Like any distilled model, Ternary Bonsai 2 27B inherits refusals and censorship from its teacher model: Qwen 3.8 27B.
Fortunately, there’s a very simple and elegant way to reduce or perhaps even completely eliminate them. 🧵
Your US stock position still carries risk even after your broker closes.
So I built Tesrune for the Bitget_AI S2 Hackathon. Tesrune is a dark-hours AI trading desk for US equities. You add the stocks you're holding.
Talking about Local Benchmaxxing - here is my 2 year old Intel i9 64 GB machine with Nvidia RTX 4090 running Qwen 3.8 Flash next at 30 t/s.
Happily using opencode at my Macbook Pro using my Older Intel Machine as an API endpoint. Quant : AD-3.84bpw-IQ4_XS-M64
Found a sweet spot between quality and speed for Qwen 3.8 Flash on 2x3090s @ 3.5bpw, running some tests
Backend is exllamav3, running pretty sweet @ around 105 tk/s average (92 tk/s prose/reasoning, 105 tk/s coding and 120 tk/s file editing) and 1635…
RTX 3060でQwen 3.8 Flash Next
net58264Sep 21
DeepSeek V4 Flash, Qwen 3.8 27B, GLM 5.2, Nemotron 3 Ultra, Laguna S 2.1, and Jev for agent routing.
All FREE: OpenRouter: DeepSeek paths, Qwen 3.8 27B, GLM 5.2, Ling 3.0 Flash VL, Laguna S 2.1, Nemotron 3.5 Lightning / 3 Ultra: OpenCode Zen:
Qwen 3.8 Max Full COURSE 1 HOUR (Build & Automate Anything)
Want the SOP? DM me. 💬
I can't believe simultaneous bilingual output now averages about 2.3s lag across 60 input languages.
Qwen 3.8 Live Translate from Alibaba_Qwen: about 2.3s average lag (down from 2.8s), 60 input languages, spoken output in 29.
Alibaba's Qwen 3.8 Omni Flash watches video the way you do
It sees the screen and hears every word at once. Not a transcript. Not screenshots. The actual video.
Drop an hour of meeting video and get action items plus edit plans, not another summary dump.
Qwen 3.8 Omni Flash from Alibaba_Qwen takes text, image, audio, and video in. Text out. About 1M context.
People underestimate what you can do with local models.
This is Qwen 3.8-27B running on an M5 Max building a 3D racing game end to end.
This new Chinese AI reads 100 hours of your calls and meetings in one shot.
Alibaba just dropped Qwen 3.8-Omni-Flash. Most AI reads text. Your business runs on calls, videos, and recordings.
Qwen 3.8 27B's advert for a lock product from Google's GenAD-Bench, baseline run, custom video creation harness ...
Hmm, well, maybe this could protect my car. Strong opener. 😂
A 27B local model spent 21 days trying to build its own CUDA inference engine on a single RTX 3090.
Qwen 3.8 27B ran mostly unsupervised, with no human-written CUDA and only ~12 nudges across the entire experiment.
Browser use that is up to 90% MORE token efficient and FAST..
AND its FREE and OpenSource?? Real time demo using Qwen 3.8 27b, hosted locally.
What if your local LLM could play your old LAN games?
This is blowing my mind honestly This is Unreal Tournament 99 on two real Windows 98 PCs. Each player is driven by a language model.
Qwen 3.8 flash next ~60 tok/s demo
Your trading agent reads the news.
Anyone who can place a headline in front of it is writing into the context of the thing that places your orders.
Google just mogged this tame impala test.
🌀 I gave 4 fast models the same Currents cover. Same prompt. One shot via aimlapi.
Constantly amazed at what you can do at home with *significant* investment (but also not boat money).
My own agent/harness running entirely on my own hardware. It's the Qwen 3.8 Flash model running on one Spark, plus all the voice models running on a…