Qwen3.8 27B
Qwen's dense 27B open-weight model for multilingual reasoning and generation.
- By
- Qwen
- Access
- Downloadable checkpoint
- License
- apache-2.0
- Deployment
- Downloadable weights
238 entities · 24 projects · 84 tools · 111 models · 19 news
Qwen's dense 27B open-weight model for multilingual reasoning and generation.
A fast Qwen3.8 checkpoint built for efficient agent and assistant workloads.
A Qwen mixture-of-experts checkpoint balancing total capacity with lower active parameters.
DeepSeek's open V4 Pro checkpoint for advanced general reasoning and generation.
An MIT-licensed DeepSeek V4.1 checkpoint for fast multimodal text generation.
A faster open DeepSeek V4 checkpoint aimed at efficient inference.
Microsoft's compact open language model for reasoning and resource-conscious deployment.
Mistral AI's compact realtime speech model for low-latency transcription workflows.
NVIDIA's 120B mixture-of-experts Nemotron checkpoint for open reasoning workloads.
Meta's open-weight Llama 4 Scout mixture-of-experts instruction model.
Hugging Face's Rust tokenizer library for research and production inference.
A smaller instruction-tuned Phi-4 checkpoint for local and edge-friendly workloads.
An optimized 30B mixture-of-experts Nemotron release for accelerated inference.
A compact Mistral instruction model sized for controlled and local deployment.
Mistral AI's 128B-class open model for capable general-purpose inference.
Meta's larger Llama 4 Maverick mixture-of-experts instruction checkpoint.
A 119B mixture-of-experts Mistral checkpoint for efficient general workloads.
An NVIDIA embedding model for retrieval, semantic search and enterprise indexing.
NVIDIA's tabular foundation models predict classifications or numeric values from labeled examples, with three sizes spanning 28M to 215M parameters.
A compact Microsoft research checkpoint focused on agentic computer-use tasks.
H Company's 27B agent model combines graphical interfaces, code, MCP and APIs; its downloadable checkpoint carries a noncommercial license.
Self-hosted AI workspace with agents, skills and tools that runs on your own provider keys.
An MIT-licensed CLI and bundled agent skill that render live task checklists beside coding agents in Rex terminal panes.
An MIT-licensed desktop document-to-Markdown workbench with a local REST API, multiple extraction engines and original-page comparison.
An Apache-2.0 framework for coding agents to build custom BI apps on reviewed models, metrics and a warehouse-backed server.
A compact English speech-recognition model with a 163 MB packed download and local MLX, PyTorch and Core ML execution options.
An unofficial Herdr reading view for coding-agent sessions, available as a Herdr app mod and a separate native Mac app.
A community training-free weight edit of Prism ML’s Bonsai 2 9B, distributed as GGUF variants with author-reported reductions in repeated reasoning.
A local source-available coding-agent plugin that indexes tool output and returns selected context, with hooks for multiple host harnesses.
A community local AI server and desktop/web interface with model management, routing and compatible APIs for connected apps.
A dense 65B open-weight model with BF16, NVFP4 and base checkpoints; Ricursive reports strong agentic coding results and up to one million tokens of context.
A community TensorFold-based recipe for quantized GLM-5.3 across four DGX Sparks, with incremental prefix caching and component-specific licenses.
Ostris's deterministic image autoencoder family uses shallow spatial compression and patching to produce latent layouts for diffusion research.
A community Android assistant with on-device Qwen-derived inference, English and Greek support, and optional phone actions.
Independent command-line utilities for searching, resuming and porting coding-agent sessions, alongside workflow helpers.
A DeepSeek-R1-Distill-Qwen-7B fine-tune trained to shorten reasoning using preferences at intermediate states.
Community GGUF quantizations of a refusal-ablated LFM3-Omni-2B derivative, distributed for safety research and red-teaming.
An open-source autoresearch harness that coordinates agent experiments, separate judging and live campaign monitoring on user-controlled machines.
SenseNova's open-weight multimodal preview reuses a shared decoder block to perform 56 layer executions with 32 distinct layers.
Community VM tooling that lets MCP coding agents and Pi create, fork and operate Linux desktops through a host-controlled credential and grants layer.
LokkenJP's community activation-guided quantizations of TenStrip's MiniMax H3-based 10Eros Max video model, including Beta7 fix1 HQ and an older compact checkpoint.
A community orchestration layer connecting Pi coding-agent conversations with Herdr terminal sessions, workspaces and background delegation.
abenzerps's community Qwen-Image-2.1-Turbo conversion package offering GGUF, Safetensors and MLX weights with separate text encoders and a VAE.
A community proxy and client plugins for context reduction, prompt-cache management and experimental tool deferral in coding-agent sessions.
Anthropic's open research artifact for cryptographic proofs that a model response matches committed weights, without publishing those weights.
A community browser dashboard for upstream llama.cpp, local GGUF management, coding-agent connections and current-versus-PR build benchmarking.
A community Rust and WebGPU inference engine specialized for small models, with browser chat demos and a Chrome or Edge DOM-agent extension.
Angronain's community Qwen3.5-2B weight edit targeting refusals and self-doubt loops, distributed as GGUF files for local inference.
A self-hosted AI-text detection application combining 23 engines with calibrated scores, document scanning and inspectable per-engine results.
IBM's Apache-2.0 generalist-agent harness for cross-application business workflows, configurable tools, execution policies and private deployment.
A local-first personal agent application with file-based memory, configurable model providers, permission modes and a paired mobile preview.
BoxyFit's community TQ2_T quantization of Groot-65B-A6B, using the Strata runtime and an expert cache to target consumer GPUs.
An MIT-licensed local SQLite memory layer for coding agents, with hooks, an MCP server and rule capture from user corrections.
Gartner's forecast that open-weight models will execute 90% of enterprise operational agent workflow steps by 2030, as reported by The Register.
A Rust inference server for local MoE models, GPU and host offloading, embeddings and multi-Spark expert routing.
Hugging Face's proposed testing format combining small frozen model fixtures, executable Markdown usage paths and human-reviewed output baselines.
A research framework for continuous video generation with released 1.3B few-step checkpoints and a multi-GPU WaveRT streaming runtime.
InterpreticAI's 1.86 GB IQ3_XXS quantization targets browser and mobile inference, with author-reported losses in general capability and more degenerate answers.
A community toy-model family with expressive small chat models and a custom Hive architecture that links a queen and shared-weight workers through hidden states.
A Mac agent workspace with local and cloud models, tools and memory; version 0.25.31 adds native reasoning tiers and improves long tool-chain compaction.
Linux desktop hotkeys for local-model chat, translation, explanations and dictation, with a default FastFlowLM setup for AMD Ryzen AI NPUs.
Experimental community GGUF conversions of Moebius image-inpainting weights for CrispEmbed's native engine, with four published precision variants.
An experimental community checkpoint described as based on Qwen3.8-Flash-Next, with BF16 weights and a Q4_K_M GGUF distribution for llama.cpp.
AgentScope's self-hosted personal AI assistant workstation, connecting local or cloud models to skills, plugins, memory and messaging channels.
An independent 27B multimodal model for conservative Safe/NSFW image classification with structured JSON output, described by its author as based on Qwen3.8-27B.
An independent dense 9B computer-use model built on Qwen3.5, with downloadable BF16 weights and native tool-call formatting for an external desktop or browser runtime.
A community Python script that copies a ComfyUI NVFP4 safetensors file while clearing full-precision matrix-multiplication flags in its quantization metadata.
A standalone Linux AI shell for operations and troubleshooting, with interactive terminal programs, configurable model endpoints and MCP tools.
Daemonrat's community Qwen3.8-27B GGUF experiment transfers a two-tensor refusal-behavior edit onto an Unsloth quantization.
An Apache-2.0 native inference engine that serves supported local language-model checkpoints through Metal or CUDA and compatible HTTP APIs.
An ISC-licensed terminal agent running inside a live Common Lisp image, with self-modification, provider connections and persistent sessions.
An MIT-licensed daemon, terminal dashboard and optional desktop client for monitoring and orchestrating multiple coding-agent sessions.
Akhaliq's community LoRA for two-step image editing on Bria Fibo 2.0 Edit Turbo, with author-declared non-commercial terms and numerical evaluation pending.
An AGPL-licensed community tool for training-free language model modification using configurable modifiers, contrasting prompts and scorer plugins.
An MIT-licensed headless coding-agent harness that preserves complete UTF-8 input, tool output and session history in its own transport.
An MIT-licensed desktop workspace for coordinating AI agents, tasks, memory and model connections.
lvladikov's community distillation adapter reduces Krea 2 Turbo sampling from eight steps to three, with stock ComfyUI and diffusers files.
WAS's community model patches combine Pika H3 Flex adapters with four-step MiniMax H3 PTD files for a ComfyUI workflow producing video and audio.
Training-free sparse attention research for video diffusion, with public kernels targeting NVIDIA Hopper GPUs.
Image-generation research combining internal distribution matching with the generator's own EMA feature statistics for one- and two-evaluation sampling.
Self-hosted Playwright browser control plane with MCP and REST access, operator takeover, saved login profiles, approvals and signed action receipts.
Local AI-agent memory over SQLite, with evidence history, corrections, dependency review and optional semantic retrieval through MCP, CLI, TypeScript or HTTP.
A community experiment that compiles integer neural networks into ZDoom maps, using game geometry and voodoo dolls for arithmetic.
YC's Pi-based harness lets model-backed agents play Minecraft through Mineflayer, exchange direct messages, keep state and use configurable shell environments.
akhaliq's community fusion of Pika H3 Flex and LightX2V MiniMax-H3 Turbo adapters, offered as four-step and eight-step LoRAs for local video with audio.
Le Chonk is Mistral's open-weight multimodal mixture-of-experts model for text and image input, with text output and configurable reasoning effort.
A community mixed-precision MLX checkpoint of Qwen3.8-Flash-Next, including vision, n-gram embeddings and an MTP head, re-quantized on October 29.
An open-source model training and local AI application with desktop interfaces, LoRA fine-tuning and model-serving workflows.
Four GGUF quantizations of Mammoth-1 with decision metadata for Magnitude's structured probability endpoints.
A community HTTP proxy that archives coding-agent history verbatim and manages prompt masking and recall in front of compatible model servers.
An open-weight model for prompted recorded-audio transcription, translation, speaker attribution and English audio questions.
An AGPL-3.0-only desktop AI studio combining local chat, media generation and agent workflows, with optional cloud services.
Ruach Studio's music-generation fine-tunes of YuE2-3B, with merged GGUF and safetensors weights including dedicated ComfyUI checkpoints.
A Llama 3.1-derived roleplay fine-tune distributed in GGUF formats, with author recommendations for character cards and repetition-control settings.
An MIT terminal coding harness and SDK with selectable model providers, permission modes and persistent local sessions.
A standalone local desktop assistant that searches an offline emergency reference library using models run through Ollama.
Tencent Hunyuan's interactive video world model, with multi-turn text instructions, camera control, downloadable weights and a documented single-GPU inference configuration.
Liquid AI's compact open-weight family for local text agents, vision, audio, multimodal interaction and structured decisions.
An Apache-2.0 compact text model for local agent workflows, with configurable thinking budgets and multiple documented inference runtimes.
An Indian AI inference-software company that adjusts model deployment and operation for application-specific speed and cost requirements.
A local Python research dashboard that coordinates multiple AI analyst roles and offers optional paper-trading integration.
CCORE's independent Qwen3.5-4B fine-tune for general coding and interactive Three.js scenes, available as preview GGUF builds.
A self-hosted Linux service that turns authorized ready GitHub issues into sandboxed Claude Code or Codex runs.
An unofficial instruction-editing LoRA for MOSS-Image-SFT with public checkpoints, comparisons and ComfyUI workflows.
OrcaRouter's 27B security-focused vision-language model, distributed as GGUF files with MTP and optional DFlash drafting.
The October 26 releases bring onboarding and connection fixes to a project that picks an open-weight model for the user's computer.
The October 25 release adds journaled memories, webhook deduplication and an opt-in Linux sandbox with a clear network limit.
The model card positions the 33B checkpoint as a teacher for future distillation and reports strong, but mixed, agent benchmarks.
The October 26 account describes an August-to-September incident fixed September 9, not a newly active breach.
ISTA-DASLab's mixed-precision GGUF builds of DeepSeek V4.1 Flash 1017, with a documented Strata runtime path and substantial disk requirements.
Community Qwen3-VL-8B research checkpoint pairing a language-model LoRA with a gated residual Q-Former on image tokens and an MM-Vet ablation runner.
Qwen3-8B NoThink post-training checkpoints released as thinking-leakage research artifacts, with a separate community imatrix GGUF conversion.
RefinedNeuro's community Qwen3.8-27B weight modification targets formulaic English prose, with author-reported writing evaluations and downloadable GGUF builds.
MOSS-Image-SFT on Apple MLX, progressing from a contributor fork to a native MFlux pipeline with an MLX Qwen3.5 text encoder and complete community quantized packs.
Apache-2.0 managed agent platform with a durable Postgres event log, model-provider choice, machine pools, tool approvals and Docker Compose self-hosting.
AGPL-3.0 Chromium companion browser driven by MCP agents, with a live dashboard, persistent browser state, local replay and a Docker/noVNC deployment route.
Source-available team agent platform with isolated session sandboxes, configuration in Git, human-reviewed change requests and pi or OpenCode harnesses.
Community EXL3 quantizations of Qwen3.8-Flash-Next, including a self-calibrated 1.66bpw branch whose author claims single-RTX-5090 agentic inference with a 64K quantized cache.
Apache-licensed coding model with 39.8 billion total parameters and about 2.5 billion active per token, with a dedicated local inference engine and PI/OpenCode recipes.
A pretrained relational foundation model for classification and regression across linked tables using historical labeled examples as inference context.
IBM's Apache 2.0 speech-language model for multilingual transcription and translation, with language prompts and keyword lists of up to 2,000 terms.
Dharma-AI's previewed method for adapting small-model vocabularies around inference cost, combining domain subwords and repeated multiword spans.
An MIT-licensed cross-harness session capture and reviewed memory system for coding agents, with local and optional hosted routes.
An MIT-licensed Python modeling platform that generates application scaffolds and offers an experimental LLM layer for customization.
An MIT-licensed Rust desktop app, CLI and MCP server for vulnerability triage, with optional local llama.cpp and Apple-silicon MLX inference.
A self-hosted TypeScript agent runtime with tools, workflows, distributed workers and OpenTelemetry GenAI tracing.
A community GGUF conversion of Google's Gemma 4 12B QAT model, rebuilt as MXFP4 after its author identified an earlier mislabeled hybrid.
An MIT-licensed filesystem substrate giving coding agents writable workspaces over shared committed history, with a DeepSeek Harness web plugin.
NVIDIA's October 2026 family of multimodal embedding, reranking and document parsing models for retrieval systems and AI agents.
An Omarchy Quattro bar plugin that monitors local model memory, context usage, runtime activity and connected clients.
An open-source AI-agent observability platform with tracing, evaluations and Signals for investigating behavior across agent runs.
A self-hosted orchestrator that runs Claude Code and Codex sessions in Firecracker microVMs, with snapshots, automations and a dashboard.
TechCrunch reports a $1.5 billion Tenstorrent financing at a $6.2 billion valuation to expand chips for open-weight inference and video generation.
Comfy-Org's component packaging of MOSS-Image Base and SFT, with BF16 and int8 diffusion weights, Qwen3.5 text encoders and FLUX.2 VAE files.
Python library and MCP server that scopes text-to-SQL schema selection to caller permissions, with OAuth identity on its HTTP transport from version 1.9.0.
An MIT-licensed native inference engine and local coding-agent stack for selected DeepSeek models and project-specific weights.
An MIT-licensed Rust penetration-testing harness that coordinates model-backed agents, evidence review and reports for authorized security assessments.
An MIT-licensed Python decision compiler that uses an LLM during compilation and a frozen local model for classification, scoring and routing at runtime.
A Mac meeting recorder with local transcription, Markdown records and an agent skill for recording and finding past meetings.
A self-hosted browser workspace for persistent terminals and structured coding-agent sessions, formerly called MidTerm.
Open Qwen3-based research agents that alternate continuous latent computation with text tool calls and include a nearest-token trace inspector.
An independent personal WhatsApp read-and-search MCP server hosted in the owner's Cloudflare account.
Community 3-, 4- and 5-bit EXL3 diffusion Transformer checkpoints for a custom AMD ROCm image-generation runtime.
An independent local MCP server that lets compatible AI hosts operate Adobe InDesign through document, layout, typography, image and export tools.
A standalone ESPHome hardware-design application with a web interface, optional Claude agent and MCP tools for design, schematic, enclosure and hardware workflows.
A 140M typed-decision model and calibration library that return probabilities with an automate or escalate flag.
A community 321M CPU-oriented classifier preview for prompt injection, tool-result inspection and output checks in English, Arabic and Persian.
Lodestones' work-in-progress community distillation of MOSS-Image-Base, with a standalone LoRA route and an eight-evaluation data-space sampler.
Infatoshi's community measurement bundle for a locally served, attention-edited DeepSeek V4.1 Flash 1017 checkpoint on four Blackwell workstation GPUs.
Local Inference Lab's alpha Qwen3.8 Flash Next package with prepared kernels and execution graphs for the Flare runtime on GB202 Blackwell hardware.
A community bundle of Prism ML's PQ2_0 Bonsai 2 model and a separate inference-time refusal-steering control vector.
A Python agent framework with MCP tools and a manifest-based project runner for tracing, terminal interaction and HTTP serving.
An independent desktop interface for installed coding-agent clients, with project sessions, worktrees and experimental remote hosts.
A downloadable English text-to-speech preview with a declared 1.03-million-parameter three-stage core and a local Python runtime.
An independent local C engine for GGUF inference, LoRA adaptation and reproducible run records.
SimpleTuner's public journal comparing MOSS-Image Base and SFT concept-training runs, with recipes, adapter checkpoints and validation previews.
A standalone headless Python agent harness with opt-in tools, SQLite-backed memory and adapters for local and hosted models.
Akhaliq's community four-step distillation of MOSS-Image-SFT, with downloadable adapters, a merged transformer and a public Gradio demo.
WaveCut's community quantization of MOSS-Image-SFT, packing the image transformer and text encoder into an 8.8 GB bundle for a documented Linux and ComfyUI workflow.
A browser simulator that previews local-model streaming speed, first-token delay and memory fit for selected hardware and quantization settings.
Qualcomm's developer-preview on-device inference runtime for Snapdragon, exposing GGUF and chipset-specific model backends through CLI, Python, Android and a local API server.
Microsoft's compact repository-level coding model derived from Qwen3.5-4B, with a community mixed-precision GGUF build for local text inference.
An MIT-licensed OpenClaw management harness with browser onboarding, gateway monitoring, crash recovery and Git-backed workspace sync.
A community llama.cpp port that serves local language and decision models from a jailbroken PS5's GPU.
An Android EPUB reader with an offline name index, configurable AI reading companion, translation and read-aloud support.
Reflection's Apache-2.0 text model with 501B total and 23B active parameters, released in BF16, FP8 and NVFP4 formats.
An Omarchy desktop plugin that serves local models through a keyed gateway and opens coding agents against them.
A self-hosted model-serving platform with local and distributed deployments and OpenAI-compatible APIs.
A multi-provider coding-agent application with terminal, editor, desktop and web interfaces, plus opt-in experimental Mods for trusted headless sessions.
A research training objective that assigns semantic-span credit to a Qwen3-8B prompt rewriter for frozen text-to-image generators.
Meta's 816B-total, 42B-active mixture-of-experts model for coding and agentic work, with multimodal inputs and community GGUF quantizations.
NVIDIA and AI4Bharat's announced synthetic Indic pretraining corpus, paired with a public NeMo Data Designer translation and evaluation plugin.
A native Android coding-agent app that edits projects, runs shell tools and connects to cloud providers or compatible local GGUF models.
A Zotero research-library connector with an MCP server, a CLI and agent skill, and a separate in-Zotero chat plugin for installed coding agents.
Community weighted GGUF conversions of OBLITERATUS's refusal-modified 28B Feyospace cybersecurity model, with listed files from 11.0 to 22.5 GB.
A self-hosted single-developer dashboard for applications, coding-agent orchestration, personal knowledge and creative media workflows.
A standalone orchestration harness that coordinates coding workers, independent acceptance and code reviewers, bounded repairs and optional Git delivery.
Pipecat's Mini and Max text models for voice-agent pipelines, with tool calling, Pipecat Flows training and short reasoning in Max.
A standalone Go personal VM cloud and retro browser desktop with local Claude Code and Codex terminals and a separate Projects issue tracker integration.
IBM's modular 8B and 30B checkpoints combining Granite chat, Decision and Guardian activated LoRA adapters, and an English speech-to-text encoder in one serving endpoint.
A community MXFP8 conversion of the distilled FLUX 3 Image Klein transformer for ComfyUI, with laptop timings, paired samples and a compatibility patch.
lethalbeats' community EmbeddingGemma 2 quantization with a 73.82 MB weight file, a custom CPU PyTorch loader and author-reported multimodal inference.
Black Forest Labs' downloadable image-generation and editing model, with a 5.6B image transformer, multiple references and prompt-based layout boxes.
Mistral's downloadable time-series forecasting family, with CPU or GPU inference and 5M, 28M, 123M and 209M parameter checkpoints.
Independent MIT-licensed Gitaly fork with a server-centric CLI, per-agent tokens, branch policies and optional external conflict resolution.
Community native NInfer conversion of Prism ML's ternary Bonsai 2 27B, with vision and speculative-decoding components for the RTX 3090 fork.
Source-available desktop meeting and interview assistant with on-device transcription, local or cloud model choices, and an overlay interface.
The Forecasting Company's approximately 256M-parameter open-weight model for probabilistic time-series forecasts with historical and known-future covariates.
A target-specific speculative draft for accelerating the Humanizer rewriting model through a tested Apple silicon MLX research runtime.
Richard Young's community GGUF builds of an abliterated Granite 4.2 3B model, with IQ quantizations added on October 15.
A Java 17+ library for local ONNX inference, with a new source-level prompt-injection detector for user input and retrieved documents.
A Rust code-graph and project-memory tool that gives coding agents cited context through one MCP server.
Ai2's climate simulation framework combines a coarse atmospheric emulator with diffusion downscaling for 3 km precipitation, wind and pressure fields.
A community weight-format conversion of the Lumimaid-Nexus narrative-model merge for inference engines expecting classic Mixtral expert tensors.
A compact Qwen3.5-4B conversion whose author reports lower GPU context memory using the dedicated ThinCtx runtime and system RAM.
A compact aesthetic photo-cropping model built on Qwen3.5-0.8B, with downloadable task weights and a public interactive demo.
A JoyFox conversational model for character roleplay and shared storytelling, released with BF16 weights and Q4-Q6 GGUF builds.
A Qwen3.5-9B derivative trained to recover from failed terminal tasks, with training code and a recovery-trajectory dataset.
Research tokenizer and generator checkpoints that produce text through parallel predictions over a coarse-to-fine hierarchy.
An open-source personal AI assistant for one user, with memories, tasks and scheduled reminders in the user's own Cloudflare account.
A local inference server tuned for Qwen3.8-Flash-Next on AMD Strix Halo, with optional NPU models.
A Mac budgeting app with an on-device AI assistant and optional connected bank syncing.
Datalab's downloadable CPU-focused OCR package combines a small recognition model with document layout and reading-order output.
A browser playground for comparing Koda-IDS defensive-security models in full precision and four-bit GGUF form.
Devseis published a quantization-aware 0.5B model build that drafts explanations for on-device endpoint audit findings.
Liquid AI released a 24B-parameter, 2B-active open-weight reasoning model for local agent and tool-use workflows.
OpenBMB's open-weight multimodal model family includes a 35B-A3B mixture-of-experts variant and a smaller 1B variant.
An independent open-weight Qwen 3.5 fine-tune built to return typed decisions rather than chat responses.
An experimental community model merge for prose and roleplay, now available in weighted GGUF quantizations.
An open-source parametric CAD application in Rust with desktop and browser interfaces and an MCP server for agent-driven modeling.
LAION's experimental text-to-speech continuation with public interim checkpoints, reference-voice controls and a listening demo.
A self-hosted queue and container runner for the pi coding agent, with job budgets, forge triggers and selected MCP servers.
A community INT8 quantization of Anima Turbo v1.1 for local anime and illustration generation with Nunchaku.
Security engineer Micah Lee urged readers to move from OpenAI and Anthropic to open-weight models or private inference, prompting an October 12 developer debate.
An experimental local language-model engine that streams mixture-of-experts weights from SSD when models exceed available RAM.
An open-source local inference server for language and media models, with a macOS app and Linux NVIDIA builds.
Underdog's 7.89 GB GGUF version of Qwen3.8-27B targets local tool calling, with creator-reported benchmark results and more than 100,000 cumulative Hugging Face downloads.
Cactus Compute's on-device speech model packages seven-language transcription, word timestamps and audio embeddings in a creator-described 16.9 MB file.
PYMNTS cites an October 10 FT report of early Nvidia-Reflection talks spanning a possible acquisition, staff-and-licensing arrangement or further investment.
Microsoft opened US preorders for its Nvidia RTX Spark-powered Surface Laptop Ultra at $2,599.99; The Verge reports an October 16 launch.
TechCrunch reports a $90 million Series B for Nous Research at a $1.5 billion valuation as the company introduces Hermes for Businesses.
Liquid AI released two downloadable decision models for text, images and audio, with edge-device speed and benchmark claims that still need independent testing.
Google's 740-million-parameter open multimodal embedding model for local search and retrieval across text, code, images, video and audio.
Mistral announced Large 4 and said its open weights would be made public on October 27; Reuters reported company claims and a contained safety-test incident.
Reflection AI introduced Beam, a 501-billion-parameter sparse model, while keeping its weights in final testing ahead of a planned October release.
Aleph Alpha's German-English open-weight reasoning model for self-hosted deployments.
Axios reports that Nvidia-backed Reflection AI is preparing an open-weight model and a related enterprise AI factory offering; no model has been released yet.
Ai2 released an open-weights 8B model for generating cited scientific reports and added it as Fast mode in Asta.
AWS-affiliated Strands Labs released an open 2B-parameter decision model for choosing among supplied options in agent workflows.
Ai2 released an open MoE training-stack upgrade and reported scaling benchmarks, while its next Olmo model remains in development.
IBM announced a self-hosted deployment option for its Bob AI software-development platform, including customer-managed and air-gapped environments.
Magnitude launched an open-source local inference engine for AI agents that tunes its kernels to a user's hardware.
An Apache-2.0 local AI engine built around Apple's MLX ecosystem.
Reasoning model release that made open reasoning systems a central developer conversation.
Google's open model family designed for capable, efficient multimodal applications.
Open-weight reasoning model intended to run in environments controlled by developers.
Alibaba's open model family with reasoning and general-purpose variants for developers.
A short brief on why open-weight releases move model choice toward infrastructure design.