Setup guide · version 2026-10-01

Terse-Coder 27B Q6_K · native MTP

A reproducible candidate and cautious setup path. This guide does not establish that it will succeed on every device.

Before you download

Inspect existing hardware, free RAM/VRAM, disk, models and runtime. Confirm the selected hardware and memory; never assume separate GPUs form one memory pool.
Check the model license, exact pinned GGUF filename, revision and checksum below. Check the report's historical-byte verification notes. Keep existing models and configs.
Use a current compatible llama.cpp build, or LM Studio with an engine supporting this architecture and template. Follow its official model-download guide; do not silently substitute another quant or latest revision.
Begin with a modest context and one session. Inspect real allocation before raising limits. Weights, cache, buffers and other apps all need memory. Recorded long-context/MTP settings are not guaranteed on other versions.
Keep any local endpoint bound to loopback. Consult your local agent’s provider documentation. Pi/Hermes need compatible provider and tool parsing; chat assistants may only be able to explain manual steps.
Run a bounded scratch task with measurable acceptance checks. Record quant/hash/runtime/template/context, elapsed time, memory, corrections and failures. Stop on memory pressure or repeated malformed tool calls. Installation and configuration changes follow your agent’s own approvals.

Settings, artifact identity & publisher claim

Medium reasoning; context 118016; q8_0 K/V; native draft-mtp max 4, backend draft sampling disabled; temperature 0.6, top-p 0.95, top-k 20, min-p 0.05, presence penalty 1.0; response cap 8192; parallel 1; no vision projector. Hermes 3aee290899e478c5fdfb6a241ef62758a49829b3. Author template Shockem/froggeric-terse-coder at c557b014b6675a87683b45d0800a68842311ffe2 (rendered anti-rumination/medium-effort use verified). Qualify MTP/tool parsing on the exact runner before relying on these settings.

Settings source · Apache-2.0

Qwen3.8-27b-Terse-Coder.Q6_K.gguf
Revision: 7602eadf6a90853492e4e8a377c2a2f531bf21e2
SHA-256: 3164a8d8cc7f5c669b7a6de57ce9e5db126d73b8d39a0f424a2b9c649f25df07

Actual local model bytes rehashed and size checked against pinned download: SHA256 3164a8d8cc7f5c669b7a6de57ce9e5db126d73b8d39a0f424a2b9c649f25df07. Author template separately pinned; runner path/alias and rendered template proof verified. This exact first-party run does not establish equivalence with other quants/runtimes.

Publisher/author claim: Shockem identifies this pinned artifact as Qwen3.8-27b-Terse-Coder Q6_K and licenses the repository Apache-2.0. No publisher speed/quality claim is used for ordering. Source

Claim is attributed, not independently confirmed; it does not determine ordering. No comparable vendor/community thinking-reduction percentage is available.

27B total parameters · dense. Full stored weights determine memory.

AI Efficiency Toolbox · 2026-10-02 · Personally tested lead; community coverage absent. One independently checked first-party FaithVoice parser run passed on Apple M5 Pro 48GiB. Exact bytes/settings are documented; no cross-variant quality transfer, isolated MTP speedup or release qualification established. Keep this separate from the historical 14-setup study and community observations.

2026-10-02: add verified latest native-MTP run; earlier no-MTP pass retained as context, not another independent community report.

Copy an agent setup handoff

Includes your selected hardware/task and the exact artifact. Your agent must check resources and its own permissions; without computer tools it can guide manual setup.

Help me evaluate this local AI setup. Guide version: 2026-10-01.
Guide: https://ai.efficiencytoolbox.com/local-llm-finder/setups/terse-coder-q6-mtp
Hardware: Unknown — ask me before downloading or changing anything; selected unified/system memory: unknown.
Execution mode: local-only. This guide covers a local candidate only; cloud/hybrid comparisons are unevaluated.
Work: Build or fix code. Current setup: not specified.
Exact candidate: Terse-Coder 27B Q6_K · native MTP, Q6_K GGUF.
Trusted artifact publisher: Shockem/Qwen3.8-27b-Terse-Coder-GGUF. Revision: 7602eadf6a90853492e4e8a377c2a2f531bf21e2.
Download: https://huggingface.co/Shockem/Qwen3.8-27b-Terse-Coder-GGUF/resolve/7602eadf6a90853492e4e8a377c2a2f531bf21e2/Qwen3.8-27b-Terse-Coder.Q6_K.gguf
SHA-256: 3164a8d8cc7f5c669b7a6de57ce9e5db126d73b8d39a0f424a2b9c649f25df07. Weight file: 22430999904 bytes. License: Apache-2.0.
Recorded runtime: llama.cpp 2.37.0. LM Studio is an optional GGUF path only if its selected engine supports this architecture/template; it does not reproduce every standalone runtime feature.

1. Inspect my OS, hardware, free RAM/VRAM, disk space, existing models, runtime and agent configuration first. Confirm unknown values and any mismatch with this guide. Do not silently pool separate GPU memories.
2. Honor your own tool/installation/network approvals. Do not replace or delete existing models/configs, expose credentials, open the server publicly, or spend money without my approval. If you lack computer tools, explain the manual steps instead of claiming installation.
3. Review model license and source. Download only the exact pinned artifact after confirming resources. Verify SHA-256 before loading. Total weights plus KV cache, context, buffers/projector and other apps must fit; active MoE parameters are not weight memory. Start with a modest context and one session, inspect actual allocation; this guide supplies no guaranteed context fit.
4. Use a compatible llama.cpp build, or LM Studio from https://lmstudio.ai/ with its official guide https://lmstudio.ai/docs/app/basics/download-model . Search the exact repository/filename; if its UI offers different bytes/revision, use the pinned download/import workflow documented by the installed version rather than substituting silently. Check architecture/template/tool-call support. Keep local endpoints bound to loopback.
5. Settings and limitations: Medium reasoning; context 118016; q8_0 K/V; native draft-mtp max 4, backend draft sampling disabled; temperature 0.6, top-p 0.95, top-k 20, min-p 0.05, presence penalty 1.0; response cap 8192; parallel 1; no vision projector. Hermes 3aee290899e478c5fdfb6a241ef62758a49829b3. Author template Shockem/froggeric-terse-coder at c557b014b6675a87683b45d0800a68842311ffe2 (rendered anti-rumination/medium-effort use verified). Qualify MTP/tool parsing on the exact runner before relying on these settings.
Settings source: https://ai.efficiencytoolbox.com/downloads/personal-benchmarks/terse-coder-q6-mtp-20261002.json
6. With Pi/Hermes or another tool-capable local-model harness, verify its model/provider configuration against its current documentation. Codex/Claude/ChatGPT/Gemini may guide setup but are not all local-model runtimes or computer-capable agents. Do not invent a provider URL or assume support. Run a bounded scratch task, validate tool arguments, outputs and tests, record runtime/version/quant/context and memory/time. Stop if memory pressure, loops or malformed tools recur.
7. Explain what worked, failed or needed correction. Keep publisher claims separate from community reports and this test. Actual local model bytes rehashed and size checked against pinned download: SHA256 3164a8d8cc7f5c669b7a6de57ce9e5db126d73b8d39a0f424a2b9c649f25df07. Author template separately pinned; runner path/alias and rendered template proof verified. This exact first-party run does not establish equivalence with other quants/runtimes.
Editorial limitation: One independently checked first-party FaithVoice parser run passed on Apple M5 Pro 48GiB. Exact bytes/settings are documented; no cross-variant quality transfer, isolated MTP speedup or release qualification established. Keep this separate from the historical 14-setup study and community observations.

Evidence and limits

One independently checked first-party FaithVoice parser run passed on Apple M5 Pro 48GiB. Exact bytes/settings are documented; no cross-variant quality transfer, isolated MTP speedup or release qualification established. Keep this separate from the historical 14-setup study and community observations.

Jason — Terse-Coder Q6_K native-MTP FaithVoice benchmark · checked 2026-10-02

Publisher claims and community reports are attributed evidence, not independent verification of your installation. No automatic install or device access is performed by this website.