DeployR02
Install Ollama
A source-based path from a versioned installer to a local API response, with a separate check for where the model actually runs.
Updated 17 Oct 2026 Read guide →
TOOLS & GUIDES
A practical next step.
Task guides and browser-local tools. No server or model action is triggered here.
DeployR03
Build llama.cpp
Keep CPU, Metal and CUDA builds separate, pin the source, and inspect placement before making an acceleration claim.
Updated 17 Oct 2026 Read guide →
DeployR06
Run Ollama locally
Separate a local endpoint, disabled cloud features and a genuinely offline workflow. Verify each boundary.
Updated 17 Oct 2026 Read guide →
HardwareR07
Estimate model memory
An explicit, editable memory budget for conventional transformer KV caches. Unknown inputs stay unknown.
Updated 17 Oct 2026 Read guide →
TroubleshootingR11
Diagnose Ollama 500
A symptom-to-check decision tool that preserves the error, changes one condition, and asks for a regression check.
Updated 17 Oct 2026 Read guide →
HardwareR13
DGX Spark preflight
A credential-free checklist for architecture, artifacts, capacity and recovery before an experiment is scheduled.
Updated 17 Oct 2026 Read guide →
ModelsR14
Choose local model sources
A source and license workflow for original weights, community conversions and runtime-specific model identities.
Updated 17 Oct 2026 Read guide →
MethodologyR15
Our collection method
Original links, pinned source identity, dated reading notes and honest comparison boundaries.
Updated 17 Oct 2026 Read guide →
EVIDENCE EXPLAINED / 18 OCT 2026
Read a configuration. Make a clearer decision.
Original source facts, practical interpretation and explicit missing evidence.
EVIDENCE EXPLAINEDA22
MiMo Pro on Eight Sparks: Scheduling, Rates and Regression Scores
Read a13/a14 cold extraction logs, the queued eighth-request fix, full model/base-engine pins, a GSM8K subset and older-build needle checks.
Updated 18 Oct 2026 Read the evidence →
EVIDENCE EXPLAINEDA21
GLM Flash Scores: Historical NVFP4 and the EXL3 Retraction
Separate a full historical MMLU run and small task subsets from the corrected EXL3 router stack, whose old task scores were withdrawn and await reruns.
Updated 18 Oct 2026 Read the evidence →
EVIDENCE EXPLAINEDA20
Full GLM-5.3 on Four Sparks: Cache Layout, Speed and Fidelity
Audit the 753B non-Flash EXL3 TP4 recipe, replicated and shared Q5 cache, million-token retrieval, BF16-reference fidelity and later RDMA fixes.
Updated 18 Oct 2026 Read the evidence →
EVIDENCE EXPLAINEDA19
Qwen Long Concurrency: Paired Memory Peaks and OOM Recovery
Read two 250K-request memory controls, a 500K single-request peak, a production OOM and invalidated/repeated recovery tests without mixing hardware or cohorts.
Updated 18 Oct 2026 Read the evidence →
EVIDENCE EXPLAINEDA18
Kimi IQ1_S on Three Sparks: Scores, Retries and Missing Pins
Audit a publisher-labelled K3 layer-split deployment, approximate memory and small GSM8K/HumanEval subsets, retaining first-pass failures and timeout retries.
Updated 18 Oct 2026 Read the evidence →
EVIDENCE EXPLAINEDA17
MiniMax-M3 on Two Sparks: The Measured W4A16 Lane
Separate a July lane-A reproduction from proposed speed/context recipes, retaining request timing, startup memory brackets and a negative NCCL result.
Updated 18 Oct 2026 Read the evidence →
EVIDENCE EXPLAINEDA16
GLM-5.3-Flash on Two Sparks: Speed Meets Quality Gates
A pinned NVFP4 target, DFlash2 draft and community image reach sustained decode and 120124-token retrieval, while strict output failures remain visible.
Updated 18 Oct 2026 Read the evidence →
EVIDENCE EXPLAINEDA15
Qwen Long Context: Concurrency, Memory and Failure Evidence
Read 128K and 490K Qwen batches, a four-GPU 128K matrix, sampled peak VRAM and a Mamba crash with original versions and measurement limits.
Updated 18 Oct 2026 Read the evidence →
EVIDENCE EXPLAINEDA14
Qwen on RTX 5090: What the 128K Recipe Tested
Understand a 104805-token Qwen retrieval, 32K-only memory telemetry and rejected MTP promotion on an RTX 5090 with WSL2 and partial CPU expert offload.
Updated 18 Oct 2026 Read the evidence →
EVIDENCE EXPLAINEDA13
Qwen Flash-Next: Concurrency Has a Latency Cost
Read Qwen Spark concurrency, per-request delay and cold-context evidence, keeping checkpoints, releases, cache settings and memory failures distinct.
Updated 18 Oct 2026 Read the evidence →
EVIDENCE EXPLAINEDA12
DeepSeek 1017: Quantization, Memory and Context Evidence
Separate a short EXL3 deployment matrix, two near-million-token needle checks and a VQ conversion that has not measured 1017 serving speed.
Updated 18 Oct 2026 Read the evidence →
EVIDENCE EXPLAINEDA11
Qwen Flash-Next: Quantization, Memory and Long Context
Read a single-Spark cold-context study and a separate PLE-memory experiment, preserving hardware, builds, memory types and unmeasured limits.
Updated 18 Oct 2026 Read the evidence →
EVIDENCE EXPLAINEDA10
Qwen3.8-27B Chisel: Deployment, Speed and the Quality Gate
Read the local serving recipe, workload-specific timing and paired BF16 quality comparisons, including the limits hidden by a simple pass flag.
Updated 18 Oct 2026 Read the evidence →
EVIDENCE EXPLAINEDA09
DeepSeek 1017 on Two Sparks: Read the TensorFold Evidence
Read the exact 1017 artifact, two historical engine configurations and a separate final prefill cohort, with original results and missing conditions.
Updated 18 Oct 2026 Read the evidence →
EVIDENCE EXPLAINEDA08
GB10 Long Context: Read the Failure Before Changing Flags
Read four attributed GB10 incidents with exact model revisions, historical build conditions, evidence states, proposed changes and verified follow-ups.
Updated 18 Oct 2026 Read the evidence →
EVIDENCE EXPLAINEDA07
Ollama, llama.cpp, vLLM or SGLang: Choose a Serving Path
Choose a runtime by artifact, backend, API and workload, with four pinned maintainer references and explicit checks before deployment.
Updated 18 Oct 2026 Read the evidence →
EVIDENCE EXPLAINEDA06
KV Cache and Speculative Decoding: The Memory Budget
Read a same-model two-Spark comparison with allocated KV pools, twelve original campaign cells and explicitly derived rate summaries.
Updated 18 Oct 2026 Read the evidence →
EVIDENCE EXPLAINEDA05
Two DGX Sparks: Read the Deployment Conditions
Read three community launch profiles with artifact revisions, KV pools, speculation, context limits, actual concurrency and transport checks.
Updated 18 Oct 2026 Read the evidence →
EVIDENCE EXPLAINEDA04
Mac Weight Precision: Speed, First Token and Memory
Compare Qwen3.5 4B and 9B weight formats on one M4 Pro, with separate workloads for generation, first-token delay and peak memory.
Updated 18 Oct 2026 Read the evidence →
EVIDENCE EXPLAINEDA03
Qwen3.8 Flash-Next: Resolve the Deployment Recipe
Choose the exact conversion, preserve the runtime qualification and resolve hardware, context, cache and optional features together.
Updated 18 Oct 2026 Read the evidence →
EVIDENCE EXPLAINEDA02
DeepSeek V4.1 Flash 1017: Deployment Evidence
Identify the exact 1017 checkpoint, resolve the vLLM recipe and keep publisher recommendations separate from measured capacity.
Updated 18 Oct 2026 Read the evidence →
EVIDENCE EXPLAINEDA01
LLM Speed: TTFT, Decode and Served Rate
Understand the clock behind a speed claim with three attributed community examples, exact workloads and separate timing definitions.
Updated 18 Oct 2026 Read the evidence →