Releases·Tool Release
Arm's new AI Portal exposes model data to agents. What would make a selection reproducible?
Arm announced its AI Portal on September 8, offering pre-optimized Qwen, Gemma, and YOLO models alongside performance data and deployment examples. Coding agents can access Portal resources through MCP, though agent-ready access is in early access. It is a vendor announcement without independent performance validation.
5 engagement1 sourceTue, Sep 8, 2026, 11:41 AM
AI Overview
- Arm's AI Portal launched September 8 with pre-optimized Qwen, Gemma and YOLO models; agent access via MCP is in early access
- Portal includes performance data and deployment examples; bring-your-own-model tooling is described as coming soon
- Arm's announcement includes no independent validation of the reported performance figures
Related
Developments·Paper
SCALPEL: Contrastive Sparse Autoencoder Targets Fine-Grained Machine Unlearning in Qwen, Llama, and Gemma
An arXiv paper (2026-09-25) presents SCALPEL, a contrastive sparse autoencoder that removes narrow target information at the representation level while preserving other model abilities. The authors identify an energy bias in reconstruction-based extraction that favors background structure over low-energy target-specific components. On TOFU across Qwen, Llama, and Gemma, SCALPEL improves substantially over NMF and standard SAE interventions and is competitive with Gradient Difference and RMU.
An arXiv paper (2026-09-25) presents SCALPEL, a contrastive sparse autoencoder that removes narrow target information at the representation level while preserving other model abilities. The authors identify an energy bias in reconstruction-based extraction that favors background structure over low-energy target-specific components. On TOFU across Qwen, Llama, and Gemma, SCALPEL improves substantially over NMF and standard SAE interventions and is competitive with Gradient Difference and RMU.
0 engagement1 sourcearxivFri, Sep 25, 2026, 09:51 AM
Products·launched
MCP server for Claude Code to delegate tasks to other LLMs
An MCP server that allows Claude Code to delegate tasks to other models like GPT-5.6, DeepSeek, GLM, and local Qwen. The creator benchmarked all models against Claude itself with 198 runs and hidden tests. It solves the problem of using multiple models from a single interface for comparison or specialized tasks.
An MCP server that allows Claude Code to delegate tasks to other models like GPT-5.6, DeepSeek, GLM, and local Qwen. The creator benchmarked all models against Claude itself with 198 runs and hidden tests. It solves the problem of using multiple models from a single interface for comparison or specialized tasks.
298 engagement1 sourceredditMon, Jul 20, 2026, 06:25 PM
Products·building
CPU-only LLM inference engine for mobile devices
A lightweight (450kb) inference engine built from scratch that runs LLMs like LFM2.5-2.6B on a OnePlus 13 at 17 tok/s using only CPU. It supports multiple model architectures and includes a device probe suite for performance monitoring via ADB.
A lightweight (450kb) inference engine built from scratch that runs LLMs like LFM2.5-2.6B on a OnePlus 13 at 17 tok/s using only CPU. It supports multiple model architectures and includes a device probe suite for performance monitoring via ADB.
235 engagement1 sourceredditWed, Aug 5, 2026, 02:20 PM
Products·launched
GIDE: Hybrid local/cloud AI coding editor with per-task model routing
GIDE is a coding editor that routes high-frequency, low-complexity tasks like autocomplete and small agent iterations to a local on-device GGUF model (Qwen, Gemma, DeepSeek, or bundled Ornith), eliminating token costs for those operations. For tasks requiring deeper reasoning, users can bring their own API key to call cloud models like Claude, GPT, or Gemini, choosing per task. It solves the cost and latency problem for developers who want AI assistance without paying per token for every keystroke.
GIDE is a coding editor that routes high-frequency, low-complexity tasks like autocomplete and small agent iterations to a local on-device GGUF model (Qwen, Gemma, DeepSeek, or bundled Ornith), eliminating token costs for those operations. For tasks requiring deeper reasoning, users can bring their own API key to call cloud models like Claude, GPT, or Gemini, choosing per task. It solves the cost and latency problem for developers who want AI assistance without paying per token for every keystroke.
3 engagement1 sourceredditThu, Aug 6, 2026, 06:38 PM
Social·Community
Users probe Gemma and Qwen logprobs to detect hallucination-prone recall
A developer experimenting with Gemma and Qwen models reports that examining logprobs during a model's first fact recall in its chain of thought may reveal unreliable recall, a potential early signal of hallucination. The approach failed when attacking hallucinations via system prompts, prompting the shift to logprob analysis.
A developer experimenting with Gemma and Qwen models reports that examining logprobs during a model's first fact recall in its chain of thought may reveal unreliable recall, a potential early signal of hallucination. The approach failed when attacking hallucinations via system prompts, prompting the shift to logprob analysis.
67 engagement1 sourceredditTue, Aug 11, 2026, 10:12 PM