15 items
Evaluated Qwen 3.6 27B across BF16, Q4_K_M, and Q8_0 GGUF quant variants with llama-cpp-python using Neo AI Engineer. Benchmarks used: HumanEval: code generation HellaSwag: commonsense reasoning BFCL: function calling Total samples: HumanE…
↯ Llama↯ Function Calling↯ Qwen 3.6humanevalfunction-callingllama+1
Hey fellow Llamas, your time is precious, so I'll keep it short. We built a GGUF port of DFlash speculative decoding.
A 1.7B model can actually turn out some code, so I'm running the training for a 9B model, then will re-run HumanEval (a full one this time). I've shown most of my homework in the article, but will be posting to github after I clean things…
Hey, I've been building side projects with Claude Code for a few months, but I'm completely new to fine-tuning — started experimenting maybe a week ago. From day one I wanted a GUI for the dataset side of the workflow, so this desktop app…
BTL-3 A 27B open-weight agent model for agentic coding and structural tool use 95.1% HumanEval · 88.5% BFCL v4 AST · 88.1% LiveCodeBench v6 (193-case run) Compact edition · Runtime source · Bad Theory Labs · Discord Introducing BTL-3 BTL-3…
ran into the q4 vs q5 wall again this morning. 70b model.
[Release] Hito 2B — structured reasoning via trained cognitive tags, +35 pts on GSM8K vs base Qwen3.5-2B (head-to-head) Been cooking this for ~6 months. Finally shipping.
Tested a bit on my XTX, a bit share hope helpful, thanks to Lucebox! Lucebox DFlash + PFlash PR #119 Reproduction Report (RX 7900 XTX) Hardware Environment Component Spec GPU AMD Radeon RX 7900 XTX (Navi 31, gfx1100) VRAM 24 GiB GDDR6 (~93…
LiveCodeBench v6, medium and hard only. 132 problems, 80 hard and 52 medium.
I've been using Claude 3.5 Sonnet for several weeks as my primary AI assistant. I wanted to share my genuine, hands‑on experience – and because this involves a comparison with another model, I've also included relevant public benchmark dat…
I have three Radeon AI PRO R9700s. I wanted to know if I could tune a quant for this specific hardware instead of just using the generic upstream formats.
In 2017, I watched OpenAI Five destroy pro players at Dota 2. That moment taught me something: games are the ultimate test of emergent intelligence.
A few months ago, I got stuck on one line in the DeepSeek-R1 paper. It said models could improve through verifiable rewards.
Ran a full eval against four local models last weekend and the spread between them is wider than I expected. All running through Ollama on CPU, no cloud, same prompts, same hardware.
↯ Ollama↯ Function Calling↯ Qwen 3.6humanevalfunction-callingollama+1