# Unofficial Intel XPU Optimization Lab (neural.download)

> Community lab notebook for fast, reproducible local LLM inference on Intel Arc Pro B-series GPUs (B50/B60/B65/B70). Exact software pins, kernel patches, speculative-decoding lanes, and benchmark evidence. Every performance claim carries an evidence label: B70-verified (reproduced in the reference lab), B70-tested (patch validated on B70), or community-reported (contributor hardware). Source of truth for all claims is the GitHub repository, not this summary.

The human landing page is https://neural.download/ (index.html), and the searchable deployment library is https://neural.download/guides.html. Markdown, JSON, and text assets are also served directly and are safe to fetch programmatically. Model weights, secrets, and raw benchmark trees are deliberately NOT published.

## Start Here (agents)

- [AGENTS.md](https://neural.download/AGENTS.md): canonical agent instructions — reading order, folder contracts, what must not change
- [CURRENT.md](https://neural.download/CURRENT.md): sole authority for live service state, active lane, protected work, next actions
- [README.md](https://neural.download/README.md): project overview, audience map, hardware scope
- [Docs index](https://neural.download/docs/README.md): navigation layer for all guides

## Core Guides

- [Model optimization guide](https://neural.download/docs/model-optimization-guide.md): start-to-finish playbook for optimizing a new model lane
- [Research workflow playbook](https://neural.download/docs/research-workflow-playbook.md): prompts, validation ladders, experiment discipline
- [Current reproducibility map](https://neural.download/docs/current-reproducibility-map.md): stable promoted reproduction catalog and host/service map
- [Model effort index](https://neural.download/docs/model-effort-index.md): cross-model status, closed lanes, where new packets go
- [Model recipes](https://neural.download/docs/model-recipes.md): which recipe folder per model/build target
- [Model intake queue](https://neural.download/model-intake/README.md): revision-pinned candidate downloads, external-store checks, and already-covered families
- [Model intake catalog](https://neural.download/model-intake/catalog.json): machine-readable revisions, artifact sizes, checksums, fit, and status
- [Distribution and packaging roadmap](https://neural.download/docs/model-distribution-and-packaging-roadmap.md): novice one/two-GPU, Docker, Windows, and contributor-recognition path
- [Reproduction certification](https://neural.download/docs/reproduction-guide-certification.md): authoritative starter/candidate/lab/record classifications and clean-host promotion gates
- [Reproduction catalog](https://neural.download/repro/guide-catalog.json): machine-readable classification and missing dependencies for every repro README
- [Model packages](https://neural.download/packages/README.md): user-facing package layer; statuses are explicit and no current package is starter-certified
- [Searchable guide library](https://neural.download/guides.html): human-facing package browser with model, quant, card-count, OS, delivery, and status filters
- [Package catalog](https://neural.download/packages/catalog.json): generated machine-readable discovery and exact-contributor metadata for every user-facing package
- [LocalMaxxing submissions](https://neural.download/docs/localmaxxing.md): competition submission records and audit trail

## Results and Evidence

- [Performance index](https://neural.download/results/scoreboard.md): promoted benchmark rows with metric definitions and validity rules — read its ground rules before comparing any two numbers
- [Claims registry](https://neural.download/claims/README.md): accepted claims with exact recipe identity, evidence, submitter credit, lifecycle status, and dated history
- [Results index](https://neural.download/results/README.md): promoted model packets and promotion criteria
- [Optimization timeline](https://neural.download/optimization-timeline.html): HTML table of how each lane got fast
- [Qwen3.8 FP8 two-B70 candidate package](https://neural.download/packages/qwen38-27b-fp8-tp2-b70/README.md): strict varied-prompt single-user headline pending after the 128-cap/selected-fixture audit; digest-pinned container, exact patches, direct verification, and separately scoped target-only aggregate/context evidence
- [Qwen3.8 Q4_K_M one-B70 candidate package](https://neural.download/packages/qwen38-27b-q4km-tp1-b70/README.md): our exact patched llama.cpp/SYCL stack, direct-verified model, and exact-output validation gate
- [Qwen3.8 Q4_K_M two-B70 candidate package](https://neural.download/packages/qwen38-27b-q4km-tp2-asrock-b70/README.md): exact 49.717503 tok/s lab replay, TTFT, direct model verification, and explicit remaining clean-host gates
- [Qwen3.8 Q8_0 two-B70 candidate package](https://neural.download/packages/qwen38-27b-q8-tp2-b70/README.md): quality-conservative target-only TP2 service, scoped contributor credit, and exact project patch chain
- [Muse-Glimmer-30B four-B70 candidate package](https://neural.download/packages/muse-glimmer-30b-q8-woq-b70/README.md): oneDNN WOQ, DFlash, distributed argmax, model hashes, and target-verified gates
- [Laguna-S-2.1 four-B70 record replay package](https://neural.download/packages/laguna-s-2.1-int4-b70-125tps/README.md): artifact-exact originating-host replay; explicitly not a portable install
- [MiniMax-M2.7 four-B70 candidate package](https://neural.download/packages/minimax-m27-int4-autoround-b70/README.md): expert INT4 build, quality, and benchmark route with its historical model-provenance limit stated

## Contributing

- [CONTRIBUTING.md](https://neural.download/CONTRIBUTING.md): required-evidence checklist for any result or patch
- [Contribution verification](https://neural.download/docs/contribution-verification.md): the eight-step reference-lab validation and evidence labels
- [community/README.md](https://neural.download/community/README.md): where runnable community work lands
- [STATUS template](https://neural.download/community/STATUS-TEMPLATE.md): required status file format for community entries
- [Field reports](https://neural.download/community/field-reports/README.md): informational measurements without runnable assets

## External

- [GitHub repository (source of truth)](https://github.com/steveseguin/b70-optimization-lab)
- [Legal and risk](https://neural.download/DISCLAIMER.md): unofficial, unaffiliated with Intel; experimental research

## Optional

- [LICENSE](https://neural.download/LICENSE)
- [Agent handoff log](https://neural.download/AGENT_HANDOFF.md): chronological cross-agent handoffs

## Learn guides (plain-language explainers with labeled lab, community, and projected evidence)

- [Learn index](https://neural.download/learn.html)
- [Models](https://neural.download/learn/models.html) · [Quantization and speed](https://neural.download/learn/quantization-and-speed.html) · [KV cache precision](https://neural.download/learn/kv-cache-precision.html) · [Context length](https://neural.download/learn/context-length.html) · [Prefill vs clock rate](https://neural.download/learn/prefill-vs-clockrate.html) · [MTP ladder](https://neural.download/learn/mtp-ladder.html) · [Speculative methods](https://neural.download/learn/speculative-methods.html) · [Runtimes](https://neural.download/learn/runtimes.html) · [Hardware](https://neural.download/learn/hardware.html) · [B70 ECC and usable VRAM](https://neural.download/learn/b70-ecc-and-vram.html) · [Host tuning](https://neural.download/learn/host-tuning.html)

## Model family coverage and deployment pages

- [Models index](https://neural.download/models/): family-first coverage plus individual deployment packets; measured speed, proof, exact pins/patches, and labeled projections stay distinct.
- [Gemma 4 family coverage](https://neural.download/models/gemma-4.html) · [manifest](https://neural.download/families/gemma-4.json)
- [Laguna S family coverage](https://neural.download/models/laguna-s.html) · [manifest](https://neural.download/families/laguna-s.json)
- [LFM2.5 family coverage](https://neural.download/models/lfm-2-5.html) · [manifest](https://neural.download/families/lfm-2-5.json)
- [MiniMax M2.7 family coverage](https://neural.download/models/minimax-m2-7.html) · [manifest](https://neural.download/families/minimax-m2-7.json)
- [Muse-Glimmer family coverage](https://neural.download/models/muse-glimmer.html) · [manifest](https://neural.download/families/muse-glimmer.json)
- [Nemotron 3.5 family coverage](https://neural.download/models/nemotron-3-5.html) · [manifest](https://neural.download/families/nemotron-3-5.json)
- [Ornith 1.5 family coverage](https://neural.download/models/ornith-1-5.html) · [manifest](https://neural.download/families/ornith-1-5.json)
- [Qwen 27B family coverage](https://neural.download/models/qwen-27b.html): Qwen3.6 and Qwen3.8 lineage, TP/MTP decisions, context curves, decode/prefill/TTFT, recipes, closures, and explicit gaps.
- [Qwen 27B family manifest](https://neural.download/families/qwen-27b.json): machine-readable revisions, variants, measured slices, coverage states, and evidence links.
- [Gemma 4 26B A4B UD-Q8_K_XL on one Intel Arc Pro B70](https://neural.download/models/gemma4-26b-a4b-q8-b70-125tps-20260701.html)
- [Laguna S 2.1 INT4 with DFlash on four Intel Arc Pro B70 cards](https://neural.download/models/laguna-s-2.1-int4-b70-125tps-20260731.html)
- [LFM2.5 2.6B Q8_0 on one Intel Arc Pro B70](https://neural.download/models/lfm25-26b-q8-b70.html)
- [MiniMax M2.7 AutoRound INT4 on four Intel Arc Pro B70 cards](https://neural.download/models/minimax-m27-b70-89tps-20260520.html)
- [Muse-Glimmer 30B Q8/WOQ with DFlash on four Intel Arc Pro B70 cards](https://neural.download/models/muse-glimmer-30b-q8-woq-b70-100tps-20260813.html)
- [Nemotron 3.5 Lightning 30B-A3B UD-Q4_K_M on one Intel Arc Pro B70](https://neural.download/models/nemotron-35-lightning-30b-a3b-b70.html)
- [Ornith 1.5 35B-A3B Q4_K_M on one Intel Arc Pro B70](https://neural.download/models/ornith-15-35b-a3b-q4km-b70.html)
- [Ornith 1.5 9B Q8_0 on one Intel Arc Pro B70](https://neural.download/models/ornith-15-9b-q8-b70.html)
- [Qwen3.8 27B 256K + vision + MTP draft on one Intel Arc Pro B70](https://neural.download/models/qwen38-27b-256k-vision-mtp-b70.html)
- [Qwen3.8 27B official FP8 on two Intel Arc Pro B70 cards](https://neural.download/models/qwen38-27b-fp8-vllm-tp2-asrock-b70.html)
- [Qwen3.8 27B Q4_K_M on one Intel Arc Pro B70](https://neural.download/models/qwen38-27b-q4km-tp1-b70.html)
- [Qwen3.8 27B Q4_K_M on two Intel Arc Pro B70 cards](https://neural.download/models/qwen38-27b-q4km-tp2-asrock-b70.html)
- [Qwen3.8 27B Q8_0 target-only on two Intel Arc Pro B70 cards](https://neural.download/models/qwen38-27b-q8-tp2-asrock-b70.html)

## Projections vs measurements

The landing page's "How much faster could these get?" section and the Hardware guide's projected comparison use the ML Bottleneck physics engine (https://mlbottleneck.com/, same author; SDK at https://mlbottleneck.com/dist/mlbottleneck-engine.umd.js). Those numbers are model projections labeled as such; they are never lab measurements and must not be cited as neural.download benchmarks.

## Related projects by the same author

- https://mlbottleneck.com/ — physics-based LLM speed and memory planner for any GPU, calibrated on community benchmarks (deep links: https://mlbottleneck.com/?model=qwen3.8_27b&hardware=Intel%20Arc%20Pro%20B70&count=2&format=Q4_K_M&runtime=vllm).
- https://style-genome.com/ — design-system generator (this site's visual style is a Style Genome output).
- https://neural.download/learn/multi-user.html : measured multi-user (1..64 simultaneous) aggregate throughput on one Arc Pro B70