A technical write-up describes a method called Dynamic Abliteration that suppresses an open-weight LLM's refusal behavior without permanently altering its weights. Using Qwen3-4B as a proof of concept, the authors intercept intermediate residual streams at runtime via PyTorch forward hooks across multiple layers, injecting a steering signal derived from an 'Engram' approach instead of the traditional weight-projection abliteration technique.
Apple has introduced LensVLM, built on Qwen3.5-9B-Base, which lets vision-language models scan compressed text-as-image inputs and selectively expand only relevant portions using learned tools rather than reading everything at full resolution. The model reportedly matches full-text accuracy at 4.3x compression and beats retrieval and compression baselines up to 10.1x compression across seven QA benchmarks, plus document and code tasks.
An analysis argues that the cost of running machine learning models is dropping by orders of magnitude annually, driven by GPU efficiency gains that double roughly every two years—a pace not seen since early Moore's Law. The piece distinguishes proprietary models like GPT-6 Astra from open-weight models such as GLM-5.3-flash, noting that hosted and locally-run versions improve at different rates, with per-token pricing for frontier models not falling as consistently as costs for smaller models.
A technical blog post demonstrates how the buzzy AI concept known as 'Jev' can be replicated using an open-source Qwen3-0.6B model and roughly 25 lines of Python code. The example loads the model, prompts it to classify an email as legitimate, spam, or phishing, and extracts probability scores directly from the model's output logits rather than relying on an external API or specialized training pipeline.
A Show HN post introduces JevBench, a benchmark that measures typed decision models by cost per 1,000 decisions rather than per token, using actual token counts from 534 v1.2 test decisions. The methodology prices systems with public tariffs at their listed per-token rates, while unlisted open-weight models are priced using OpenRouter or DeepInfra hosting rates for the same or comparable weights, explicitly avoiding raw GPU rental costs. Reported figures include Jev 1.13.0 at $0.0399 per 1,000 decisions, SemIf at roughly $0.022, and Winnow-12B Q8 at roughly $0.037.
A developer has released Kev, a family of small decision-making models (0.8B, 4B, and 9B parameters) built on Qwen3.5, inspired by the architecture behind Jev. The models handle yes/no, multiple-choice, and rating questions within a single request and are compatible with TypeSafe's System One API, meaning developers can run Kev locally instead of relying on that hosted service. Full training code, evaluation data, and a web playground are included, and the models run on both CUDA and Apple Silicon hardware.
A screenshot circulated online showing the US Federal Register's search interface offering Alibaba Cloud's Qwen3:0.6B language model as a search option alongside the default semantic search. About a day after the screenshot spread, the Qwen options disappeared from the site, though an archived snapshot from September 16 confirms they existed.
OpenJev is a browser-based experiment that runs a local AI model entirely on a user's own GPU, with no backend or waitlist required. It lets users compare how quickly a model can read probability scores for a set of allowed options versus generating the same distribution token by token, using models like MiniCPM5 2B or the lighter Qwen3 0.6B for smaller devices.
PrismML, a startup founded by Caltech researchers led by CEO Babak Hassibi, has released Bonsai 2 27B, a compressed version of Alibaba's open-source Qwen3.8 27B model that shrinks it from full size down to 5.9 GB. The compressed model retains about 98% of the original's benchmark performance, up from 95% in the prior Bonsai release, while being small enough to run on PCs and possibly high-end smartphones. The company has raised a $22.25 million seed round and counts Databricks co-founder Ion Stoica as an advisor.
A new ternary-quantized model, Ternary Bonsai 2 27B, has been released, built on Qwen3.8 27B and using {-1,0,+1} weights with FP16 group scaling to shrink the model to about 1.76 effective bits per weight and a 5.9GB footprint. Despite being over 9x smaller than its full-precision counterpart, it retains 98.2% of aggregate benchmark performance across reasoning, coding, vision and agentic tasks, and supports a 262K-token context window under an Apache 2.0 license.
Nari Labs announced its Qwen3-TTS and Qwen3-ASR models achieved leading positions on Coval's voice AI benchmark, which measures latency and word error rate for speech AI systems. The ASR Fast model hit 44ms latency with 3.6% error rate at $0.12/hour, while the TTS Fast model achieved 63ms latency with a best-in-class 3.8% error rate, undercutting competitors like AssemblyAI and Deepgram on price.
A developer outlines a workflow for running local large language models on an Apple MacBook Pro M5 with 48GB RAM, combining Ollama for model management, OpenCode as the development harness, and Docker's sbx sandbox tool for isolation. The setup uses models like Qwen 3.8 27B and Gemma 4 31B, pulled via Ollama and configured through a custom sbx kit for each project.
GoKawiil