Real-world
Coding
Concrete workflows, including the tradeoffs and limitations reported by their authors.
this collection
Coding a SaaS product on an M1 Max, one well-scoped feature at a time
Running Qwen 3.8 27B on the same Mac they work on, this developer credits per-feature domain glossaries and tight scoping — not the model — for making local coding reliable.
98% local, on a self-extended Pi Agent 'software factory'
DeepSeek V4 Flash Vision on two DGX Sparks, wrapped in a heavily extended Pi Agent setup, handles nearly all of this person's hands-off work.
The first local model this person actually trusts with real work
On a 12GB 5070, a heavily quantized Qwen 3.8 27B through a custom Hermes backend is, in their words, the first local model that's felt confidently productive.
Switching back a model version to escape endless self-checking loops
Qwen 3.8 kept spiraling into overlong, indecisive reasoning on this person's coding tasks, no matter the configuration — so they moved back to Qwen 3.6.
Using a local model to stress-test the agent a coding agent is building
An unusual layered setup: the local model isn't the coding agent — it's the thing an outer coding agent uses to build and stress-test a separate 'inner' agent.
Rotating between three harnesses to find what actually works
After trying Claude Code with task-farming and a Qwen-as-orchestrator experiment, this person settled into two parallel workflows — Orca as the local workhorse, and OpenCode or Bionic for solo agent development.
Good enough for 80% of coding, honest about the other 20%
Daily local coding on a 4070 is explicitly not as good as Claude for complex refactoring — but the privacy and iteration speed still make it worth keeping.
Deep coding work on a 128GB Strix Halo
An experienced software engineer finds sub-30 tok/s local inference on a 128GB Strix Halo more than adequate for deep coding work.
Finishing work integration tests with a local OpenCode setup
After hitting a cloud coding usage limit, a developer finished a real work task with Ling 3 Flash on an Asus GB10 through OpenCode.
80% local writing and scripting on a 128GB MacBook Pro
A MacBook workflow keeps most writing, review, summaries, and shell/Python/YAML scripting local, with Apple Shortcuts loading different models for different tasks.
Primary work coding on a 3090 and 5060, cloud for deep expertise
A developer uses Qwen 3.8 27B for primary coding on a 3090-plus-5060 setup and pays for frontier inference only when speed or model depth matters.
Iterative coding with Qwen 27B on a 4070 Ti Super
A developer uses Qwen 27B as a carefully supervised coding partner inside Cline, with explicit planning and changelog-based context management.
Building a game iteratively with a local Pi harness
A non-professional developer uses a local model in small, iterative steps to build a long-term personal game project.
Qwen 27B coding and infrastructure on two 3090s
A two-3090 local setup runs Qwen 27B for coding and infrastructure work, including bug fixing and quality-of-life automation.
Building Dynatrace apps with local Qwen and a codebase manifest trick
A site reliability engineer builds custom apps and a theming engine inside Dynatrace with local Qwen, using a per-repo manifest so the model only reads the code it actually needs.
A senior engineer with no need for cloud on a 64GB Strix Halo
An experienced software engineer runs Qwen3-Coder 30B on a 64GB Strix Halo and finds local working better than newer, larger models for well-scoped tasks.
One R9700 with card-specific kernels for local coding
A single Radeon Pro AI R9700 runs coding work using Lucebox and a recommended 4-bit quant, relying on custom optimized kernels for the card.
A professional developer who prefers a good local model because it keeps him in the loop
With access to both Claude Fable and high-end local models, this developer chooses local because a less autonomous model keeps him involved in debugging and research work.
Local did not pay off: the electricity bill outweighed a cheap cloud subscription
An honest counterpoint: with public code and a 5kWh-per-day rig at 28 cents per kWh, paying for a stronger cloud model was cheaper than running local.
A llama.cpp fork reaching 90+ tok/s through a 100K context on a 3090
A custom llama.cpp fork for Ampere cards pushes Qwen 3.8 27B to 90+ tok/s through a 100K context, aimed at caching-heavy agentic work.
Three concurrent coding sessions at 150 tok/s on one 5090
A developer with 20 years of experience runs three concurrent coding sessions at about 150 tok/s each on a single RTX 5090 and has been cloud-free for three months.
Two Radeon RX 9070 XT cards as a 32GB budget rig
A compact configuration note: two RX 9070 XT cards give 32GB total VRAM and 50-60 tok/s at a 200K context on Qwen 3.8 27B in MXFP4.
Qwen 27B on SGLang at NVFP4 with a 180K context
A short but specific configuration note: Qwen 3.8 27B served with SGLang in NVFP4, deliberately capped at a 180K context to stay out of the degradation zone.
Qwen 3.8 27B at 100K context on a single RTX 4080 16GB (NInfer 4080)
A purpose-built engine (NInfer 4080) runs ISTA-DASLab's Qwen 3.8 27B GSQ at 100K+ context on one RTX 4080 16GB, measuring up to ~2,720 tok/s prefill and ~262 tok/s decode through DFlash2 speculative decoding.
Shipping seven Kotlin Multiplatform projects on a fully custom local harness
Not vibe coding: a custom harness on a patched llama.cpp build has shipped seven real KMP projects.
A context-aware router working on the first try
A router in front of compiled llama.cpp servers, feeding Qwen 3.8 27B through OpenCode with downloaded skills, worked essentially out of the box.
Eight A100s in tensor-parallel, inside a company with a hundred more
A large enterprise deployment: eight A100 80GBs running Qwen 3.8 under SGLang with tensor parallelism, while the company brings newer B200s online for bigger models.
Qwen Flash Next on a 12GB RTX 3500 Ada laptop
A laptop workflow combines Qwen Flash Next, llama.cpp and a DeepSeek harness, using system memory and SSD-backed loading to make a MoE model practical on 12GB of VRAM.
A 524K-context Qwen 27B building a Godot game for the family
A long-context Qwen 3.8 27B under a DeepSeek harness handles multi-hour agent runs, building a Minecraft-style voxel game, a React site and a physics module.
A cloud-free workstation: RTX Pro 6000, vLLM and a DeepSeek harness
A custom Linux workstation with an RTX Pro 6000 and 48GB of RAM runs programming, copy editing and summarisation through vLLM, with no cloud use all year.
Coding on an 8GB 5060, where tuning the GPU/CPU split matters more than tok/s
A daily local coder on an RTX 5060 8GB and 32GB of RAM finds responsiveness under memory pressure matters more than benchmark speed.
A workstation GPU as a drop-in subagent for Codex and Claude
Qwen 3.8 Flash Next on an RTX Pro 6000 takes implementation and review tasks delegated from Codex or Claude, replacing Sonnet and Opus for that slice.
Maintaining a DAW, an adblock proxy and a notes app on a Strix Halo
A developer runs Qwen 3.8 Flash Next on a Strix Halo through a Pi harness and reports 40-50 tok/s decode plus 1300 tok/s prefill while maintaining three substantial projects.
Fitting Qwen 27B and a 96K context onto a single 3090
A 3090 owner runs Qwen 3.8 27B daily in Cline at a 96K context by quantising the KV cache to q8_0, reporting about 43 tok/s once the model stops thinking.
Agentic coding on a 512GB Mac Studio with oMLX
A 512GB M3 Ultra Mac Studio running models through oMLX handles most agentic coding work, with measured speeds across several models listed by the owner.
Qwen3.8 Flash Next on a Radeon 7900XTX and 64GB DDR5 with Strata at about 60 tok/s
A Radeon 7900XTX with 24GB VRAM and 64GB DDR5 runs Qwen 3.8 Flash Next through Strata at about 60 tok/s up to 250k context while leaving VRAM for gaming.
Qwen3.8 Flash Next 177B at 11-15 tok/s on a single RTX 5070 12GB with 32GB DDR4-2400
A llama.cpp expert-streaming build runs the 177B Qwen 3.8 Flash Next MoE at roughly 11.5 tok/s on a 12GB RTX 5070 and 32GB DDR4-2400 by streaming experts from SSD.
Coding locally with Qwen 3.8 27B Q6 on an RTX 5090: 160k context, ~80 tok/s, Hermes harness
The author runs Qwen 3.8 27B at Q6 with a 160k context and vision on an RTX 5090, using the Hermes agent harness and SearxNG in Docker, and reports around 80 tok/s but about 15 minutes for a one-shot website.
Local coding and homelab on a first-gen AMD Framework 13 (Ryzen 7 7840U)
A Framework 13 with a Ryzen 7 7840U and 64GB of unified memory runs privacy-first coding and homelab work through llama.cpp and hermes, where the larger model is too slow to use interactively.
Qwen 3.8 27B at ~6 tok/s on an 8GB RTX 4060 via RAM offload and MTP
A gaming PC with an RTX 4060 8GB and 32GB DDR5 runs a 17.5GB Qwen 3.8 27B by splitting it between VRAM and system RAM and adding speculative decoding, then drives it as a remote coding agent; the same machine also runs Qwen3-VL for invoice extraction.
Qwen 3.8 27B at a true 128K context with vision on a single 16GB RTX 5080 (NInfer)
A purpose-built engine (NInfer v1.3) fits Qwen 3.8 27B at a real 131,072-token context and KV capacity, plus vision, on one RTX 5080 16GB, qualifying ~71.6 tok/s decode and ~1,381 tok/s prefill on a 118,001-token prompt.
Qwen 3.8 27B on a single RTX 4090, measured (llama.cpp)
A documented, measured local coding setup: Qwen 3.8 27B on one RTX 4090 with llama.cpp and OpenCode, holding a 128k context and reporting around 92 tokens/s generation with speculative decoding.
Ternary Bonsai 2 27B on an RTX 4090: speed at a quality price
The same project's measured run of a ternary-quantised 27B model: about 92 tokens/s generation with no speculative decoding, from 6.7 GB of weights, at a documented cost in agentic coding ability.
Building a full SaaS product on Qwen 3.8 27B with vLLM
Switching from llama.cpp to vLLM on AMD hardware was the unlock for running a complete SaaS build loop through a local model.
Building robotics simulators and interactive lessons on a Mac Studio
A teacher uses local models to build topic-specific simulators, interactive lectures, and fast examples for students, while sending larger changes to cloud models when more context is needed.
Qwen 3.8 27B on a MacBook Pro M4 Max 36GB: context scaling, TTFT and decode through Ollama
Context-scaling benchmark of Qwen 3.8 27B in NVFP4/MLX through Ollama on a 36 GB M4 Max for agentic coding, measuring TTFT, prefill and decode from 4K to 115K.
Qwen 3.8 Flash Next on a 2021 M1 Max: 44 tok/s and 35 tok/s at 398K context with Splash
Long-context decode benchmark of Qwen 3.8 Flash Next on an M1 Max 64 GB through an unofficial Splash/ds4 fork in real OpenCode agent sessions.
Full app development on a single 3090 with OpenCode
A four-bit Qwen 3.8 quant at 200K context handles complete app builds through OpenCode, left running until it hits a decision point.
Background coding at 10-15 tok/s on an Intel Arc laptop GPU
A NUC12 with a 16GB Intel Arc A770M handles a 262K-context Qwen build entirely in VRAM, coding in the background while the owner does something else.
Qwen for personal coding alongside Claude at work
A software engineer uses Qwen locally for personal projects and compares it directly with Claude used professionally.