I've been hosting Qwen3.8-27B myself. On my endpoint it's $0.30/1M in, $0.10 cache, $2.03 out - so those agent turns that re-send the same prefix get a lot cheaper when cache hits. UI at inference.tiyuvta.ai/app if you want to try it. Hosted is up to 210 tok/s and 280ms TTFT with reasoning off.
most of the time when a local model feels dumb its not the quant, its the chat template. a lot of gguf mints just drop the template from the metadata and the runtime silently falls back to chatml. model still talks fine so nobody notices, it just gets noticeably dumber. got burned by this myself serving qwen, now i grep the gguf for the template tokens before i blame anything else. second place is sampling, people run whatever defaults their ui ships instead of what the vendor recommends and then compare that to benchmark numbers that were run greedy or with the official settings
I've been comparing against TextGen and llama.cpp while I port to LocalAI and have been surprised by what's happening over the API, even with the defaults and jinja. It's been a fair reminder not to eschew familiarizing myself with the repos.
Nothing beats the classic of figuring out something yourself with your brain, but I also like dictating to LLMs a stream of consciousness with what I'm interested in (while forcing it to NOT give any answers or opinions), and getting back file names it suggests I look at and explore.
Modern frontier LLMs can still be used as rubber ducks, and it's a great.
vllm on 4x 5090 is getting ~20 tok/s with mtp on (their own thread on the hf card). i had qwen3.8-27b up the day after release, one rtx pro 6000, 140 tok/s spec, 0.156s first token, full 262k. image and video on the same api. numbers: https://github.com/avifenesh/memra
try it: https://inference.tiyuvta.ai/app
$0.38 in / $0.20 cache / $2.60 out. openrouter's only host right now is 23 tok/s at $0.45 / $3.20.
Hey HN - I built agnix because I kept losing time to the same class of bug: AI tool configs that are almost right but silently wrong.
The trigger: I had a Claude Code skill named `Review-Code`. It never auto-triggered. No error message. I spent 20 minutes debugging prompt engineering before realizing the Agent Skills spec requires kebab-case names (`review-code`). The spec is clear about this, but Claude Code doesn't validate it - it just silently ignores the skill.
This happens across the entire AI tool ecosystem. Cursor has .mdc rule files with YAML frontmatter - invalid YAML means your rule metadata is silently dropped. MCP server configs can use deprecated transport types with no warning. Claude Code hooks support `type: "prompt"` only on Stop and SubagentStop events - use it on PreToolUse and nothing happens. GitHub Copilot instruction files need valid glob patterns in their `applyTo` field - a malformed glob matches nothing.
None of these tools validate their own configuration files. It's the same gap that ESLint filled for JavaScript or clippy fills for Rust: catching things that are syntactically valid but semantically wrong.
agnix currently has 156 validation rules across 28 categories, covering 11 tools (Claude Code, Cursor, GitHub Copilot, Codex CLI, Cline, MCP, OpenCode, Gemini CLI, and more). Every rule is traced to an authoritative source - official specs, vendor documentation, or research papers.
Technical choices:
- Written in Rust, parallel validation via rayon
- LSP server for real-time IDE diagnostics (VS Code, JetBrains, Neovim, Zed)
- 57 auto-fixable rules (`agnix --fix .`)
- SARIF 2.1.0 output for CI/security workflows
- GitHub Action for CI integration
- Deterministic benchmarks via iai-callgrind in CI (blocks merge on perf regression)
- Single file: <10ms. 100 files: ~200ms.
Try it: `npx agnix .` (zero install, zero config)
I'd like feedback on two things: (1) whether the rule coverage feels right - are there config mistakes I'm not catching? and (2) whether anyone has thoughts on the cross-platform validation approach (detecting conflicts between CLAUDE.md, AGENTS.md, and .cursorrules in the same project).
For users of Node.js, Java, Python and very soon Go -
Valkey-Glide:
https://github.com/valkey-io/valkey-glide
It will stay out of the hand of Redis, can promise that, and if you use another key-val db and want to contribute to its compatibility with it, we will be happy, come talk to us.
Want to understand what dev's will appreciate and will lead decision makers to choose our client.
What will make you choose a client library over the other, and what will make consider refactoring to a new one?
ValKey-Glide, ValKey/redis-OSS client library. Multilingual wrap over Rust core (available in python and java, nodejs is 1.0 in two weeks, and go is under developments, active roadmap for C# and PHP).
OSS under Valkey org, part of Linux foundation, backed by AWS and GCP.
