https://devotd.net/en/Clear, source-backed reporting on IT, AI, LLMs and local AI.enSun, 18 Oct 2026 03:42:10 GMThttps://devotd.net/en/news/llama-cpp-b11717-hexagon-npu-q1-0-q2-0-bonsai-2/https://devotd.net/en/news/llama-cpp-b11717-hexagon-npu-q1-0-q2-0-bonsai-2/Sun, 18 Oct 2026 03:42:10 GMTITSupport for matrix multiplication with ultra-low-bit weights, row extraction, and Hadamard operations has been added to Snapdragon NPUs. A developer said the Bonsai-2 1-bit and 2-bit paths are usable, but further optimization is needed to better exploit the reduced memory footprint.https://devotd.net/en/news/pplx-decider-v1-2-joint-request-decision-model/https://devotd.net/en/news/pplx-decider-v1-2-joint-request-decision-model/Sun, 18 Oct 2026 01:53:31 GMTLLMsPerplexity changed its request format and training data to assess multiple questions about the same context at once. Grouping related attributes helps, but independent candidates fare better separately. Long inputs and execution-path differences need checking.https://devotd.net/en/news/moss-image-5b-layout-control-text-rendering-open-weights/https://devotd.net/en/news/moss-image-5b-layout-control-text-rendering-open-weights/Sun, 18 Oct 2026 01:26:09 GMTAIOpenMOSS has released code and weights for image generation models that let users position text and objects with coordinates. It offers base and fine-tuned models. Adoption requires weighing layout input effort against memory and inference time.https://devotd.net/en/news/cloudflare-openai-crawler-discovery-pilot-results/https://devotd.net/en/news/cloudflare-openai-crawler-discovery-pilot-results/Sun, 18 Oct 2026 01:24:25 GMTAICloudflare and OpenAI say a pilot using network signals discovered hundreds of millions of new webpages and added them to a search index. Signals of changes and potential harm aim to reduce unnecessary crawling. Quantitative results on load reduction and actual business outcomes have not yet been provided.https://devotd.net/en/news/dspine-qwen3-8b-parallel-drafter-speculative-decoding/https://devotd.net/en/news/dspine-qwen3-8b-parallel-drafter-speculative-decoding/Sat, 17 Oct 2026 23:21:48 GMTLocal AITencent has released a draft model that generates answer candidates in parallel, along with inference code. Tests showed longer accepted drafts than the comparison methods. Actual speed gains must be weighed against extra resource costs.https://devotd.net/en/news/nvidia-indic-pretraining-2-8t-tokens-open-pipeline/https://devotd.net/en/news/nvidia-indic-pretraining-2-8t-tokens-open-pipeline/Sat, 17 Oct 2026 23:18:14 GMTLLMsNVIDIA introduced roughly 2.8 trillion tokens of translation-based training data for 22 Indian languages and an open generation process. The planned release supplements locally authored text. Reported gains come from a training stage using both types of data.https://devotd.net/en/news/cloudflare-verified-bots-continuous-accountability/https://devotd.net/en/news/cloudflare-verified-bots-continuous-accountability/Sat, 17 Oct 2026 23:17:47 GMTITCloudflare introduced a system that continually evaluates verified bots' compliance with instructions and responses to requests to slow down. It is building an internal observation history, with public scores and customer-selected trust criteria planned for the future.https://devotd.net/en/news/llama-cpp-b11714-discover-models-hf-hub-draft-sidecars/https://devotd.net/en/news/llama-cpp-b11714-discover-models-hf-hub-draft-sidecars/Sat, 17 Oct 2026 21:33:33 GMTITUsers can now search Hugging Face Hub models in the web interface and download inference auxiliary files matched to a quantization. Paused downloads survive restarts, but start over if files change or resume requests are rejected.https://devotd.net/en/news/nemotron-labs-dory-1-0-4b-looped-elastic-compute/https://devotd.net/en/news/nemotron-labs-dory-1-0-4b-looped-elastic-compute/Sat, 17 Oct 2026 21:32:19 GMTITNVIDIA has released a 4-billion-parameter language model that adjusts accuracy and compute by changing internal loop counts without retraining. It offers last-loop caching and multiple precision formats, but latency and cost need task-specific evaluation.https://devotd.net/en/news/opencode-console-provider-key-breach-distillation-attack/https://devotd.net/en/news/opencode-console-provider-key-breach-distillation-attack/Sat, 17 Oct 2026 21:17:10 GMTITA security breach affected 1.4% of all console users, specifically some users who connected external AI services. Notified users should revoke and replace keys in provider dashboards and review recent usage and connected applications.https://devotd.net/en/news/nemotron-3-5-bolt-74b-a9b-24gb-2bit-local-agent/https://devotd.net/en/news/nemotron-3-5-bolt-74b-a9b-24gb-2bit-local-agent/Sat, 17 Oct 2026 17:48:03 GMTLocal AINVIDIA has released a 24GB vision-language model designed to handle screen understanding and tool calls on personal devices. It compresses a large model and corrects quantization errors through training. Bolt outscored Qwen3.6 on terminal tasks but lagged on some knowledge accuracy and hallucination suppression measures. In separate low-precision speed tests, it was slower than two smaller mixture-of-experts models.https://devotd.net/en/news/reflection-beam-1-0-open-weights-self-hosting/https://devotd.net/en/news/reflection-beam-1-0-open-weights-self-hosting/Sat, 17 Oct 2026 17:29:39 GMTLLMsReflection released its first open-weight model in three precisions, enabling self-hosting and further training. Deployment control leaves operations teams responsible for checking hardware needs, task quality, and safety.https://devotd.net/en/news/claude-code-2-1-302-auto-mode-classifier-idle-compaction/https://devotd.net/en/news/claude-code-2-1-302-auto-mode-classifier-idle-compaction/Sat, 17 Oct 2026 17:21:54 GMTITClaude Code added settings to select the auto mode decision model and toggle idle conversation compaction. Alongside the previous version's unattended retry wait cap, this version rejects image sources larger than 128 MB.https://devotd.net/en/news/ktransformers-0-7-2-deepseek-v4-flash-mxfp4-lora/https://devotd.net/en/news/ktransformers-0-7-2-deepseek-v4-flash-mxfp4-lora/Sat, 17 Oct 2026 15:19:31 GMTLocal AIA new path keeps large language model expert weights compressed on the CPU while training only small correction weights. Saving and inference integration are supported, but full-model quality and throughput results are not provided.https://devotd.net/en/news/intel-llm-scaler-vllm-1-5-qwen-gemma-4-fp8-mtp/https://devotd.net/en/news/intel-llm-scaler-vllm-1-5-qwen-gemma-4-fp8-mtp/Sat, 17 Oct 2026 15:18:04 GMTLocal AIIntel has updated the engine behind its local language model inference package and expanded model support and low-precision execution paths. It also offers faster response generation and customized model serving. However, users must check model-specific requirements and note that the image is for demos.https://devotd.net/en/news/granite-switch-4-1-decision-guardian-speech-one-endpoint/https://devotd.net/en/news/granite-switch-4-1-decision-guardian-speech-one-endpoint/Sat, 17 Oct 2026 13:27:20 GMTLLMsIBM’s new open-weight model offers chat, choice-based decisions and speech recognition through one endpoint. It can simplify server setups, but response-specific handling and English-only speech input require consideration.https://devotd.net/en/news/llama-cpp-b11709-kolibri-1-gguf-support/https://devotd.net/en/news/llama-cpp-b11709-kolibri-1-gguf-support/Sat, 17 Oct 2026 13:20:41 GMTLocal AISupport now lets users convert and run a German-English open-weight model with a local inference tool. Only some parameters are used per token, but memory for all weights and validation in real-world use remain separate considerations.https://devotd.net/en/news/strata-0-1-45-qwen3-6-vision-parallel-requests/https://devotd.net/en/news/strata-0-1-45-qwen3-6-vision-parallel-requests/Sat, 17 Oct 2026 12:24:07 GMTITThe local language model runner adds image input and up to 8 concurrent requests while improving response generation speed. With limited memory, concurrency can slow a single request, so settings must suit the hardware and usage.https://devotd.net/en/news/fish-audio-drama-3-35b-a3b-open-tts/https://devotd.net/en/news/fish-audio-drama-3-35b-a3b-open-tts/Sat, 17 Oct 2026 11:27:01 GMTAIFish Audio previewed a speech synthesis model combining multiple voices, natural-language direction, and partial audio editing. Adoption requires checking GPU memory, latency under concurrency, removed training tools, and differing license labels.https://devotd.net/en/news/flux-3-image-klein-5-6b-open-weights/https://devotd.net/en/news/flux-3-image-klein-5-6b-open-weights/Sat, 17 Oct 2026 11:23:25 GMTAIBlack Forest Labs has released weights for a 5.6 billion-parameter image generator that handles both generation and editing. Self-hosting requires checking the memory cost of a separate prompt expansion model and the scope of conditional commercial use.https://devotd.net/en/news/deepseek-v4-1-flash-1017-open-weights-search-fetch/https://devotd.net/en/news/deepseek-v4-1-flash-1017-open-weights-search-fetch/Sat, 17 Oct 2026 11:23:23 GMTLLMsDeepSeek released a model it says improves coding scores while reducing task time and output length. Web search and page fetching are now public APIs, with temporarily free server-side calls through designated tools.https://devotd.net/en/news/openai-agents-python-0-24-mcp-annotations-tool-errors/https://devotd.net/en/news/openai-agents-python-0-24-mcp-annotations-tool-errors/Sat, 17 Oct 2026 09:28:13 GMTITCode can now read annotations supplied by external tools and format execution errors from eligible functions through a shared handler. Automatic history compaction is deferred while approval is pending. Existing error handlers need checks for the new error kind.https://devotd.net/en/news/lfm2-5-24b-a2b-on-device-agent-moe/https://devotd.net/en/news/lfm2-5-24b-a2b-on-device-agent-moe/Sat, 17 Oct 2026 08:47:15 GMTITLiquid AI has released a model for on-device agents that reasons before answering while using only part of its parameters for computation. It showed strengths in tool-use evaluations. Knowledge and coding limitations, weight storage demands, and commercial-use terms also warrant review.https://devotd.net/en/news/minicpm-v-4-7-35b-a3b-on-device-moe-vision/https://devotd.net/en/news/minicpm-v-4-7-35b-a3b-on-device-moe-vision/Sat, 17 Oct 2026 08:45:38 GMTITOpenBMB's new vision-language model preloads needed computation blocks to reduce phone delays in reading weights. The team reports results combining quantization and speculative generation. Standard inference setup and device-optimized deployment require separate approaches.https://devotd.net/en/news/offrail-27b-uncensored-bonsai-2-mtp-graft/https://devotd.net/en/news/offrail-27b-uncensored-bonsai-2-mtp-graft/Sat, 17 Oct 2026 05:50:21 GMTITA build adds a predictive aid to a ternary-compressed model with fewer refusals. The distributor's tests showed about 20% faster generation under the same conditions, but extra memory and a dedicated runtime are required.https://devotd.net/en/news/unsloth-0-1-908-beta-nemotron-muse-spark-prompt-cache/https://devotd.net/en/news/unsloth-0-1-908-beta-nemotron-muse-spark-prompt-cache/Sat, 17 Oct 2026 05:16:48 GMTITMore large models can now run locally, with less conversation reprocessing for coding assistants. Minimum memory assumes short contexts. Faster response starts do not mean faster generation overall.https://devotd.net/en/news/ray-2-59-llm-ga-token-auth-default/https://devotd.net/en/news/ray-2-59-llm-ga-token-auth-default/Sat, 17 Oct 2026 04:38:52 GMTITRay has made language model data processing and serving generally available and enabled token authentication by default for local clusters. Authentication depends on how a cluster starts. Remote cluster operators should prepare token distribution for a future expansion of default authentication.https://devotd.net/en/news/claude-code-2-1-300-mods-otel-advisor/https://devotd.net/en/news/claude-code-2-1-300-mods-otel-advisor/Sat, 17 Oct 2026 04:27:21 GMTITThe update adds input-only decorations and standard attributes for logging generative AI work. The advisor requires opt-in. Shared logs now contain more information for reproducing requests and warrant review.https://devotd.net/en/news/sglang-0-5-21-release-new-models/https://devotd.net/en/news/sglang-0-5-21-release-new-models/Sat, 17 Oct 2026 00:53:14 GMTITThe new version expands model support and changes defaults for repeated-input caching and request handling. Conflicting setup guidance for new models calls for separate checks of deployment configurations and existing service compatibility.https://devotd.net/en/news/open-webui-0-12-3-end-to-end-encrypted-chats/https://devotd.net/en/news/open-webui-0-12-3-end-to-end-encrypted-chats/Sat, 17 Oct 2026 00:50:55 GMTITTemporary chats now encrypt messages in participants' browsers without storing them on the server. Administrators must configure feature activation and user permissions separately and check communication settings.https://devotd.net/en/news/hy-memory-30b-a3b-open-memory-writer/https://devotd.net/en/news/hy-memory-30b-a3b-open-memory-writer/Fri, 16 Oct 2026 18:05:56 GMTITTencent Hunyuan has released a dedicated model that selects and updates information for long-term memory systems. It replaces the writing model in existing memory management setups. Evaluation results need to be considered alongside the fixed samples used and the scope of validation.https://devotd.net/en/news/tensorrt-llm-1-4-0rc0-glm-5-3-flash-kvcache-v2/https://devotd.net/en/news/tensorrt-llm-1-4-0rc0-glm-5-3-flash-kvcache-v2/Fri, 16 Oct 2026 15:09:38 GMTLLMsSupport for a new model's text, image, and video inputs has been added. The cache manager's Python implementation is gone, so existing settings and event collection paths need checking before an upgrade.https://devotd.net/en/news/codex-cli-0-163-config-search-mid-turn-steering/https://devotd.net/en/news/codex-cli-0-163-config-search-mid-turn-steering/Fri, 16 Oct 2026 14:56:59 GMTAIThe terminal coding tool adds a settings search panel and support for follow-up instructions during responses. Saving settings differs from applying them, and interrupting work does not guarantee automatic rollback.https://devotd.net/en/news/nodejs-26-12-wasm-required-websocket-mask/https://devotd.net/en/news/nodejs-26-12-wasm-required-websocket-mask/Fri, 16 Oct 2026 14:56:58 GMTITError handling now better distinguishes failures when features require unavailable WebAssembly or executable memory. New built-in WebSocket masking and unmasking functions let library developers move these operations out of add-ons.https://devotd.net/en/news/vscode-1-142-copilot-harness-default-canvases/https://devotd.net/en/news/vscode-1-142-copilot-harness-default-canvases/Fri, 16 Oct 2026 10:46:36 GMTITVisual Studio Code 1.142 gradually changes the default runtime for new agent chats and adds a preview of a workspace that users and agents edit together. Check policy-driven chat migration and the sandbox's actual limits.https://devotd.net/en/news/transformers-5-20-muse-spark-open-ling-3-native/https://devotd.net/en/news/transformers-5-20-muse-spark-open-ling-3-native/Fri, 16 Oct 2026 10:45:05 GMTLLMsTransformers adds support for Meta's multimodal model and native loading for InclusionAI's model. Input handling and loading are unified. Deployments still need separate checks for large-model memory needs and expert-parallel inference slowdowns.https://devotd.net/en/news/github-copilot-local-models-mai-code-1-1-flash-mxc-sandbox/https://devotd.net/en/news/github-copilot-local-models-mai-code-1-1-flash-mxc-sandbox/Fri, 16 Oct 2026 05:51:24 GMTITGitHub Copilot plans to introduce automatic device–cloud routing by late October 2026. Developers must assess execution location and command permissions separately. Device measurements and multi-model research evaluations use different conditions.https://devotd.net/en/news/nemotron-3-5-super-vl-120b-a12b-mtp-vision/https://devotd.net/en/news/nemotron-3-5-super-vl-120b-a12b-mtp-vision/Thu, 15 Oct 2026 16:42:00 GMTLLMsNVIDIA released weights for training and evaluation, plus memory-saving deployment weights, for a model handling long documents, images, and video. Built-in prediction can accelerate generation, but actual speed and quantization's quality impact must be assessed by deployment conditions and task.https://devotd.net/en/news/boole-1-7b-one-bit-storage-quality-tradeoff/https://devotd.net/en/news/boole-1-7b-one-bit-storage-quality-tradeoff/Thu, 15 Oct 2026 16:40:58 GMTLLMsA small language model stores weight signs in 1 bit, with a 292.1MB distribution file that also runs on CPUs. Its average across ten company-evaluated tasks fell by 8.28 points. Runtime memory and generation speed were not measured.https://devotd.net/en/news/asterinas-0-19-0-self-hosting-vmx/https://devotd.net/en/news/asterinas-0-19-0-self-hosting-vmx/Thu, 15 Oct 2026 14:52:54 GMTITThe Asterinas project reports that version 0.19.0 successfully builds its own kernel and boots it under nested QEMU. The set of development tools that run on Asterinas has broadened, but the documented validation is limited to an x86-64 boot test in an outer VM allocated 40 GiB of memory. Intel VMX-based support does not provide a KVM-compatible host interface.https://devotd.net/en/news/flashinfer-071-quantized-expert-parallel-moe-rubin-blackwell/https://devotd.net/en/news/flashinfer-071-quantized-expert-parallel-moe-rubin-blackwell/Thu, 15 Oct 2026 14:52:54 GMTLLMsSupport for distributing model expert computations across GPUs has expanded on Rubin and Blackwell workstations. The workstation path leaves input delivery and output summation to the caller. Published single-layer performance figures do not represent speedups for an entire distributed service.https://devotd.net/en/news/github-actions-workflow-lockfiles-preview/https://devotd.net/en/news/github-actions-workflow-lockfiles-preview/Thu, 15 Oct 2026 14:52:54 GMTITGitHub Actions lockfiles preserve readable tags while pinning executed code to recorded commits. They apply to selected workflows, and teams review and commit the generated changes. Dependencies inside reusable workflows in other repositories are managed by those repositories’ lockfiles.https://devotd.net/en/news/semaphore-open-source-1-6-mcp-sem-ai/https://devotd.net/en/news/semaphore-open-source-1-6-mcp-sem-ai/Thu, 15 Oct 2026 14:52:54 GMTITThe announcement for Semaphore Open Source 1.6 says AI agent integration will expand to CE and EE, but whether an upgrade is available now still needs verification. Semaphore says the release includes an MCP server and supports sem-ai without additional configuration, but 1.6 was absent from the public releases and tags checked on October 15. The MCP setup documentation reviewed is for the cloud and does not establish the self-hosted deployment procedure.https://devotd.net/en/news/subversion-115-pristines-on-demand-large-files/https://devotd.net/en/news/subversion-115-pristines-on-demand-large-files/Thu, 15 Oct 2026 14:52:54 GMTITSubversion 1.15 offers an option to download cached BASE file versions on demand. Apache says this can reduce working-copy storage by up to 50%. It may suit large, rarely changed files, but increases repository communication and full-file uploads for modified files. These working copies require clients running 1.15 or later.https://devotd.net/en/news/kiro-ide-1-3-local-command-sandbox/https://devotd.net/en/news/kiro-ide-1-3-local-command-sandbox/Thu, 15 Oct 2026 14:03:49 GMTITA new sandbox limits access for local commands run by AI, while saved workflows can now be launched from chat. Execution status displays have also improved, but the sandbox does not guarantee full isolation. Proposed commands still need review.https://devotd.net/en/news/aibrix-v0-8-0-kubernetes-inference/https://devotd.net/en/news/aibrix-v0-8-0-kubernetes-inference/Thu, 15 Oct 2026 08:09:41 GMTLLMsAIBrix v0.8.0's experimental ModelClaim separates model lifecycles from Pods and puts idle engines to sleep to share GPU memory. However, the first request to a sleeping model receives HTTP 503 rather than waiting for a result. If its reserved space is also shared while it sleeps, waking up depends on available capacity.https://devotd.net/en/news/vllm-tpu-inference-0310-hierarchical-moe-w4a16/https://devotd.net/en/news/vllm-tpu-inference-0310-hierarchical-moe-w4a16/Thu, 15 Oct 2026 08:09:41 GMTLLMsvLLM TPU Inference v0.31.0 adds optional dispatch that reduces inter-chip MoE communication and W4A16 linear operations on the JAX path. Development PRs reported a 4.5% reduction in total decoding step time for one MoE configuration and fit Gemma 4 31B QAT on one v6e chip. These results address different bottlenecks; they do not establish general speedups or preserved quality.https://devotd.net/en/news/openssh-10-6-upgrade-compatibility-impact/https://devotd.net/en/news/openssh-10-6-upgrade-compatibility-impact/Thu, 15 Oct 2026 07:46:47 GMTITBefore upgrading to OpenSSH 10.6, check automation that passes usernames containing $ or \ on the command line, along with workloads that rely on SSH compression. scp -R still works in this release; the new deprecation warning and the future change to the copy path are separate steps.https://devotd.net/en/news/sqlite-354-cli-prompt-escape-sequences/https://devotd.net/en/news/sqlite-354-cli-prompt-escape-sequences/Thu, 15 Oct 2026 07:26:27 GMTITSQLite 3.54.0’s interactive CLI shows the database name and open transaction status in its prompts, along with the characters needed to finish incomplete input. SQLITE_PS1, SQLITE_PS2 and .prompt let users customize the display, but this information does not guarantee SQL safety or a successful commit.https://devotd.net/en/news/embeddinggemma-2-modular-encoders-vector-storage/https://devotd.net/en/news/embeddinggemma-2-modular-encoders-vector-storage/Wed, 14 Oct 2026 17:13:14 GMTAIEmbeddingGemma 2 can load only the input encoders it needs and shorten retrieval vectors. The first choice affects model memory; the second affects storage and search quality. Google reports substantial quality loss for image, video and audio retrieval at 128 dimensions.