same speed. The 20–40 tok/s power claim comes from a commenter, not the project page.
Gyro-S quant squeezes Qwen3.8-Flash-Next onto a 32GB GPU
Agention’s updated Gyro-S quant tries to make a 125B MoE feel like a single-GPU model. The card claims Qwen3.8-Flash-Next fits a 32GB card at 27.64 GiB, hits about 221 tok/s on an RTX 5090, and reduces looping in stress tests. Loader support is limited to their forks.
Fluid distills clef-flash into a 0.70B local decision model
Fluid Inference distilled Cloudflare’s 9B clef-flash decision model into a 0.70B Qwen3-based Core ML package. The pitch is local classification at support-ticket speeds, with GPU or ANE execution and calibrated option probabilities. The model card is clear about the tradeoff: good routing/classification, weak knowledge and reasoning.
Fireworks discloses credential incident affecting some eval-job secrets
Fireworks disclosed a security incident involving unauthorized use of internal credentials and access to some customer evaluation-job environment variables. The company says it has not detected access to models, datasets, training data, personal information, prompts, outputs, or inference traffic. Builders should still rotate exposed secrets and watch follow-up scope updates.
Tools & repos
4 selectedmksglu/context-mode
Context-mode attacks a real agent cost problem: tool-output sprawl. It claims sandboxing, session memory, and MCP/hooks routing across 17 platforms, with a headline 98% reduction in tool output.
Onepin
Onepin is quality-control glue for AI voiceover pipelines. It works across existing voice providers, checks pronunciations, expands prices and dates, scores lines, and fixes individual words without re-rendering the full take.
mattpocock/skills
Matt Pocock’s skills repo is a peek into an agent-facing workflow directory rather than a conventional library. The value is in reusable task patterns builders can adapt for their own coding agents.
Claude Dashboards & Motion
Claude’s beta artifacts are moving from chat answers into working outputs: live BI dashboards and code-editable explainer videos. Useful if your team already trusts Claude, but the data-connector surface deserves careful governance.
Blogs worth your time
2 reads
Sebastian Raschka extends his reasoning-model build with GRPO stability work
Raschka’s part 7 is a practical debugging pass over toy-scale RLVR: track the right GRPO metrics, add clipped policy ratios, understand when KL helps, and avoid over-rewarding cosmetic think tags.
Akarsh Kumar argues evolution is a serious search primitive for AI
This MLST conversation is a useful antidote to “evolution equals random search.” Kumar connects artificial life, foundation-model critics, Core War program evolution, and why path-dependent regularities may matter for building more robust intelligence.
Community discussions
2 threadsA 4x Radeon R9700 box sparks the local-inference build-vs-cloud debate
A startup-scale local inference build claims useful throughput on four 32GB Radeon R9700 AI Pro cards. The comments quickly ask the right production questions: eval parity, GPU interconnects, batching, and single-stream latency.
Data scientists ask whether notebooks survive agentic workflows
The thread pushes back on replacing notebooks with prompt-result cells. Builders still value notebooks for exploration and preserved outputs, while critics point to git diffs, PR review pain, and production misuse.
Funding & acquisitions
1 moves
Apple hires from Huxe and licenses its personalized audio tech
Apple disclosed a reverse acqui-hire-style deal with Huxe: employment offers to some employees plus a non-exclusive IP license. The strategic read is obvious, but unconfirmed: Apple may want more AI-generated audio capability.
Bengaluru radar
1 events
Reimagining education in the age of AI with Dr. Prathosh A.P.
A free Bengaluru talk on AI’s impact on classrooms, teaching, assessment, equity, and original thinking, led by IISc’s Dr. Prathosh A.P.