English | 简体中文

First LLM Studio is a local-first LLM workbench for Apple Silicon. It brings local MLX runtimes, remote API targets, Agent sessions, Compare, Fine-tune, Benchmark, model discovery, runtime recovery, release evidence, and admin monitoring into one operating surface.

It is not another chat shell. It is built for people who need to compare behavior, debug runtimes, run evals, prepare adapters, and keep local and remote model work inside one product loop.

Current source version: v1.5.1, with source status COMPLETE and local acceptance PASS. The machine-readable release-state.json deliberately separates the active source milestone from the latest public GitHub release (v0.4.0) and the latest desktop candidate (v1.1.0-rc.2).

Previous operational-lifecycle source gate: v2.4.0-v2.5.4. Local projection contains 8 passing, 5 attention, and 2 external-only signals; independently verified external records remain 0/15, so this evidence does not change distribution or production status.

Previous governed-autonomy and interoperability plan: v2.6.0-v2.7.4. Its 15 source slices read existing feature-owned evidence and preserve independent external review, distribution HOLD, and production BLOCKED as separate facts.

Previous governed-autonomy source gate: v2.6.0-v2.7.4 records 9 passing, 4 attention, 0 unavailable, and 2 external-only signals. Independently verified external records remain 0/15.

Previous operational remediation plan: v2.8.0-v2.9.4. Its initial owner-controlled projection records 6 passing, 7 attention, 0 unavailable, and 2 external-only signals. The attention queue is explicit product work; independently verified external records remain 0/15, distribution remains HOLD, and production remains BLOCKED.

Previous validated source gate: v2.8.0-v2.9.4 records 119/119 tests, 73/73 CI route checks, full cross-surface smoke, desktop/mobile browser QA, and the machine-readable export while preserving the same ATTENTION/HOLD/BLOCKED truth.

Previous remediation-control and service-readiness plan: v3.0.0-v3.1.4. The initial projection reports 5 passing, 8 attention, 0 unavailable, and 2 external-only signals. Its control plane classifies 2 items as satisfied, 3 open, 8 blocked, and 2 external-only; independently verified records remain 0/15, distribution remains HOLD, and production remains BLOCKED.

Previous validated source gate: v3.0.0-v3.1.4 records 125/125 tests, 75/75 CI routes, full smoke, production build, security preflight, and desktop/mobile browser QA while preserving ATTENTION/HOLD/BLOCKED truth.

Latest remediation-execution and operational-acceptance plan: v3.2.0-v3.3.4. Seven direct owner actions now carry deterministic idempotency keys, bounded leases, fencing, rollback, and evidence fingerprints; the current projection is 0 satisfied, 3 ready, and 4 blocked. Independently verified records remain 0/15, distribution remains HOLD, and production remains BLOCKED.

Latest validated source gate: v3.2.0-v3.3.4 records 131/131 tests, 77/77 CI routes, full smoke, production build, security preflight, and desktop/mobile browser QA while preserving ATTENTION/HOLD/BLOCKED truth.

Current owner-workload and operational-decision plan: v3.4.0-v3.5.4. Seven owner workloads now have strict request and candidate receipt contracts; the source projection is 5 pass, 8 attention, and 2 external-only. External evidence remains 0/15, distribution remains HOLD, and production remains BLOCKED.

Current validated source gate: v3.4.0-v3.5.4 records 137/137 tests, all 11 changed TypeScript partitions, 79/79 CI routes, full smoke, production build, security preflight, and desktop/mobile browser QA while preserving ATTENTION/HOLD/BLOCKED truth.

Current receipt and exception lifecycle plan: v3.6.0-v3.7.4. The repository now owns authenticated append-only receipt events, quarantine, compensation, escalation acknowledgement, bounded waiver expiry, and deterministic decision packages. The source projection is 6 pass, 7 attention, and 2 external-only; no real owner receipt was submitted during source validation, external evidence remains 0/15, distribution remains HOLD, and production remains BLOCKED.

Previous validated source gate: v3.6.0-v3.7.4 records 146/146 tests, all 11 changed TypeScript partitions, 81/81 CI routes, full smoke, production build, release-truth validation, architecture/durable-state checks, and security preflight.

Current competitive product plan: v3.8.0-v3.9.4 implements all 15 repository-owned contracts and exposes their source/local/external/production truth in /experiments. Real provider accounts, workload scorecards, runtime/model transfers, OS sandbox receipts, licensed multimodal assets, managed RAG, non-MLX training workers, multi-user identity and independent acceptance remain explicit evidence gates; distribution remains HOLD and production remains BLOCKED.

Current validated source gate: v3.8.0-v3.9.4 records 156/156 tests, all 11 changed TypeScript partitions, 85/85 CI routes, full smoke, production build, architecture/durable-state checks, and security preflight. Source is 15/15 pass; local evidence is 12/15 pass with routing-quality, licensed multimodal, and real Hub-transfer evidence deliberately held.

Current evidence-operations plan: v4.0.0-v4.1.4 adds a strict, durable candidate receipt contract and 15 visible source/local/external/production gates in /experiments. Candidate receipts can close only local evidence; independent signatures, immutable archive read-back, distribution approval, and production authority remain HOLD or BLOCKED.

Latest validated source gate: v4.0.0-v4.1.4 records 164/164 tests, all 11 changed TypeScript partitions, 86/86 CI routes, full smoke, a 53-page production build, architecture/durable-state checks, and an archived repaired zero-vulnerability security preflight. The latest online-audit rerun failed closed on an npm registry timeout, not a detected vulnerability. Source is 15/15 pass; observed local evidence is 3/15 pass, candidate receipts are 0, external verification is 0/15, distribution remains HOLD, and production remains BLOCKED.

Current evidence-collection plan: v4.2.0-v4.3.4. All 15 source mechanisms are implemented: queue controls, Provider catalog execution, 20-sample routing/shadow replay, review/export, and eight feature-owned domain collectors. The domain collectors reject reference, fixture, and derived read models before receipt submission; four deterministic control gates pass locally, while real workload evidence remains needed. Distinct reviewer IDs are not verified independent principals. The validation record separates isolated transport tests from real workload evidence. Distribution remains HOLD; production remains BLOCKED.

Current evidence-adjudication plan: v4.4.0-v4.5.4. The source-gate record records all 15 implemented contracts and seven deterministic local controls. No accepted workload package, externally verified reviewer, immutable archive read-back, or independent re-observation is present, so the external and production states remain HOLD / BLOCKED.

Current evidence-acceptance plan: v4.6.0-v4.7.4. The source-gate record records all 15 source contracts, eight deterministic lifecycle controls, and the production-build route-smoke receipt. No controlled authority, immutable archive read-back, current externally verified package, independent observer, or distribution approval is present, so external and production states remain HOLD / BLOCKED.

Current evidence-authority campaign plan: v4.8.0-v4.9.4. The source-gate record records all 15 source contracts, nine deterministic controls, and production-build route-smoke evidence. No controlled credential resolver, live authority profile, campaign, externally attested receipt, immutable archive read-back, independent observer, or distribution approval is configured, so external and production states remain HOLD / BLOCKED.

Internal post-release evidence has advanced through the v1.6.6 benchmark decision-intelligence checkpoint: all 500 stored results are classified (160 correct and 340 mathematical mismatches, with 100% answer extraction), confidence-aware risk highlights Intermediate Algebra, Precalculus, and Level 5, and the conservative 500-sample detectable effect is 8.27 percentage points. Only one distinct complete run exists, so candidate promotion remains EVIDENCE NEEDED. This does not change the active source version or claim hosted leaderboard parity, independent-host reproduction, multimodal execution, distribution, or production promotion.

Latest desktop candidate notes: v1.1.0-rc.2. Distribution remains HOLD and production remains BLOCKED; this repository does not represent missing Apple, organization, non-loopback, distributed-worker, or collaborative receipts as completed evidence.

Reviewed against first-party product and model documentation on 2026-08-31. Core means a first-party primary workflow; integrated means the capability exists but is not the product's deepest specialization; ecosystem means it is normally assembled through adjacent clients or plugins. Vendor benchmark claims are not treated as First LLM Studio evidence.

Current official model/Agent watchlist: OpenAI gpt-5.6-sol; Anthropic claude-fable-5 / claude-opus-5 and Managed Agents; Google gemini-3.7-flash and Antigravity; DeepSeek deepseek-v4-pro / deepseek-v4-flash; MiniMax MiniMax-M2.7; Kimi kimi-k2.6; Zhipu's glm-5.2 quickstart; and Qwen3-Coder plus Qwen Code Agent Teams. This is an integration watchlist, not a claim that every model is configured or paid for locally.

First LLM Studio's advantage is the connected evidence chain across Agent, Compare, Retrieval, Benchmark, Fine-tune, adapter export, Workflow Studio, and release review. The 15 v3.8.0-v3.9.4 source contracts now cover capability discovery, explainable routing, Agent conformance/cache, isolated multi-agent plans, sandbox policy, multimodal admission, Model Hub/Local Server v3, governed RAG, training backend federation, Team Studio policy, and a <=14-day competitive freshness gate. Missing real workload and external receipts remain visible as HOLD. See the full bilingual competitive landscape and product direction.

• Compare MLX local models against hosted APIs under aligned context budgets.
• Inspect runtime cost, prewarm, release, recovery, and hardware pressure without leaving the app.
• Decide which local model is actually usable for daily coding and analysis workflows.

• Validate tool-calling, repo-grounded behavior, replay, and patch flows in one workbench.
• Turn Compare runs into benchmark handoff without switching products.
• Separate model-quality failures from provider quirks and local runtime instability.

• Run formal and focused benchmark suites with repeatable profiles.
• Review baselines, deltas, run notes, failure classifications, and release evidence.
• Keep local and remote targets inside one comparable target catalog.

• Unified local + remote target catalog.
• Compare Lab for model-vs-model output review.
• Fine-tune workflows for datasets, recipes, training, evaluation, adapter proof loops, export, best-checkpoint selection, and LoRA release evidence.
• Visual Workflow Studio for typed Agent/RAG/eval graphs, immutable recipes, guarded tool execution, replay, and OpenAI-compatible deployment.
• Benchmark operations with history, progress, baselines, reports, and release evidence.
• Foreground Retrieval for document import, chunk inspection, and grounded evidence probes.
• Experiments timeline for session/run lineage, artifact navigation, and retention policy.
• Replay, trace review, patch inspection, and exportable review notes.
• Runtime operations for prewarm, release, restart, log inspection, telemetry, and recovery.
• Dynamic local/community model discovery plus remote provider health scanning.

• Local Qwen3 0.6B
• Local Qwen3 4B 4-bit
• Local Qwen3.5 4B 4-bit
• Local Gemma 3 4B It Qat 4-bit

• OpenAI Codex
• OpenAI GPT-5.5
• Claude API
• DeepSeek API
• Kimi API
• GLM API
• Qwen API

Selection and stability guidance: docs/benchmark-lane-comparison.md.

Contributor onboarding: English · 中文快速上手 · GitHub setup checklist.

The 2026-08-30 refresh adds fresh, high-resolution evidence from the running application:

• Runtime Fabric: real MLX, Ollama, and llama.cpp passed 3/3 backends, 6/6 adapter contracts, and 42/42 normalized operations.
• Local Server: real Ollama 0.31.1 + Qwen3 0.6B passed 15/15 slices, 6 requests, 95 tokens, and 169 ms average latency.
• Benchmark smoke: Local Qwen3 0.6B completed 3/3 runs at 192.67 ms first token, 504.67 ms total latency, and 234.02 tokens/s.
• MATH-500: the complete 500-item run scored 500/500 outputs and replayed 500/500 evaluator decisions with 32% local accuracy.
• Fine-tune: the archived real Qwen3 4B LoRA run preserves 816 steps, save/eval markers, and the selected best checkpoint at step 800.

Full metrics, digests, capture dimensions, and evidence boundaries: v3.1.4-high-resolution-live-machine-capture-2026-08-30.md. Local source acceptance passes where stated; production promotion remains HOLD or BLOCKED where independent external evidence is still absent.

Captured from the running local app after type, test, build, route, and screenshot validation. The 1920x1200 capture viewport uses 2x DPR for 3840x2400 route images; evidence panels retain their native high-resolution crop, and the LoRA chart is exported from SVG at 2x DPR.

Agent workbench with target catalog, runtime rail, and tool-enabled composer:

Workflow Graph Studio with draggable typed nodes, version/revision controls, execution recovery, and promotion evidence:

Reproducible motion capture workflow: docs/demo-video-workflow.md.

Watch the Agent workbench MP4 demo · SHA-256 metadata

Fine-tune Studio with workflow tabs, training controls, and report/evidence panels:

Fine-tune completed run with live loss curves, train/validation traces, and handoff actions:

Real Qwen3 4B LoRA release evidence with save/eval markers and selected best checkpoint:

Vector version: fine-tune-qwen4b-lora-chart.svg. Full run archive and manifest: docs/release-evidence/finetune-qwen4b-lora-2026-07-01.

Benchmark Studio with run controls and historical evidence cards:

Benchmark run evidence from the fresh real local smoke run:

Complete MATH-500 result with subject/difficulty breakdown, Wilson confidence interval, evaluator replay, and latency/token performance:

Models Studio with immutable Hub/storage receipts, real Ollama Local Server acceptance, and the real MLX/Ollama/llama.cpp Runtime Fabric matrix:

Fresh local Runtime Fabric and Local Server acceptance panels:

MCP and secure extension acceptance with signed lifecycle, real tool discovery, quarantine defenses, and explicit production gates:

Compare, Retrieval, and Admin surfaces:

Operational remediation and service-readiness controls, with unresolved external gates kept visible:

• macOS on Apple Silicon
• Node 22.x
• Python 3.12
• MLX-compatible local environment

nvm install 22
nvm use 22
npm install
cp .env.example .env.local

npm run dev

Default routes:

• http://localhost:3011/agent
• http://localhost:3011/compare
• http://localhost:3011/fine-tune
• http://localhost:3011/models
• http://localhost:3011/benchmarks
• http://localhost:3011/retrieval
• http://localhost:3011/experiments
• http://localhost:3011/admin

python3 -m venv .venv
source .venv/bin/activate
pip install -U pip
pip install mlx mlx-lm
python scripts/local_model_gateway_supervisor.py

If your preferred Python is outside PATH, set LOCAL_AGENT_PYTHON_BIN before starting the app or gateway.

Gateway health:

• http://127.0.0.1:4000/health

npm run typecheck:changed
npm run smoke:routes
npm run smoke:screenshots

Copy .env.example to .env.local and fill only the providers you want to use.

Important notes:

• .env.local is ignored by git.
• Remote providers are optional.
• Several targets use OpenAI-compatible or Claude-compatible endpoints.
• Public defaults in this repository are sanitized placeholders.

app/                      Next.js app routes and thin API transports
components/               Shared UI and compatibility shells
features/                 Feature-owned routes, contracts, state, actions, and application ports
lib/agent/                Agent runtime, providers, benchmark, gateway helpers
lib/finetune/             Fine-tune store facade and split operation services
scripts/                  Local gateway, runtime, verification, and release scripts
docs/                     Architecture, release notes, launch notes, roadmap, and assets
modelscope/               ModelScope profile/readme metadata
public/                   Public assets and social cover art

• GitHub: https://github.com/ChrisChen667788/Your-First-LLM-Studio
• ModelScope profile: https://www.modelscope.cn/profile/haozi667788
• Default ModelScope repo id: haozi667788/first-llm-studio

The ModelScope package script exports the committed Git tree so GitHub and ModelScope can stay file-identical for each synced version.

• Sensitive local actions require confirmation.
• Secrets belong in .env.local.
• Public repository defaults are sanitized.
• New public commits should use a GitHub noreply address where possible.
• See SECURITY.md.

Issues and PRs are welcome.

• CONTRIBUTING.md
• CODE_OF_CONDUCT.md
• docs/open-source-backlog.md

• Current version: VERSION
• Release notes: docs/releases
• Release process: docs/release-process.md
• Active source: v1.5.1; latest public GitHub release: v0.4.0; latest desktop candidate: v1.1.0-rc.2

First LLM Studio 是一个面向 Apple Silicon 的本地优先 LLM 工作台。它把本地 MLX 运行时、远端 API 目标、Agent 会话、Compare 对比、Fine-tune 微调、Benchmark 评测、模型发现、runtime 恢复、发布证据和后台监控统一到一个产品界面里。

它不是另一个聊天壳，而是给真正需要比较模型行为、调试 runtime、跑评测、准备 adapter，并把本地/远端模型工作流收在同一个产品循环里的开发者使用。

当前源码版本：v1.5.1，源码状态为 COMPLETE、本地验收为 PASS。机器可读的 release-state.json 明确区分当前源码里程碑、最近公开 GitHub Release（v0.4.0）与最近桌面候选包（v1.1.0-rc.2）。

上一阶段受治理自治与互操作计划：v2.6.0-v2.7.4。这 15 个源码切片只读已有 feature-owned 证据；独立生态签收继续为 HOLD，生产状态继续为 BLOCKED。

上一阶段受治理自治 source gate：v2.6.0-v2.7.4 记录 9 个通过、4 个需关注、0 个不可用和 2 个仅外部可满足的信号；独立外部签收仍为 0/15。

上一阶段运行整改计划：v2.8.0-v2.9.4。首轮 owner-controlled 投影记录 6 个通过、7 个需关注、0 个不可用和 2 个仅外部可满足的信号；attention 队列属于明确的产品整改项，独立外部签收仍为 0/15，分发保持 HOLD，生产保持 BLOCKED。

上一阶段已验证 source gate：v2.8.0-v2.9.4 记录 119/119 测试、73/73 CI 路由、完整跨页面 smoke、桌面/移动浏览器验证与机器可读导出，同时保持相同的 ATTENTION/HOLD/BLOCKED 事实边界。

上一阶段整改控制与服务就绪计划：v3.0.0-v3.1.4。首轮投影为 5 个通过、8 个需关注、0 个不可用和 2 个仅外部可满足；控制面进一步分类为 2 个 satisfied、3 个 open、8 个 blocked 和 2 个 external-only。独立外部签收仍为 0/15，分发保持 HOLD，生产保持 BLOCKED。

上一阶段已验证 source gate：v3.0.0-v3.1.4 记录 125/125 测试、75/75 CI 路由、完整 smoke、生产构建、安全预检和桌面/移动浏览器验证，同时保持 ATTENTION/HOLD/BLOCKED 事实边界。

最新整改执行与运营验收计划：v3.2.0-v3.3.4。7 个直接 owner 动作现已具备确定性幂等键、短租约、围栏、回滚和证据指纹；当前投影为 0 个 satisfied、3 个 ready、4 个 blocked。独立外部签收仍为 0/15，分发保持 HOLD，生产保持 BLOCKED。

最新已验证 source gate：v3.2.0-v3.3.4 记录 131/131 测试、77/77 CI 路由、完整 smoke、生产构建、安全预检和桌面/移动浏览器验证，同时保持 ATTENTION/HOLD/BLOCKED 事实边界。

当前 Owner 工作负载与运营决策计划：v3.4.0-v3.5.4。7 类 owner 工作负载现已具备严格请求与候选回执合同；源码投影为 5 个通过、8 个需关注和 2 个仅外部可满足。外部证据仍为 0/15，分发保持 HOLD，生产保持 BLOCKED。

当前已验证 source gate：v3.4.0-v3.5.4 记录 137/137 测试、全部 11 个变更 TypeScript 分区、79/79 CI 路由、完整 smoke、生产构建、安全预检和桌面/移动浏览器验证，同时保持 ATTENTION/HOLD/BLOCKED 事实边界。

当前回执与异常生命周期计划：v3.6.0-v3.7.4。仓库现在具备带鉴权的 append-only 回执事件、隔离、补偿、升级确认、限时豁免到期和确定性决策包；源码投影为 6 个通过、7 个需关注、2 个仅外部可满足。本轮源码验证未提交真实 owner 回执，外部证据仍为 0/15，分发保持 HOLD，生产保持 BLOCKED。

上一阶段已验证 source gate：v3.6.0-v3.7.4 记录 146/146 测试、全部 11 个变更 TypeScript 分区、81/81 CI 路由、完整 smoke、生产构建、发布事实、架构/持久化边界和安全预检。

当前竞品产品计划：v3.8.0-v3.9.4 已实现 15 个仓库自有契约，并在 /experiments 分开显示 source/local/external/production 状态。真实 Provider 账号、workload scorecard、模型传输与 runtime、OS 沙箱、多模态授权数据、托管 RAG、非 MLX 训练 worker、多用户身份和独立签收仍是明确门槛；分发保持 HOLD，生产保持 BLOCKED。

当前已验证 source gate：v3.8.0-v3.9.4 记录 156/156 测试、全部 11 个变更 TypeScript 分区、85/85 CI 路由、完整 smoke、生产构建、架构/持久化检查和安全预检。源码为 15/15 通过；本地证据为 12/15 通过，路由质量样本、多模态授权执行和真实 Hub 传输证据继续保持 HOLD。

当前证据运营计划：v4.0.0-v4.1.4 新增严格、可持久化的候选回执合同，并在 /experiments 分开显示 15 项 source/local/external/production 门禁。候选回执只能闭环本地证据；独立签名、不可变归档回读、分发批准与生产授权继续保持 HOLD 或 BLOCKED。

最新已验证 source gate：v4.0.0-v4.1.4 记录 164/164 测试、全部 11 个变更 TypeScript 分区、86/86 CI 路由、完整 smoke、53 页生产构建、架构/持久化检查和已归档的修复后零漏洞安全预检；最新在线审计重跑因 npm registry 超时而正确 fail-closed，并非检出漏洞。源码为 15/15 通过；真实本地证据为 3/15，候选回执为 0，外部签收为 0/15，分发继续 HOLD，生产继续 BLOCKED。

当前证据采集计划：v4.2.0-v4.3.4。15 个源码机制均已实现：队列控制、Provider 目录、20 条路由质量/Shadow 执行器、复核/导出和 8 个 feature-owned 领域执行器。领域执行器会在回执提交前拒绝 reference、fixture 与 derived read model；本地确定性控制门禁为 4/15，真实 workload 证据仍待采集。不同复核 ID 不等于已认证的独立身份；验证记录 区分隔离测试与真实 workload 证据。分发保持 HOLD，生产保持 BLOCKED。

当前证据审定计划：v4.4.0-v4.5.4。源码门禁记录 固化了 15 个已实现源码合同与 7 个本地确定性控制；尚无已接受 workload package、外部验证的复核身份、不可变归档读回或独立复观察，因此外部与生产状态继续为 HOLD / BLOCKED。

当前证据验收生命周期计划：v4.6.0-v4.7.4。源码门禁记录 固化了 15 个源码合同、8 个本地确定性生命周期控制与生产构建 route-smoke 回执；尚无受控外部权威、不可变归档读回、当前已核验 package、独立观察者或分发批准，因此外部与生产状态继续为 HOLD / BLOCKED。

当前权威证据活动计划：v4.8.0-v4.9.4。源码门禁记录 固化了 15 个源码合同、9 个本地确定性控制与生产构建 route-smoke 证据；尚未配置受控凭据解析器、live authority profile、campaign、外部签收回执、不可变归档读回、独立观察者或分发批准，因此外部与生产状态继续为 HOLD / BLOCKED。

内部后续证据已推进到 v1.6.6 Benchmark 决策智能检查点：500 条结果全部完成分类（160 正确、340 条数学不等价，答案提取覆盖 100%），置信区间规则标出 Intermediate Algebra、Precalculus 与 Level 5 风险，500 样本下的保守可检测变化约为 8.27 个百分点。当前只有一个不同 run id 的完整运行，因此候选晋级继续为 EVIDENCE NEEDED。该检查点不改变当前源码版本，也不宣称托管榜单等价、独立主机复现、多模态全量执行、分发或生产晋级。

最新桌面候选说明：v1.1.0-rc.2。分发状态继续为 HOLD、生产状态继续为 BLOCKED；仓库不会把缺失的 Apple、组织、非回环、分布式 worker 或多人协作 receipt 表述为已完成证据。

本表基于 2026-08-31 可查的官方产品与模型文档。核心表示产品原生主流程，已集成表示具备能力但不是最深的专长，生态表示通常依赖相邻客户端或插件组装。厂商自报 benchmark 不直接算作本项目证据。

当前官方模型/Agent 雷达包括 OpenAI gpt-5.6-sol、Anthropic claude-fable-5 / claude-opus-5 与 Managed Agents、Google gemini-3.7-flash 与 Antigravity、DeepSeek V4、MiniMax M2.7、Kimi K2.6、智谱 glm-5.2 以及 Qwen3-Coder / Qwen Code Agent Team。它只是 provider 集成雷达，不代表本机已配置或已购买全部模型。

First LLM Studio 的优势是统一证据链。v3.8.0-v3.9.4 的 15 个源码契约现已覆盖 Provider 型号漂移、Agent conformance/cache、多 Agent 隔离计划、沙箱策略、Local Server、RAG 生命周期、训练后端联邦、团队资产策略与不超过 14 天的竞品 freshness gate；缺少的真实 workload 和外部回执继续显示为 HOLD。完整方法、优劣势、官方来源和版本顺序见：竞品格局与产品方向。

• 在统一上下文预算下，对比 MLX 本地模型和托管 API。
• 不离开应用就能查看 runtime 成本、prewarm、release、恢复动作和硬件压力。
• 判断哪个本地模型真的适合日常 coding / analysis 工作流。

• 在一个工作台里验证 tool calling、repo-grounded behavior、replay 和 patch 流程。
• 直接把 Compare 结果送入 Benchmark，不必切换产品。
• 区分失败来源：模型质量、provider 行为，还是本地 runtime 不稳。

• 用可复现 profile 跑 formal 和 focused benchmark suites。
• 查看 baseline、delta、run note、失败分类和发布证据。
• 让本地与远端 target 落在同一个可比较的 target catalog 里。

• 本地 + 远端统一 target catalog。
• Compare Lab 支持模型对模型审阅。
• Fine-tune 工作流覆盖 dataset、recipe、training、evaluation、adapter proof loop、export、best-checkpoint 选择和 LoRA 发布证据。
• 可视化 Workflow Studio 覆盖 Agent/RAG/eval 类型图、不可变 recipe、受保护工具执行、回放与 OpenAI-compatible 部署。
• Benchmark 运维覆盖 history、progress、baseline、report 和 release evidence。
• Retrieval 前台覆盖文档导入、chunk 检查和 grounded evidence probe。
• Experiments 时间线覆盖 Session/Run lineage、artifact 导航和 retention policy。
• Replay、trace review、patch inspection 与可导出的审阅记录。
• Runtime 运维覆盖 prewarm、release、restart、日志检查、telemetry 和 recovery。
• 支持本地/社区模型发现和远端 provider health 扫描。

• Local Qwen3 0.6B
• Local Qwen3 4B 4-bit
• Local Qwen3.5 4B 4-bit
• Local Gemma 3 4B It Qat 4-bit

• OpenAI Codex
• OpenAI GPT-5.5
• Claude API
• DeepSeek API
• Kimi API
• GLM API
• Qwen API

Target 选择、稳定性与适用任务对照：docs/benchmark-lane-comparison.md。

贡献者入口：English · 中文快速上手 · GitHub 仓库设置清单。

2026-08-30 的证据刷新从当前运行中的应用生成了新一批高清实机材料：

• Runtime Fabric：真实 MLX、Ollama、llama.cpp 3/3 后端通过，adapter contract 6/6，标准化操作 42/42。
• Local Server：真实 Ollama 0.31.1 + Qwen3 0.6B 通过 15/15 项，完成 6 个请求、95 tokens，平均延迟 169 ms。
• Benchmark smoke：Local Qwen3 0.6B 完成 3/3 次运行，首 token 192.67 ms，总延迟 504.67 ms，吞吐 234.02 tokens/s。
• MATH-500：完整 500 题输出全部判分，官方等价判分器决定 500/500 重放一致，本地准确率 32%。
• Fine-tune：真实 Qwen3 4B LoRA 归档保留 816 steps、save/eval 事件和 step 800 最佳 checkpoint。

完整指标、摘要、截图尺寸与证据边界见：v3.1.4-high-resolution-live-machine-capture-2026-08-30.md。已通过的项目仅代表本地源码/实机验收；缺少独立外部证据的生产晋级继续保持 HOLD 或 BLOCKED。

以下截图来自本地运行版本，并经过类型、测试、构建、路由与截图完整性验证。1920x1200 截图视口使用 2x DPR，路由图达到 3840x2400；证据面板保留原生高清裁切，LoRA 图从 SVG 以 2x DPR 导出。

Agent 工作台：target catalog、runtime rail 与工具化输入区：

Workflow Graph Studio：可拖拽类型节点、版本/修订控制、执行恢复与 promotion evidence：

可复现动态演示流程：docs/demo-video-workflow.md。

查看 Agent 工作台 MP4 演示 · SHA-256 元数据

Fine-tune Studio：工作流 tab、训练控制与 report/evidence 面板：

Fine-tune 完成作业：真实 loss 曲线、训练/验证轨迹与 handoff 操作：

真实 Qwen3 4B LoRA 发布证据：包含 save/eval 事件标记和自动选择的最佳 checkpoint：

矢量版本：fine-tune-qwen4b-lora-chart.svg。完整 run archive 与 manifest：docs/release-evidence/finetune-qwen4b-lora-2026-07-01。

Benchmark Studio：运行控制与历史证据卡片：

Benchmark：本轮本地实机 smoke run 生成的评测证据：

完整 MATH-500 结果：学科/难度分层、Wilson 置信区间、判分器重放、延迟与 token 性能：

Models Studio：不可变 Hub/外置盘证据、真实 Ollama Local Server 验收，以及真实 MLX/Ollama/llama.cpp Runtime Fabric 矩阵：

本轮真实 Runtime Fabric 与 Local Server 验收面板：

MCP 与安全扩展验收：签名生命周期、真实工具发现、隔离/检疫防御和显式生产门禁：

Compare、Retrieval 与 Admin：

运行整改与服务就绪控制面板，尚未完成的外部门禁保持可见：

• Apple Silicon macOS
• Node 22.x
• Python 3.12
• 可运行 MLX 的本地环境

nvm install 22
nvm use 22
npm install
cp .env.example .env.local

npm run dev

默认入口：

• http://localhost:3011/agent
• http://localhost:3011/compare
• http://localhost:3011/fine-tune
• http://localhost:3011/models
• http://localhost:3011/benchmarks
• http://localhost:3011/retrieval
• http://localhost:3011/experiments
• http://localhost:3011/admin

python3 -m venv .venv
source .venv/bin/activate
pip install -U pip
pip install mlx mlx-lm
python scripts/local_model_gateway_supervisor.py

如果 Python 不在 PATH 中，可以在启动前设置 LOCAL_AGENT_PYTHON_BIN 指向实际解释器。

网关健康检查：

• http://127.0.0.1:4000/health

npm run typecheck:changed
npm run smoke:routes
npm run smoke:screenshots

把 .env.example 复制成 .env.local，只填写你要启用的 provider 即可。

需要注意：

• .env.local 已被 git 忽略。
• 远端 provider 是可选的。
• 部分 target 走 OpenAI-compatible / Claude-compatible endpoint。
• 本仓库公开版本已经做过脱敏，占位值需要替换成你自己的 endpoint。

app/                      Next.js app routes 与 thin API transports
components/               共享 UI 与兼容 shell
features/                 feature-owned routes、contracts、state、actions、application ports
lib/agent/                Agent runtime、providers、benchmark、gateway helpers
lib/finetune/             Fine-tune store facade 与拆分后的 operation services
scripts/                  本地网关、runtime、验证和发布脚本
docs/                     架构、release notes、launch notes、roadmap 和 assets
modelscope/               ModelScope 主页/readme 元数据
public/                   对外资源和社媒封面图

• GitHub: https://github.com/ChrisChen667788/Your-First-LLM-Studio
• ModelScope profile: https://www.modelscope.cn/profile/haozi667788
• 默认 ModelScope repo id: haozi667788/first-llm-studio

ModelScope 打包脚本会导出已提交的 Git tree，因此每次同步都可以让 GitHub 和 ModelScope 保持同一份文件快照。

• 敏感本地操作默认需要确认。
• Secret 应保存在 .env.local。
• 公开仓库默认配置已经做过脱敏。
• 新增公开提交应尽量使用 GitHub noreply 地址。
• 见 SECURITY.md。

欢迎 issue 和 PR。

• CONTRIBUTING.md
• CODE_OF_CONDUCT.md
• docs/open-source-backlog.md

• 当前版本：VERSION
• Release notes：docs/releases
• 发布流程：docs/release-process.md
• 当前源码：v1.5.1；最近公开 GitHub Release：v0.4.0；最近桌面候选包：v1.1.0-rc.2