Salvatore Sanfilippo released DwarfStar 4 (ds4), a local LLM inference engine that runs DeepSeek V4 Flash on a MacBook at 26–39 tokens per second without Python, Docker, or an Ollama daemon. The engin…
Salvatore Sanfilippo, the creator of Redis, published DwarfStar (repo `antirez/ds4`), an MIT-licensed inference engine that runs DeepSeek V4 Flash, GLM 5.x, and DeepSeek V4 PRO entirely on consumer ha…
Better Stack demonstrated a simulated Open Duck Mini robot duck controlled by Needle 3, a 15MB, 4-layer tool-calling model from Cactus Compute that maps natural-language requests to five actions: walk…
Apple announced it will require "very explicit user action" before apps can read a user's full filesystem, messages, and browsing history, citing risks from increasingly autonomous AI agents, followin…
ARC Prize reported that the cost to reach 75% accuracy on ARC-AGI-1 fell 99.95% over 19 months, from $26 per task with o3-preview to $0.01 per task with DeepSeek V4 Flash. On ARC-AGI-2, ARC Prize said…
A study posted to arXiv (2610.00314v1) found that predictive credit for research-agent explanations remained unconfirmed across 336 prospective states in controlled learning, 12 Tox21 endpoints, and 2…
Clustering two NVIDIA DGX Sparks over a 200 GB RoCE RDMA cable and running tensor-parallel inference with vLLM lets the pair serve models in the 150 to 160 GB range, such as DeepSeek V4 Flash, that do…
Maneshwar, building LiveReview, cut the cost of an AI code-review run from over 50 cents to about 4 cents by replacing Gemini's hosted File Search with a local Chroma vector store embedded with Qwen3-…
DeepSeek released DeepSeek-V4.1-Flash, an asymmetric Causal Encoder-Decoder Mixture-of-Experts model with a 552 billion parameter backbone that activates only 8 billion parameters during prefill and 1…
Debian launched the Debian Inference Portal, a self-service service that lets Debian contributors log in with their Salsa account, create API keys, and track spending on shared LLM inference through a…
Parity engineers ran a self-hosted inference trial of the open-weight DeepSeek V4 Flash model that handled more than 144,000 requests and almost 12.9 billion tokens from 25 engineers between 16 August…
DeepSeek has deprecated DeepSeek V4 Flash, the deepseek-v4-flash model released on 2026-04-24 with a 1000k context window and 384k maximum output, according to a provider announcement tracked by model…
A September 2026 comparison of open-source coding models found GLM-5.3-Flash from Z.ai best for agentic coding, DeepSeek V4 Flash cheapest per token, and MiniCPM5-2B best for on-device use. GLM-5.3-Fl…
Ds4 now supports directional steering, a runtime activation edit that applies a normalized f32 direction per transformer layer during inference, with steering files shaped 43x4096 for DeepSeek V4 Flas…
Cactus Compute released Needle 3, an open-weight foundation model that ships as a single 8-29MB file for on-device tool calling, structured extraction, and embeddings, with a runtime engine under 1MB …
Cactus Compute's Needle 3, a foundation model shipping as a single 8-29 MB file, beats models ten times its size on mobile tool-calling accuracy and matches models two to three times larger on structu…
Cactus Compute released Needle 3 on September 17, an Apache 2.0-licensed automation foundation model that ships as an 8 to 29MB binary and runs tool calling, structured extraction, and text embeddings…
Anthropic does not include Claude Code in its free plan, listing it as "No" on Free and including it from Pro at $20 a month or $17 a month billed annually, and in Max from $100 a month, according to …
A developer analysis of OpenRouter's LLM routing gateway warns that the abstraction between model and provider leaks in production, with the same model ID yielding significantly different performance,…
Abacus.AI released three open-weight Smaug models — Smaug Flash on DeepSeek V4 Flash, Smaug Mini on Qwen3.8 27B, and Smaug Agentic on Kimi K3 — tuned for long-running agent loops and published on Hugg…