Meta returns to open weights with strings attached; DeepSeek ships anyway

The big story

Meta's Muse Spark Open is the model Zuckerberg promised in September: 816B total, 42B active, 1M context, text/image/video/audio in, text out, and Meta's own claim of "most capable US open-weight model to date." The weights are real and on Hugging Face, but gated, and the Muse Model License is Llama-style rather than open: derivatives must carry a "Built with Muse Spark Open" notice, and anyone selling model access as a service with more than $20M in trailing-twelve-month revenue needs a separate agreement with Meta. For a hobbyist or an internal deployment that is a non-issue; for inference hosts it is the whole contract. Meta also co-launched with Apple, whose MLX team showed the NVFP4 build (~493 GB) decoding at 25 tok/s across two M5 Ultra Mac Studios with tensor parallelism and 36.7 tok/s across four, over RDMA-on-Thunderbolt 5. That is a genuinely new hardware story for this size class.

The reason to temper the celebration arrived two days later. DeepSeek-V4.1-Flash-1017 is billed as a "minor update" to September's V4.1-Flash, but the model card's table has it beating the prior checkpoint on all eleven benchmarks, and beating Kimi K3, GLM-5.3 and Claude Opus 5 on DeepSWE v1.1 (77.0), Automation-Bench and Agents' Last Exam, with shorter outputs. It is MIT-licensed, ungated, and the community had GGUFs, an abliterated variant and a ds4 PR within 24 hours. Meta's benchmark scorecard has not yet been independently reproduced, and nobody has posted a head-to-head. The pattern of the last year holds: US labs release big weights with conditions; a Chinese lab releases slightly better ones without.

New open-weight releases

• Muse Spark Open (816B MoE / 42B active, Muse Model License) — Meta's first frontier-scale open-weight release since the Llama era; 1M context, natively multimodal, gated repo with attribution and $20M MaaS clauses. Build guide.
• Beam 1.0 (501B MoE / 23B active, Apache-2.0) — Reflection delivered the weights on October 16 as promised, in BF16, FP8 and NVFP4 plus a speculative-decoding drafter; text-only, 1M context, pitched on inference efficiency rather than peak capability. llama.cpp support is already merged. HF blog.
• DeepSeek-V4.1-Flash-1017 (552B backbone plus engram parameters, 763B on the HF counter; 8B prefill / 16B decode active, MIT) — point release with an updated vision tower and a new reference prompt-encoding library. Model card.
• Step 5 Preview (600B MoE / 27B active, StepFun Model License 1.0) — StepFun hit its October 15 date after a shaky morning. Not Apache-2.0 like Step 3.5 Flash: attribution kicks in above 100M MAU or $20M monthly revenue, and hardware makers and large MaaS operators need a commercial agreement. FP8 checkpoint is 650 GB. HF.
• FLUX 3 Image Klein (5.6B DiT, FLUX Kommunity License) — Black Forest Labs' small open release from the FLUX 3 family: text-to-image, editing, up to 10 reference images, 4K output, INT8 build on 8 GB cards. Free commercial use only under $5M annual revenue. Announcement.
• LFM2.5-24B-A2B (24B MoE / 2B active, Liquid's own licence) — reasoning-only agent model with 131K context; Liquid claims 184 tok/s on an M5 Max and parity with Qwen3.6-35B-A3B on tool-use benchmarks. HF blog.
• Nemotron 3.5 Bolt 74B-A9B and Super-VL 120B-A12B — NVIDIA shipped Bolt as a 2-bit QAT checkpoint sized for a 24 GB card, and a 120B vision model with LiLiCorr draft weights; early LocalLLaMA testing on data-analysis tasks was lukewarm. Licence per NVIDIA's cards.
• Also out: OpenMOSS's MOSS-Image (base and SFT, Apache-2.0), OpenBMB's MiniCPM-V 4.7 (35B-A3B and 1B), inclusionAI's Ling 3.1 Flash (560B / 25B active), Aleph Alpha's Kolibri-1 (78B / 3.4B active), Fish Audio's Drama 3 TTS preview, and Jais 3 from 0.9B to 375B for Arabic.

Policy & politics

On October 16 China's commerce and science ministries published a draft adding 23 AI technologies — model weights and code, agents, compute clusters, chip design — to the export-restricted catalogue, covering models above 2T parameters or trained with 10^26 FLOPs or ranked top-10 globally. Two carve-outs matter: technology already public via open source or papers is exempt, and API/app/SaaS access is not an export. Comments close November 15, but an interim pre-export registration rule took effect the same day. The open-source exemption is being read as reassurance, and the ministry says it "encourages" open development. Read it the other way: the exemption covers what is already open, and the licensing requirement attaches to the decision to open the next thing. Kimi K3 at 2.8T is the only model on our index above the parameter line; a K4 or a trillion-scale Qwen would need someone in Beijing to agree that releasing it is fine. That is a lever that did not exist last week.

Separately, Anthropic's head of international told Bloomberg TV in Dublin that open models carry "more risk" by definition, a broader line than Amodei's July position. It landed the same week Nvidia's open-weights coalition letter — which Anthropic still has not signed — kept picking up signatories, and LocalLLaMA filed it under free marketing. Mistral's Large 4 (1T / 49B active) remains API-only until October 27, and under a custom licence rather than Apache-2.0 this time.

Ecosystem

The practical news is that 500B–800B MoEs are now runnable on desks. Apple's numbers above are the official story; the unofficial one is Reddit: DeepSeek 1017 at EXL3 2.9bpw on two DGX Sparks at 57 tok/s, the Strata-derived Xolotl engine claiming ~27 tok/s for the 552B model on a single 4090 with SSD streaming, and Muse Spark Open limping to 4 tok/s on a 256 GB M3 Ultra via ds4. Treat the single-GPU numbers as unverified until someone posts KL divergence, which is exactly the complaint running through the Strata threads this week: the speed is real, the quality cost is undocumented.

In llama.cpp, Muse Spark Open support merged in four hours, Beam and Kolibri-1 landed, and a CUDA MMVQ rewrite gave a large speedup for Q2_K/Q3_K decoding. Hugging Face shipped 200+ WebGPU kernels, and AI2 published BenchMIRT, an item-response analysis of what benchmarks actually measure. A LocalLLaMA thread alleged that AutoTrust's trending "decision models" are riding farmed likes; the accusation is unproven, but HF Trending is now a marketing surface and should be read as one.

From the index

Our DeepSeek-V4.1-Flash listing (510 GB, MIT) is the September checkpoint; the 1017 refresh has a new vision tower and is not a drop-in file swap, so it needs its own torrent and we will seed it. Beam is Apache-2.0 and the easiest add of the week. Muse Spark Open is the one that most needs a torrent, since the HF repo is gated behind a legal-name form, and the licence explicitly permits redistribution provided the licence text and notice travel with the weights — so they will. Meanwhile the thread on Kimi K3 three months after launch is worth reading next to our 1,561 GB K3 listing: it remains the largest file on the index and the model fewest people can actually run, which is precisely why an archive copy matters. GLM-5.3 at 753B sits comfortably below Beijing's proposed threshold; K3 does not.