1.0.3 — HunyuanVideo, reliable tool calls, and Desktop LAN and folder chat

Rapid-MLX 1.0.3 adds HunyuanVideo 1.5 text-to-video, faster decode on qualified models, and more reliable tool calling and structured answers. Desktop adds opt-in LAN access, folder Q&A and a memory overview.

• HunyuanVideo 1.5. Text-to-video at 480p on Apple Silicon Macs with 64 GB or more unified memory. hunyuan-video-1.5-480p and hunyuan-video-1.5-480p-distilled use the existing video jobs, progress, cancellation and MP4 output. The base checkpoint downloads about 53 GB directly from Hugging Face. The CFG-distilled variant uses one transformer pass per step instead of two, still at 50 steps. This family supports text-to-video only; the Tencent Hunyuan Community License excludes the European Union, United Kingdom and South Korea. (#5679, #5687)
• Faster decode on qualified models. Dense Qwen3.5 models gain guarded decode fusions with speculative decoding off; Qwen3.5/3.6 35B-A3B MoE models gain device-specific tuning for plain decode. With a 564-token prompt, 256 greedy output tokens and a fresh server per repetition, Qwen3.5 4B 4-bit measured 166.89 → 187.86 tok/s on M3 Ultra and 81.76 → 85.68 tok/s on M4 Pro. Separately, 35B-A3B 4-bit singleton tuning measured 157.19 → 177.00 tok/s on M3 Ultra and 167.27 → 171.81 tok/s on M5 Max. These gains use separate baselines and are not additive. Compatible dense models gain fused decode operations; MTP avoids unused draft-head work and improves verification. Guarded fused 4-bit MoE experts use approximate arithmetic, so output can differ slightly. (#5690, #5728, #5695, #5674, #5689, #5694)
• More reliable tool calls. Dotted or slash-separated tool names resolve to a unique declared tool after separators are normalized to underscores; exact matches take precedence. Gemma 3 12B and 27B support structured, streamed and parallel tool calls and tool-result history; the 1B and 4B catalog variants no longer advertise tool support. Forced and required Anthropic-compatible tool choices generate arguments, and automatic choice works with parallel calls disabled. Existing Qwen and GLM parsers forward undeclared calls so clients can return tool errors; the built-in agent still refuses to run unknown tools. (#5788, #5715, #5750, #5724, #5723, #5700)
• Desktop LAN access. Settings → Tools can share inference with devices on your network using a persistent bearer key saved in Keychain and copyable connection details. Access is off by default and uses plain HTTP. Model management, administrative routes and Local Search, including grounded chat, stay on this Mac. (#5676, #5749)
• Ask this folder… Start a text-only chat scoped to one registered Local Search folder, with a removable folder chip and saved passage previews below answers. Requires Built-in Agent Harness (Experimental); tools are disabled. Without matching evidence, the app abstains without inference, including for unrelated questions in small folders. Saved excerpts survive file or index changes but do not guarantee a correct answer. API clients can request grounded chat from explicitly selected registered folders, with citation IDs and source locators. (#5654, #5769, #5652, #5638)
• Desktop memory and local checkpoints. Performance settings shows app, wired, compressed and unused memory alongside engine allocations, prompt-prefix cache use, running and waiting requests, and a five-minute trend. Engine allocations overlap the Mac totals; missing measurements show a dash. Settings → Models → Open existing checkpoint… registers a complete local MLX chat checkpoint without copying weights or changing the download folder; Remove reference leaves the files intact. (#5651, #5677, #5660)
• Benchmarks and model recommendations. Desktop Benchmark is a regular sidebar tab; nothing runs until you start. rapid-mlx benchmark run --configs N (2–16) compares the current Rapid Lattice configuration with nearby configurations on the same Mac, using interleaved fresh-load repetitions (--repetitions, 3–10). Recommendations and CLI models and doctor show offline speed references matched to chip, RAM and model configuration; Generation details saves supported execution facts and timing with answers. Unmatched data stays unavailable; references are not performance guarantees. Base and Pro Macs with 48 GB or more lead with Qwen3.6 35B-A3B MoE; Max and Ultra Macs keep the dense 27B first. MoE labels show total and active parameters. (#5781, #5763, #5647, #5655, #5659, #5784)
• First run, downloads and speech. The installer and cold-cache CLI chat use the LFM2.5 2.6B starter, about a 1.6 GB download; installer recommendations use the smaller starter below 8 GB or when RAM is unknown. pull --check includes free disk space, corrupt cached files get one targeted repair attempt, and Desktop retries an unforced failing mirror through Hugging Face. Idle transcription, alignment and speech-synthesis models can unload while the API stays online, then reload on the next request; CLI users opt in with --audio-lane-idle-ttl. Standard TTS model names resolve to the default local speech model, and audio-enabled installers prepare Kokoro’s English prerequisites. (#5631, #5632, #5725, #5736, #5666, #5635, #5712, #5713)
• Structured answers and agent setup. JSON output requests follow the same thinking preference as plain chat. Structured output preserves reasoning and the system-prompt prefix, including for vision; constraints start after reasoning. Strict Responses streams no longer duplicate JSON or append reasoning to it, and strict schema violations return HTTP 422 json_schema_violation. Vision reasoning budgets are enforced during generation. Agent setup skips servers needing a missing key or serving non-chat models and gives a working next step. (#5756, #5675, #5678, #5682, #5707, #5706, #5727, #5730, #5594)
• Desktop recovery and privacy. Custom-folder models remain recognized after unloading. Model search results appear below the search box; long and colliding names keep their distinguishing text. External-volume access failures show an explanation and Retry; disk-full downloads retain partial files. Server-backed conversation deletion remains pending across relaunch until confirmed, then revokes saved-passage access; original files and indexes remain, and retained database and attachment bytes are not physically purged. Consented anonymous metrics include Desktop launches, daily foreground use and coarse startup and agent-readiness failure categories, excluding raw messages, paths and credentials; concurrent first runs share one install identity. (#5770, #5633, #5669, #5733, #5670, #5766, #5746, #5658, #5697, #5634, #5698, #5668, #5709)
• Upgrade notes. Run rapid-mlx upgrade, brew upgrade rapid-mlx or pip install -U rapid-mlx==1.0.3. Forty-four niche or superseded aliases leave discovery; old names warn and forward for one release, with repository execution profiles preserved. glm4.7-flash-4bit replaces glm4.7-9b-4bit and is labeled MoE; existing downloads are reused. Always-reasoning models default to 4,096 CLI output tokens when no explicit limit is given; --max-tokens still wins. Dynamic loading defaults to the smaller of 80% of physical RAM and 90% of Metal’s recommended working set; explicit memory limits retain their behavior. pull and chat accept --force-disk-check. (#5684, #5748, #5747, #5701, #5754)
• Desktop 1.0.3. Published 2026-11-11 as rapid-mac-v1.0.3: opt-in LAN access, folder-scoped chat, memory overview, opening existing checkpoints, measured speed references, a Benchmark sidebar tab, idle speech unloading and model-selection, download and conversation-recovery fixes. Download Desktop 1.0.3.