DeepSeek released its smallest model in a new Causal Encoder-Decoder family: ~552B total / 8B active MoE with native multimodal understanding, up to 1M contextcars and roughly a quarter of the prior generation's KV-cache HBM. MIT weights are on the Hub and old v4-flash IDs now route to this model.
Qualcomm detailed its next-generation Hexagon NPU: a new Element Accelerator for transformer workloads, 50% more shared memory, KV-cache acceleration, context support to 32Kcars and roughly 50% faster INT4 prefill than the current Snapdragon 8 Elite Gen 5 — targeting 30B-parameter MoE models that activate only ~3B parameters on phones.
Wiz scanned 3,074 internet-exposed LiteLLM gateways and found 9.6% accept the default master key or no auth at all, chaining with the MCP auth bypass (CVE-2026-59822) and guardrail RCE (CVE-2026-59821) into root-level code execution and cloud credential theft. Rotate the master key and upgrade to LiteLLM 1.84.0+.
NVIDIA's system enables online co-training of speculative drafts (EAGLE-3, DFlash, DSpark) by merging rank-local branch attention into packed zigzag-ring context-parallel attention, delivering 1.50–1.88x end-to-end roll-out speedups on targets up to 122B while preserving the RL trajectory.
Weight-Redundancy Pruning removes whole Transformer blocks without any forward passes, scoring inter-block similarity from weight redundancy instead of activation statistics — cutting inference cost while avoiding calibration-data forward passes entirely.
RiLM removes the output softmax matrix entirely for sub-million-parameter models by decoding over squared geodesic distance on a Riemannian manifold. HypRiLM, at ~290k params, reaches 54.2 validation perplexity on WikiText-2 versus 87.6 for the flat variant, aimed at edge deployment.
An FPGA-native neural network built from six-input lookup tables with learnable connectivity, trained end-to-end via differentiable training rather than converting pretrained quantized models — eliminating multiply-accumulate operations in favor of LUT-based on-chip execution.
A study under strict memory and latency constraints for on-device in-vehicle assistants shows readable natural-language schemas outperform fixed opaque keys for small-model function-call accuracy in resource-limited deployments.
IBM Research's ~385M-parameter open-weight time-series foundation model is dual-licensed under Apache 2.0 and OpenMDW 1.0, and as of September 8 it tops GIFT-Eval among replicable zero-shot models under a permissive license.
AK highlights the AuK technical report, an open-source foundational model for speech generation and editing, with the paper and weights hosted on Hugging Face.
An encoder-only alternative to generative VLM backbones trained from scratch with masked discrete-diffusion over text and raw image patches. Apache 2.0 checkpoints ship with a visual-RAG retriever head for page-image document retrieval.
Qwen's team open-sourced ZG, a local-first CLI fusing ripgrep, BM25 and vector search for intent-based code search. Qwen benchmarks report up to 58.6% fewer agent tool calls and up to 82.7% fewer input tokens on SWE-QA Bench and real repos like pylint.
A Show HN open-source tool for spawning and remotely managing multiple Claude Code instances, aimed at teams running fleets of local coding agents.
A provider-agnostic CLI that runs a team of specialized agents (Architect, Planner, Builder, Tester, Debugger) locally against your codebase with your own API keys. MCP tools are permission-gated per tool with deny-by-default policies.
Kryzz AI was open-sourced (MIT) as a private, local-first Android AI workspace for chat, voice and tool-driven agents. v1.1 adds MiniMax as a provider, a multi-step Agent Auto tool loop and starter skills.
A Show HN CLI that audits context cost and conflicts across your agent skill definitions, helping local agent setups size and de-conflict their skill packs.
n8n's stable 2.38.6 bounds peak memory during source-control push, while 2.39.x pre-releases add fixes for the sandboxed built-in AI assistant (Instance AI), MCP tool handling and prevent Anthropic agent threads from breaking permanently.
A community experiment prefilled 45 prompts per model with GPT-5.5 Pro reasoning and found Qwen3.8's answers became dramatically more GPT-5.5-like (+18.18pp) while other local models barely moved — strong evidence Qwen3.8 was trained on that reasoning data.
A homelab user benchmarked Flash-Next NVFP4 against the 27B FP8 on a single RTX PRO 6000: Flash-Next was faster with flawlessly strict JSON and injection resistance, but denied partial credit on rubric-edge items where the 27B consistently complied.
A systems essay argues running local AI well is engineering, not model choice: unbounded concurrency against one local model causes requests to evict each other's KV memory, so bounded queues, visible streaming and deliberate memory planning are required for multi-agent homelabs.
Brazilian modder fmklab rebuilt the never-released RTX 3070 Ti 16GB by transplanting a 3070 Ti core onto an RTX 3070 PCB with GDDR6 chips, verifying the full 16GB pool by running a local LLM that used nearly all of it — a cheap path to more VRAM.
Leaker Hongxing2020 claims NVIDIA is preparing an RTX 5070 revision using the larger GB203 die instead of GB205, keeping core count, 12GB GDDR7 and the 192-bit bus — a supply-flex signal as Blackwell wafers get binned down-market.
Framework secured a limited batch of cheaper Micron 32GB and 64GB LPCAMM2 modules and is retroactively reducing memory prices for shipped and upcoming Laptop 13 Pro orders — a rare downward move in the RAM price crunch inflating local rig builds.
AMD's top-end RDNA 4 card overtook other AMD GPUs to become the most popular Radeon discrete card in Valve's Steam hardware survey, as NVIDIA RTX 50 supply stays constrained and street prices keep climbing.
Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.
Source:
