Matching stories Newest first
Reddit user runs Qwen 3.8 Flash Next locally on an 8GB GPU under CachyOS
The poster reports about 30 tokens per second on average and bursts up to 48 tokens per second, using a 3-bit Unsloth quant with a 42k context that nearly exhausts VRAM.
Simon Willison describes Battle Drones, a vibe coded RTS built with Claude
Willison built the browser RTS mostly with coding agents, mainly Claude Code, over 12 iterations and 321 commits. It includes audio-based multiplayer for two devices in microphone range, and he says he reviewed almost none of the code.
Hugging Face model Spark-X2.6-4B, a compact on-device language model, is released
Spark-X2.6-4B keeps the Spark-X2.5-4B architecture and gains from improved post-training, with three thinking levels (none, low, high). The developer claims it matches or beats models around 9B parameters on open-source benchmarks.
Qev-0.8b tiny local decision model claimed as new state of the art
8b as a new state-of-the-art tiny local decision model. The post itself is a brief community submission with no benchmark details in the excerpt.
Anthropic's Claude designs protein binders against scorpion neurotoxins
Anthropic published research showing Claude can design protein binders targeting scorpion neurotoxins.
Gradium releases Phonon, open-source on-device text-to-speech models
Gradium released Phonon, two small TTS models of 31M and 40M parameters, with a fully open-source inference stack targeting CPUs, GPUs and NPUs across 10+ backends, mostly written in Rust.
Intel Arc Pro B65 four-GPU rig posts long-context throughput sweeps for Qwen3.8 models
8 Flash-Next and 27B models on 1x to 4x Intel B65 setups with vLLM, reporting output throughput across context lengths up to 200K. Results include queueing effects and are not model-quality scores.
Pocket TTS with voice cloning shrinks to 1.9 million parameters
A community post says the Pocket TTS text-to-speech model, which supports voice cloning, is now only 1.9 million parameters.
Six-way benchmark compares Qwen3.8-Flash-Next locally against Claude Opus 5.5 on a Rust feature
A community test of 12 PRs on a LlamaStash feature found Opus 5.5 averaged 82% against 68% for Claude Code with Flash-Next and 66% for Pi with Flash-Next. Flash-Next ran locally at about 4.16 bpw, so the author paid only electricity, and results are a single developer's unverified experiment.
Ornith-1.5-122B, a Qwen-based finetune, released in GGUF and NVFP4 formats
Ornith AI introduces Ornith-1.5, which continuously generates its own training tasks and improves through reinforcement learning, extending Ornith-1.0 built on Qwen3.5 and Gemma4. GGUF and NVFP4 builds are on Hugging Face, with more details in the company blog.
Paper introduces COUNT, an n-gram method that scales language models on host memory
COUNT stores precomputed n-gram counts on host RAM instead of training learned embeddings, and reportedly matches a model with 57% more compute for 1B-parameter pretraining. The results come from an arXiv paper shared on Reddit and are not independently verified.
Community post shares a Qwen prompt that makes the model draft and refine iteratively
A Reddit user reports that adding a fractal draft-test-refine prompt to Qwen 3.8 Flash Next made it faster and more decisive in their own projects. The claims are anecdotal and no benchmarks are given.
Vals AI says DeepSeek V4.1 Flash 1017 beats GPT-6 Sol on a multi-step offensive cybersecurity task
Vals AI reports that DeepSeek V4.1 Flash 1017 outperformed GPT-6 Sol on its multi-step offensive cybersecurity task. Sol was tested through OpenAI's Daybreak trusted access program.
ISTA-DASLab releases a Qwen3.8-27B prefiller model that streams from SSD for low-quant accuracy
The demo recovers accuracy for low-bit quantized Qwen3.8-27B while keeping VRAM usage low and decoding fast. It reportedly needs an Nvidia Blackwell GPU, and the poster has not tested it. For context longer than about 4k tokens, the poster expects no speed penalty.
Reddit user asks how to annotate 500 images for fine-tuning Meta DETR
A community post seeks advice on annotation tools, using Grounding DINO or OWLv2 for pre-annotation, and quality control for a small object-detection dataset. It contains no new model news.
DeepSeek open-sources DeepInfer, the serving stack it runs in production
It ships under Apache 2.0 and for now supports only DeepSeek-V4.1-Flash in offline mode. It leads Hacker News with 1,862 points.
Berth open-source tool lets a single RTX 4090 be shared between local AI engines
Berth is a system service that swaps and loads local AI engines such as llama-swap, strata and ComfyUI on one GPU, and can serve them as an API. The project is shared as a GitHub repo.
HermesConsole and VoiceManager open-source a low latency local speech-to-speech demo
A hobbyist released VoiceManager, a streaming STT and TTS server under MIT, with a Python client and a Kotlin console reference app. Barge-in is off by default because echo cancellation is hard, and the default setup uses Nemotron and Breeze TTS 2.
Tess-4.5-27B, a Qwen finetune, claims 25% fewer tokens at the same score
Community finetune by migtissera built on Qwen3.8-27B, with GGUF files on Hugging Face. The author reports 25% fewer tokens across 150 BenchLocal scenarios and 43% fewer on data extraction, with the same scores claimed.
Developer proposes entropy-gated adaptive thinking for LLMs with parallel reasoning trees
A Reddit post describes using entropy per token to trigger extra thinking blocks, with parallel branches chosen by JEV logic. The author shares an open-source GitHub repo and says the idea targets local models that can run multiple streams.