Claude Code and OpenCode with local Qwen: compatible or supported?
OpenCode documents local llama.cpp. Claude Code can traverse a compatible gateway, but Anthropic does not support routing it to non-Claude models.
Syntalith blog
Practical writing on AI agents, applications and workflow automation. Assess cost, risk and business fit before committing to a build.
Codex with local Qwen3.8-27B over the Responses API
Codex can use a custom provider over the Responses wire. This guide shows the minimum profile, support boundary and the Qwen result.
Polish company RAG: two studies, two different winners
Bielik led one Polish RAG study and PLLuM led another. Here are the numbers, why the order reversed, and how to build an auditable company brain.
qwen3.8company brainragknowledge base
How to benchmark a local coding agent honestly
A useful coding-agent benchmark needs work requests written like ordinary issues, clean workspaces, hidden checks, retained metrics and a separate review. This method evaluates a patch before acceptance.
llama.cpp or vLLM for Qwen3.8-27B on an RTX 3090?
A patched vLLM profile won one complete Qwen Code task. llama.cpp remained the simpler fallback. Here is what the evidence covers and what it leaves open.
Local Qwen Code: the medium profile from a home test
Wire-level reasoning effort, no fixed max_tokens, preserved thinking and task-scoped sessions removed observed failures in a home-PC experiment.
Local Qwen3.8-27B on RTX 3090: a runbook from the working experiment
A practical local Qwen3.8-27B startup pattern: verify the profile, inspect the API, watch GPU memory, switch to a fallback and run a restart test.
qwen3.8deploymentrtx 3090local llm
Long context or RAG? Qwen and Polish RAG measurements
Qwen recalled facts from 230,085 tokens but needed 563.4 seconds. A public 146-question Polish RAG study shows what retrieval adds.
qwen3.8raglong contextcompany brain
Maintaining local Qwen: turn failures into tests
Eight failures from a home Qwen and RTX 3090 experiment, with controls for upgrades, version changes, handover, testing and operator training.
qwen3.8maintenanceevaluationstraining
How DFlash2 speeds up Qwen3.8 and shortens context
In our measurement, DFlash2 raised median decoding from 93.3 to 120.9 tok/s while the context limit fell from 150,000 to 65,536 tokens.
qwen3.8mtpdflash2optimization
One RTX 3090 for a team: measured Qwen concurrency
One card can process requests in sequence or in parallel. We pair our task times with a public 1, 2, 4 and 8-stream RTX 3090 test.
qwen3.8rtx 3090throughputdeployment
One Qwen, four clients: Qwen Code, Codex, Claude Code, OpenCode
Four clients drove the same local Qwen3.8-27B. We compare the tasks, elapsed time, hidden tests and actual code choices. OpenCode won one import; Qwen Code won the long run.
Qwen Code in practice: CSV, pagination and import
Four small tasks revealed more than one large benchmark. Qwen Code read contracts well at medium effort, but hidden tests still found a missing validation rule.
Qwen Code: 26 minutes to a frontend, 48 to acceptance
An empty folder, one sentence and local Qwen on a home RTX 3090. The first ATS dashboard took 26:03. Here are the prompt, output and complete repair cost.
Qwen3.8-27B: 60k, 120k, 150k or 250k context?
Every configuration returned three planted facts, but 250k was almost ten times slower than 60k. Here is how to choose a useful context window.
Qwen3.8-27B quantization: Q4, Q5 or W4A16?
Q4 leaves more room for context, Q5 uses more VRAM, and W4A16 needs a different stack. Here are the profiles that fit and what the test cannot prove.
Qwen3.8-27B on an RTX 3090: results, metrics and limits
A complete record of one home experiment with Qwen3.8-27B: serving profile, timings, tokens, energy, VRAM, coding tests, failures and limits.
Qwen3.8, Bielik or PLLuM? A task-based comparison
Qwen3.8-27B, Bielik-11B-v3.0-Instruct and PLLuM-12B-chat-2512 compared by parameters, architecture, Polish use, licensing, hardware and deployment.
qwen3.8bielikpllummodel comparison
How we isolated the local Qwen agent experiment
Clean fixtures, separate worktrees, hidden checks outside the agent folder, an explicit tool list, telemetry and retained diffs. We also name the gap: shell was not a full sandbox.
local llmqwen3.8securityexperiment
Safe Qwen deployment: threat-model the path from laptop to tool
We decompose the real laptop -> private endpoint -> Qwen -> tool path, showing what the home experiment protected and what a company still needs.
qwen3.8securitydeploymentlocal llm