Qwen 3 8B

Evidence (17)

• 4Hacker News·10/10/2026research
  Build Your Own Decision Model

  The post explains how to turn a language model into a decision model by constraining outputs to a fixed set of options and selecting the highest-probability choice in a single pass. It also shows a simple implementation with Qwen/Qwen3-1.7B and reports accuracy on a CommonsenseQA sample, including a small gain after fine-tuning.

• 4Simon Willison·10/4/2026research
  Qwen3.8-27B Tested on Word-Return Addition

  Simon Willison reports an experiment running Qwen3.8-27B locally to compute sums while returning the answers in words, comparing results with and without reasoning enabled. The model was highly accurate overall, especially with reasoning on, achieving 167 correct results out of 169 attempts in the one-shot run.

• 8Hacker News·10/4/2026open_source
  Qwen 3.8 Flash Next Runs on RTX 4090 at 100T/s

  The article and accompanying repository claim that Qwen 3.8 Flash Next, a 125B model, can run on consumer hardware such as an RTX 4090 while reaching 100 tokens per second. The post generated significant discussion on Hacker News around model performance, inference efficiency, and practical local deployment of large models.

• 7Hugging Face Blog·10/2/2026open_source
  Ai2 open-sources AstaBrief report-generation model

  Ai2 is releasing AstaBrief 8B, a small open model designed to generate cited scientific reports from research questions and retrieved literature excerpts. The company says it is faster than the proprietary mode used in Asta, and is also publishing the weights, training data, and an example workflow for local use.

• 5Hacker News·9/24/2026research
  Dynamic Abliteration suppresses refusals in Qwen at runtime

  The post describes a proof of concept for dynamically suppressing refusal behavior in the open-weight Qwen3-4B model without changing its weights. It uses multi-layer steering via PyTorch forward hooks to inject signals into residual streams during inference, aiming to preserve the base model while altering responses to sensitive prompts.

• 2Hacker News·9/23/2026open_source
  Parody post shows a 25-line Python Jev classifier

  This Hacker News item presents a parody implementation of “Jev” as a small local Python script that loads a GGUF model and classifies an email into labels such as Legitimate, Spam, or Phishing. The post argues that the approach is simple, fast, and privacy-preserving because it runs locally without sending data to an external API.

• 6The New Stack·9/21/2026open_source
  Kev: a tiny Jev-like decision model family built on Qwen3.5

  Kev is an open-source family of small decision models inspired by Jev, built on top of Qwen3.5. The project appears aimed at lightweight model experimentation and practical decision-making tasks, and has drawn substantial Hacker News attention.

• 7Hacker News·9/17/2026model_release
  PrismML launches Bonsai 2 27B with 9x smaller footprint

  PrismML released Ternary Bonsai 2 27B, a compressed multimodal model based on Qwen3.8 27B. The company says it retains 98.2% of full-precision benchmark performance while using ternary weights and a 5.9GB footprint, targeting local deployment for coding, vision, and agentic workloads.

• 6Hacker News·9/8/2026research
  Qwen3.8 27B 4-bit quantization keeps benchmark quality

  A benchmark comparison of Qwen3.8 27B quantizations found that the 4-bit Q4_K_M model performs essentially the same as BF16 on GPQA Diamond, IFBench, and Terminal-Bench 2.1 while using far less GPU memory. Performance drops sharply at very aggressive compression, with 1-bit quantization collapsing to near-random results on GPQA Diamond and degrading reasoning further on longer tasks.

• 4Hacker News·9/8/2026research
  Benchmarking model and harness combos on a Three.js task

  A Hacker News post reports tests of 10 model and harness combinations on the same prompt to build a single-page Three.js sci-fi hangar scene. The write-up compares runtime, token usage, tool calls, browser checks, and estimated costs across models such as GLM 5.3 Flash Max, GPT-5.6 variants, and Qwen 3.8 27B under different harnesses.

• 4Hacker News·9/3/2026model_release
  Cerebras adds Qwen 3.8 27B to public endpoints

  Cerebras says Qwen 3.8 27B is now available on its public endpoints at about 1,500 tokens per second, alongside OpenAI GPT OSS 120B. The page also clarifies that public models are served unpruned, with selective weight-only quantization used for storage and full-precision activations, attention, and KV cache.

• 7Hugging Face Blog·9/3/2026research
  Reproducing watercolor painting training with TRL and OpenEnv

  Hugging Face details an open implementation of a coding model that generates watercolor-style images by writing JavaScript to control p5.brush. The post publishes the dataset, RL environment, training scripts, and trained models, and compares three reward mixes run end to end on Hugging Face infrastructure.

• 7Hacker News·9/1/2026open_source
  Slotstream runs a 125B Qwen3.8 model on a 48GB Mac

  Slotstream is a new approach for running the 125B-parameter Qwen3.8-Flash-Next model in 4-bit on low-memory Macs, starting from 16GB RAM. It uses expert offloading and SSD streaming to reduce memory requirements while keeping the system mac-native through MLX and Swift, with an auto mode that balances speed and memory use.

• 7NVIDIA GenAI·8/25/2026product_launch
  Perplexity brings Computer agent to local desktop hardware

  Perplexity, working with Nvidia, has moved its Computer agent from the cloud to a local desktop version called Portable Computer. The system runs on Nvidia DGX Spark or high-end PCs with Ubuntu and at least 24GB of Nvidia GPU VRAM, and uses a local sandbox plus a deterministic orchestrator to manage tool use and sensitive cloud fallbacks.

• 7Hacker News·8/25/2026model_release
  Qwen 3.8-Flash-Next scheduled for release tomorrow

  Qwen 3.8-Flash-Next, a new model from Qwen, is reported to be releasing tomorrow. The listed model appears to be a 125B parameter model with 6B active parameters, drawing strong attention in the Hacker News discussion.

• 7Hacker News·8/23/2026research
  Qwen 3.8 27B completes a reverse-engineering task in 30 minutes

  A hands-on test found Qwen 3.8 27B could work through a complex local reverse-engineering task on a single workstation and produce both a report and a bypass. The model relied on static analysis and Bash-based tools, and it corrected its own mistakes during the process. The article frames this as evidence of strong local model capability, while also highlighting cybersecurity concerns.

• 7Hacker News·8/21/2026research
  Qwen3-TTS implementation reaches sub-50 ms voice latency

  The authors report a Qwen3-TTS 1.7B CustomVoice implementation that achieves sub-50 ms p95 time-to-first-audio and about 10 requests per second on a single NVIDIA H100 SXM while maintaining real-time playback. They benchmarked multiple serving engines, tuned them for low-latency streaming, and released the implementation and benchmark as open source.