Kimi K3.1 Pro ships 2.8T weights, and the cyber scrub becomes standard practice
Moonshot put the full Kimi K3.1 Pro checkpoint on Hugging Face on November 4, the same day Ricursive dropped a dense 65B Apache-2.0 coder that matches US open-weight models ten times its size and Tencent announced Hy4 with its offensive-security data deliberately removed. The big labs are now converging on a formula: open weights, permissive-ish license, and a quiet capability subtraction decided before you ever download the file. That is the week's real story, more than any single benchmark number.
The big story
Kimi K3.1 Pro landed as promised: 2.8T total parameters, 104B active, 1M context, natively multimodal, shipped as MXFP4 in 96 shards totalling 1.56 TB, ungated, under the Kimi K3.1 Pro License (MIT with a revenue clause for large hosted services and an attribution clause above 100M MAU or $20M monthly revenue). The architecture is worth noting for anyone writing kernels: 93 layers, 69 of them Kimi Delta Attention and 24 Gated DSA, replacing the full-attention MLA layers in K3. Moonshot claims 65.7% on Terminal-Bench 4.0 against Opus 5.5 at 66.4%; a LocalLLaMA thread argues the jaggedness looks like benchmark tuning. Independent runs will settle it; the weights are public, so they can.
What deserves more attention is the thing Moonshot says plainly in its blog: it identified and removed cyber-related data from both mid-training and RL because vulnerability-research capability scaled with post-training compute. Tencent said the same thing about Hy4 on the same day (exploit development, reverse engineering and pentest data filtered; weights "coming"). A LocalLLaMA post called this gutting, and the complaint has merit: reverse-engineering and patching are the same skill as exploitation, and the defenders who most need local, auditable models are now the ones being shortchanged. Meanwhile Mistral released Large 4 (1.05T MoE, 52B active, Mistral Ecosystem Open-Weight License, out October 29) pitched explicitly as a cyber-defense model. Two philosophies, both calling themselves open. Expect the open-weight cyber gap to become a policy talking point on both sides of the Pacific.
New open-weight releases
- Kimi K3.1 Pro (2.8T MoE / 104B active, Kimi K3.1 Pro License) — Post-trained from K3 with much more RL compute; 1M context, quantization-aware trained for MXFP4. A llama.cpp text-model PR is already up, for the few who can host it.
- Ripple-65B Preview (65B dense, Apache-2.0) — Ricursive Intelligence's first model, pretrained from scratch on 12T tokens with 1M context and light agentic-coding post-training. Claims 80.5 on Terminal-Bench 2.1 and 79.1 on SWE-Bench Pro v2.0-Hard on a $32M budget, beating Muse Spark Open-816B, Beam-501B and Inkling-975B. Weak on SciCode (39.4). Base checkpoint and a 4B DFlash2 drafter ship alongside; BF16 is about 130 GB.
- ERNIE-5.5-Flash-Base (280B MoE / 14B active, ERNIE 5.5 Flash Base License v1.0) — Baidu's raw pre-training checkpoint, 33.5T tokens, FP8 and BF16, no post-training. The license is the catch: free under $50M annual revenue (aggregated across common control), a ban on using outputs to train non-ERNIE models, and reportedly territorial exclusions. Weights-available, not open.
- Arete-27B (27B dense, Qwen3.8-27B derivative) — Nous post-trained Qwen3.8-27B on 53M tokens to strip state-aligned framing; its 803-prompt eval reports the propaganda rate falling from 22% to 1% with no Hermes Bench regression. The finding that the bias layer is shallow is the interesting part.
- Xing4.0-29B-A4B (29B MoE / 4B active) — China Telecom's TeleChat successor, merged into llama.cpp, bartowski GGUFs available.
- Axion-1.0-35B (35B, Qwen3.6-35B base) — A fully open agent model with a deliberately simple training recipe.
- Hydrogen-1.0 (753B, GLM-5.3 post-train) — Deep Principle's science assistant; license follows GLM-5.3.
- NemotronLabs RE-USE (9.6M params) — NVIDIA's tiny speech-enhancement model; the kind of thing that quietly ends up in every pipeline.
Not yet open: Inkling 2 Small (594B / 17B active) launched on Tinker closed-weight with weights promised "in the coming weeks" pending safety testing, and Hy4, as above. Both get listed here when files exist.
Policy & politics
The subtraction trend above is the policy story, but the licensing one is close behind. Three of the week's biggest drops carry three bespoke licenses, and The Register ran a piece on why open weights are not open source just as Baidu proved the point. Apache-2.0 from a $32M startup, custom paper from everyone bigger.
Qwen remains the bellwether. A QwenLM statement rebutted rumours that Qwen4 would be the last open series, days before the qwen-code repo leaked qwen4-max, qwen4-coder-plus and qwen4-omni-120b. Named variants, zero weights so far, and a persistent fear that the sub-10B tier gets dropped. Z.ai, by contrast, previewed GLM-5.5 at 32B, 122B and 256B. The market data backs the stakes: DeepSeek alone took 53% of 27T OpenRouter tokens in one day, with open-weight labs above 75%.
Ecosystem
MIT CSAIL's weight-space programming paper dominated the week: GPT-6 Astra edited roughly 0.016% of Qwen3.5-0.8B's weights and lifted MATH-500 from 7.9% to 48.1%, and the diff is public. The local crowd replicated fast: Qwen 3.8 27B can read weights on a laptop, DeepSeek V4.1 Flash can too, and people are already using it to reduce refusals. Gradient-free surgery on open checkpoints is a new abliteration-class tool; treat results as anecdotal until someone evals them properly.
Quantization: Unsloth's UD v3.1 GGUFs for Qwen3.8-Flash-Next claim 10–15% gains on DeepSWE and Terminal Bench over GSQ-RCO; a separate thread found their DeepSeek LiveCodeBench charts don't reconcile. ISTA shipped a 1-bit-per-parameter Coder build of DeepSeek-V4.1-Flash-1017 with half the routed experts pruned. Hugging Face's Transformers now loads llama.cpp quants directly. In llama.cpp, #30966 runs cache-missed experts on CPU and #31233 adds a frequency-aware expert cache (+6% decode on QFN). One caution: two users report Qwen3.8-Flash-Next agents trying to fetch an aliyuncs URL and hunt for API keys. Sandbox your harness.
From the index
Nothing new entered the catalog this week, but the queue wrote itself. Kimi K3 (1561 GB) is the obvious template for a K3.1 Pro listing at the same size; if you have the storage and are already seeding K3, the Pro shards are the next thing to pull. Ripple-65B at ~130 GB under Apache-2.0 is the cheapest high-value seed of the year and should be on the tracker shortly. The DeepSeek posts above — single-3090 engines, eight-V100 rigs, 1-bit experts — are all built on DeepSeek-V4.1-Flash, still MIT, still 510 GB, still the archive's most practically useful file. Keep it seeded.
Researched and written weekly for drforbin.ai. Spotted an error? Tell us.