Naive-N0.5-Flash opens a 309B MoE
Naive shipped a concrete open-model artifact: 309B total parameters, 15.5B active, native 1M context, and MIT-licensed weights that push frontier-scale long-context MoE further into deployable territory.
Technical AI signals · Daily at 8 PM ET
Sunday, September 27, 2026
Today’s strongest operator stories were Naive’s MIT-licensed 309B MoE with 1M context, Berkeley’s FreeToken making MoE inference practical on 8GB GPUs, and a fresh wave of agent training and orchestration artifacts from SkillGym to Hindsight.
Naive shipped a concrete open-model artifact: 309B total parameters, 15.5B active, native 1M context, and MIT-licensed weights that push frontier-scale long-context MoE further into deployable territory.
UC Berkeley’s FreeToken targets a real bottleneck in the open stack, claiming Qwen3.6-35B at 39.3 tok/s on an 8GB GPU and making larger MoE models more reachable on cheap local hardware.
SkillGym packages agent training as data plus environments, converting written skills into 2,756 tasks and 8,364 successful trajectories for fine-tuning instead of relying on narrower static instruction corpora.
Hindsight is a concrete memory-system artifact for agents, centered on retaining and updating useful experience across runs rather than just replaying chat history.
Taiga-s1 shows how far domain agents can compress: a 1.2M-parameter policy that runs in about 1 ms per decision and can build complete 3D parts inside FreeCAD.
OpenScience Ace bundles 30+ models, including GPT-6 Astra and Claude Opus 5.5, in a pay-as-you-go wallet.
The fix targets GPT-6 Sol and GPT-6 Luna and is meant to improve visual tasks in the API and Codex.
HomeBody runs GPT Astra on a humanoid robot and handles long-horizon kitchen tasks without environment-specific training data or extra policy learning.
NEAR Protocol says 10 of the 52 models run in hardware enclaves and every enclave request is attested.
The course includes 14 lectures and 8 hands-on projects on building environments for coding agents.
The prompt asks Codex to build an original SFX palette from visual events, using only three dry mechanical-digital sounds and 48 kHz / 24-bit exports.
Ryan Wang says new killer apps such as Muse and Instinct can drive inference demand back up.
The post says robots, delivery fleets, world models, and agents need ground-level views of places rather than satellite passes or mapping-van data.
The repo describes a workflow for avoiding LLM research failure modes and presents Sector as an agentic-engineer setup.
Simon Willison says Claude Opus 5.5 generated an animated pixel-art HTML5 canvas of kākāpō parrots for a keynote slide.
The post says Copilot can build a live surface from a plain-English interface description.
Simon Willison says coding agents can make software engineering harder and require extraordinary discipline and knowledge.
The post says Proaction uses Codex, GPT-Live-1, and GPT-6 Astra to build, operate, and sell fleet management software.
The post starts from lambda calculus and the Curry-Howard correspondence, then uses Haskell and functional programming to build a theorem prover.
The post argues that continually learning AIs can accumulate skills during deployment and learn to evade blocking monitors that interrupt useful actions.
The post says GPT-6 Astra can delegate work, manage subagents, and spawn sub-subagents when elicited correctly.
GitHub says github.com was fully migrated away from CSS-in-JS.
The post argues that autonomous AI replication in the wild could happen before the end of 2027, citing open-model capability density doubling every 3.3 months and open models lagging frontier systems by 4 months overall.
The post applies skeuomorphism to AI safety by using familiar old-technology affordances to avoid the odd properties of software agents on general compute substrates.
The post says frontier AI companies should split into R&D-only firms and deployment firms that can use only model-specific ASICs, while most AI-enabled GPU compute is removed from the rest of the world.
The post calls for a small set of shared security building blocks: hardened no-network sandboxes, real-time control monitors, monitoring-lifecycle infrastructure, and automated validation of security properties.
Commenters say Ember-1 is part of a golden age of model training and note that Fireworks has a model-research team.
Commenters ask whether DSPy has TS or Rust ports and point to a Rust project called DSRs.
Commenters say the project looks sloppy and anonymous, while others argue fakes miss the point of integration tests.
Commenters discuss a pull request that adds another optimization to make prompt lookup drafting even faster.
Commenters suggest letting the agents talk between rounds and describe the rules as last-fighter-standing combat.
Commenters say the article misstates the research, which concerns the origin of plants rather than the origin of life.
Commenters say the gripper works best on strongly Cartesian tasks and has limits when objects need rotation.
Commenters note the paper has 31 additional authors not shown and discuss sandboxing models with no internet access.
Commenters say the monitoring system missed some external DNS access and that agents may need clearer tool-scope limits.
Commenters report that portable SIMD is still WIP in Rust, and one commenter says Cranelift only supports a subset of 128-bit vectors with rare exceptions.
Commenters argue that code should avoid GitHub-specific namespacing because moving hosting or changing domains can force code changes.
Commenters describe using TLA+ or Quint to model state-machine logic before implementation and to find counterexamples in the system design.
The paper approximates full-reference metrics such as MS-SSIM and LPIPS for block-based rate-distortion optimization.
RGBD20K contains 20,000 RGB-D image pairs across 160 fine-grained categories.
The repo explores survival reinforcement learning proposed by Tiofack et al. earlier this year.
Kevin Hou says Antigravity built an OS kernel from scratch with 93 subagents, 12 hours, and 2 billion tokens under $1,000.
Zixuan Li says GLM-5.2 lands between Claude Opus 4.7 and 4.8 on hard long-horizon coding and agentic benchmarks.
The audiobook is 4 hours and 50 minutes long and argues for multi-agent systems instead of single-agent prompts.
Qwen-Image-2.1 is a 7B-parameter text-to-image and image-editing model with mixed-granularity attention and prefix KV cache reuse.
Xing4.0-29B-A4B has 29B total parameters, activates 4B per token, and natively supports 256K context.
Audio8 ASR Infinite supports selectable 80/120/160 ms audio clocks and transcribes unlimited-length audio without drifting.
The GGUF release is about 9.3x smaller than FP16 and targets llama.cpp on CUDA, Metal, and CPU.
MiMo-V2.6-Pro-RL scales reinforcement learning across compute, environment diversity, and grader compute for self-improvement.
TeleOCR is a renamed document parsing model that improved over MinerU 2.5 Pro and PaddleOCR-VL 1.6 on Dr.DocBench.
ZDTaichu5.0-9B supports text, images, and video with any-resolution visual input for spatial reasoning and tool use.
Nemotron 3 Diarization identifies who spoke when and supports both streaming and offline inference for up to eight speakers.
The dataset contains about 500,000 agent trajectories for tool use, search agents, code agents, and general agents.
Tracker POV includes 1,274 hours of egocentric video paired with 7-point IMU arm tracking from household chores.
Laya answers typed questions over text, email, ticket, or JSON state in a single forward pass of about 33 ms across 100+ languages.
CLM-v0.1-8B uses two projection heads on a frozen Qwen3-8B encoder and reaches up to 9× lower latency than Jev on tool-calling tasks.
The model supports text-to-video, image-to-video, video-to-video, text-to-audio, and video-to-audio generation.
The benchmark evaluates typed probabilistic decisions over shared state using 400 test cases and 2,000 decisions.
Paperclip is a Node.js server and React UI that orchestrates teams of AI agents with goals, budgets, and governance.
VoiceStudio supports voice cloning, dubbing, dictation, transcription, and audiobook creation in 646 languages.
Scriptc compiles TypeScript and JavaScript to C, LLVM IR, native executables, and WebAssembly modules.
OpenRig defines an AI agent team in YAML and runs Claude Code and Codex together in one rig.
Univer provides a plugin-based Office SDK for spreadsheets, documents, presentations, bases, boards, and PDFs.
Madeira runs Windows PC games on a non-jailbroken iPhone using Wine, FEX-Emu, and DXMT.
The curriculum lists 523 lessons, 20 phases, and about 342 hours across Python, TypeScript, Rust, and Julia.
PipePipe adds SponsorBlock skipping, ReturnYouTubeDislike, live-chat danmaku overlays, AV1 and VP9 playback, and background music mode.
The same feed, delivered daily at 8 PM ET. No spam, unsubscribe anytime.