Ollama

Erik Markus Kannike released qwen3-4b-deslop-v4, which needs no system prompt. On 40 held-out prompts, em dashes fell from 20.5 to zero and 68% of answers landed in human-writing range.

Commenters found it still reads as AI, and one asked why anyone would want that.

Tobi Knaup argues that capable open-weight models are becoming platforms rather than merely self-hosted alternatives. The emerging stack already includes vLLM, SGLang, llama.cpp, Ollama, MLX, fine-tunes, adapters, and deployment tooling.

HN readers liked the interoperability argument but challenged the analogy’s missing ingredient: frontier model training demands capital on a scale that open-source software never did.

The three flagship models are small by frontier standards: Google's Gemma 4 E2B (10.28 GB, 128K context), Cohere's North Mini Code (19.38 GB, 500K context) and Liquid AI's LFM2.5-VL (3.2 GB) — all sized to run comfortably on a single MacBook.

Not everyone bought the pitch: one reader questioned whether 'frontier' still means anything at this size, and another couldn't say what Nativ does that LM Studio and Open WebUI don't already do today.

The closest anyone gets: GLM and Qwen run agentically on self-hosted boxes and feel near Haiku, but the thread's experienced voices agree none of them reach Opus 4.8 or GPT 5.5 for real work.

The most-upvoted reply isn't an answer, it's a complaint about the genre: 'I use Qwen and get great results' tells you nothing without the quantization, llama parameters, context window, GPU and VRAM behind it.

One reader running a local Gemma and Qwen setup on a GPU-less box says he does plenty of AI coding for almost nothing; another shipped three client apps this month on 15 cents of DeepSeek API.

The real economics the thread surfaces is blunter: there's no point running local models while VCs subsidize frontier subscriptions to absurd token ceilings. 'Going broke' is mostly self-inflicted by long, unmanaged sessions.

The guide wires up an offline coding agent on a 64GB M1 Max: llama.cpp built with Metal, Gemma 4 in Q4, and a Multi-Token Prediction draft model that lifts generation from 58 to 72 tokens/sec.

The thread's sharpest critique wasn't the setup but the metric: every local-AI post benchmarks tokens per second, none benchmark answer quality. 'Quickly serving me slop doesn't make it more useful.'