DeepGrove released Maple-Preview, an open-weight 20B-A1B ternary-weight mixture-of-experts reasoning model that runs at 218 tokens per second on an Apple Mac mini M4, with a 5.31 GB checkpoint and a 1…
LM Studio Bionic, in partnership with Meta, launched support for Muse Glimmer, a 30B-parameter open-source model optimized for agentic tasks, available for local download under the Apache 2.0 license.…
Google DeepMind released DiffusionGemma, an experimental open-weight model that converts the Gemma 4 architecture into a discrete text-diffusion system using less than 10% of Gemma 4's original traini…
Google DeepMind retrofitted Gemma 4 into a text diffusion model, DiffusionGemma, using less than 10 percent of the original training budget, and it generates 256 tokens in parallel at about 1,500 toke…
Local models such as Qwen 3.6 and Gemma 4 are close to being useful for Kotlin Multiplatform code generation but still fall short for many developers, according to Mark Murphy's experiments. Murphy su…
A mechanistic interpretability study found that ablating specific MLP neurons in Google's Gemma 4 model causes it to revert to its dormant Bard identity, with the model answering 'My name is Bard' wit…
A developer using the LFM 2.5 2.6B model with llama-cpp found that adding detailed descriptions to JSON schema parameters improved tool call correctness from 0% to nearly 100%, enabling the use of a f…
Fireworks, an AI inference platform, does not support voice AI models such as Parakeet, Kokoro, and Qwen ASR because voice workloads require different optimization strategies than text LLMs, according…
Google DeepMind released DiffusionGemma, an open-weight language model that generates text via discrete diffusion, achieving roughly 1,500 output tokens per second on a single H100 compared to about 3…
Google DeepMind introduced DiffusionGemma, an experimental open-weight language model that uses discrete diffusion to generate text at high speed, achieving roughly 1,500 output tokens per second on a…
An open-source test suite for voice assistants, developed by Aquila, is now available, with reproducible tests and a leaderboard at https://git.cicero.sh/aquila/ha-voice-test-suite/. The creator, who …
Developer Ryan Findley ran Google's Gemma 4 26B MoE model on a 2013 HP StoreVirtual server (dual Xeon E5 + DDR3, no GPU), achieving about 5.2 tokens/sec after quantization and inference tuning, with t…
Every notable open-weight frontier model released in the past year uses a Mixture-of-Experts (MoE) architecture, according to an analysis by Vetted Consumer. The shift means total parameters range fro…
Researchers introduced a model-side aggregation interface using Hash-based HyperLogLog (HLL) sketches that maintains compact 2 KiB states alongside frozen language models, achieving 1.6% mean relative…
TurboFieldfare, a Swift-and-Metal runtime, runs Google's Gemma 4 26B-A4B model on Apple Silicon Macs with as little as 2 GB of RAM by streaming expert weights from SSD, achieving 5.1–6.3 tokens per se…
A developer tested Google DeepMind's open-weight Gemma 4 against Anthropic's Claude Opus 5 and OpenAI's GPT 5.5 on real Python projects, finding that the free, local model performed comparably to the …
Higgs, an unofficial, agent-first CLI for Proton Mail designed to be driven by language models, has been released as a local-only tool with no cloud dependencies or telemetry. The project, built by th…
A developer testing Nanbeige 4.2, a 3-billion-parameter looped-transformer model from a Chinese company, found it well-suited for tool calling with a Rust MCP server, outperforming Gemma 4 12B in that…
Google DeepMind's Gemma 4 family, launched on April 2, 2026 under the Apache 2.0 license, anchors a growing industry coalition including Meta, Microsoft, and Nvidia pushing back against Anthropic and …
A developer known as slvDev has run a 28.9 million parameter language model on an ESP32-S3 microcontroller that costs about $8, generating text at roughly 9.5 tokens per second entirely on-device. The…