Topic
Open Models
19 stories, newest first.
- 01Open Models18 Oct 2026
Kimi K3 vLLM decode up to 2x faster with coding agents
In an Oct 16 vLLM blog post, Benjamin Chislett (NVIDIA) describes a unified CuTe DSL decode path for Moonshot's open 2.8T-parameter Kimi K3 on GB300 GPUs. Per layer, stock vLLM took 130 µs and 23 kernels vs 69 µs and 13 specialized. With DSpark speculative decoding (synthetic AL=6) it reached 842 tokens/s at batch 1 on 16 GB300s (1.73x stock); without speculation, 256.4 tokens/s (1.88x). The team writes that agent-optimized inference code is outpacing review, and proposed a fast-track RFC. Caveats: GB300-specific, decode batches 1-64, gains in several tested configurations.
Source: vLLM Blog (opens in a new tab) - 02Open Models18 Oct 2026
Perplexity open-sources AI that only decides
Perplexity released Decider v1.2 on Oct 16 as open weights under Apache 2.0: pplx-decider-v1.2 (27B) and pplx-decider-flash-v1.2 (4B, distilled from the larger models). Instead of chatting, they answer yes/no, multiple-choice and scored questions about text or images with calibrated probabilities, the kind of small judgment calls agents and AI workflows make constantly. Version 1.2 reads a whole request in one pass, so several questions about the same input cost about as much as one, and the company says inference is up to 5x faster. On its own run of the public Decision Index, the overall score rises from 62.3 (v1.1) to 68.1 and vision from 73.2 to 77.4, and the 4B Flash model leads all 4B models on the Decision Index leaderboard. Decisions API pricing is $0.02 (27B) and $0.01 (Flash) per 1M input tokens, with output tokens free. Caveats: probabilities remain somewhat over-confident on held-out tasks, and training now includes official training splits of some Decision Index benchmarks (decontaminated against the test sets).
Source: Perplexity on Hugging Face (opens in a new tab) - 03Open Models18 Oct 2026
NVIDIA's looped Dory 4B lifts SWE-bench to 61.8
NVIDIA released Nemotron-Labs Dory1.0-4B on Hugging Face on Oct 16, 2026, a 4B-parameter looped (latent-recurrent) model: 8 input layers, a shared 20-layer block that repeats with the input reinjected every loop, and 8 output layers. One checkpoint serves different compute budgets with no retraining: by NVIDIA's own numbers SWE-bench Verified rises from 50.3 at 1 loop to 61.8 at 4 loops (Qwen3.5-9B scores 61), and Terminal-Bench 2.1 from 12.5 to 22.2 with about half the output tokens. Optional last-loop KV caching avoids extra KV-cache memory as loops grow, and it was pre-trained in 4-bit NVFP4 for the first ~4.5T of 6T tokens, with a 256K context. Caveats: extra loops cost compute per token, and Qwen3.5-4B still leads GPQA Diamond (77.7 vs 72.4). Weights under OpenMDW-1.1; vLLM support is under review.
Source: NVIDIA on Hugging Face (opens in a new tab) - 04Open Models18 Oct 2026
DeepSeek's open model tops Claude Opus 5
DeepSeek released DeepSeek-V4.1-Flash-1017 on Oct 17, 2026, a minor update to V4.1 Flash that it says finishes tasks faster with shorter outputs. By DeepSeek's own numbers it scores 77.0 on DeepSWE v1.1 versus 74.0 for Claude Opus 5, and leads Opus 5 on Automation-Bench (60.5 vs 50.3) and Apex Bench (72.1 vs 65.8); versus the previous V4.1 Flash, Terminal-Bench 4.0 rose from 31.2 to 42.3. Opus 5 still leads on Terminal-Bench 4.0 (51.8) and HLE without tools (56.3 vs 42.6). Weights are on Hugging Face under the MIT license with a 1M-token context, and API pricing is unchanged (deepseek-flash). Vendor-reported benchmarks; independent tests to come.
Source: DeepSeek API Docs (opens in a new tab) - 05Open Models17 Oct 2026
Reflection's Beam: a 501B open-weight model with only 23B active
Reflection, a new US lab, released open weights for Beam, its first open-weight model, on Hugging Face under Apache 2.0: a mixture-of-experts with 501B total parameters, 23B active per token and a 1M-token context. Built for coding and agents; Reflection's model card reports 81.6% on SWE-bench Verified and 80.1% on Terminal Bench 2.1, and says it needs ~3x fewer reasoning tokens than leading open models for comparable quality. Caveat: those are Reflection's own numbers, and even the NVFP4 checkpoint is ~280 GB (Hugging Face suggests four B200s).
Source: Reflection (opens in a new tab) - 06Open Models16 Oct 2026
Meta open-weights Muse Spark Open: 816B / 42B active
Natively multimodal with a 1M-token context, built for coding and agentic work. First in its class to run on two Apple M5 Ultra Mac Studios with MLX; APIs and Hugging Face weights live under the Muse Model License.
Source: Meta AI Research (opens in a new tab) - 07Open Models16 Oct 2026
StepFun open-weights Step 5 Preview: 600B, 27B active
Sparse MoE with up to 1M tokens of context for repo-level coding, long-horizon agents and finance. Serving examples start at four B300 GPUs (FP8) via vLLM or SGLang; StepFun Model License 1.0.
Source: Hugging Face / StepFun (opens in a new tab) - 08Open Models16 Oct 2026
Ant Ling open-sources 560B Ling-3.1-flash under MIT
Ling-3.1-flash is a 560B hybrid-linear MoE (about 25B active, up to 1M context) released with MIT-licensed open weights on Hugging Face and ModelScope, aimed at agents, office work and coding.
Source: inclusionAI / Ant Ling (Hugging Face) (opens in a new tab) - 09Open Models16 Oct 2026
Mistral's Timestral: 209M forecasting model tops 3 benchmarks
Mistral's first time-series foundation models (5M–209M params) run on a CPU; Mistral says the 209M model ranks first on GIFT-Eval, TIME and fev-bench with two-thirds the parameters of the previous best.
Source: Mistral AI (opens in a new tab) - 10Open Models16 Oct 2026
Span-02: open-weight decisions with a thinking deadline
Respan released Span-02 (Oct 15): open-weight decision model with budget_ms anytime reasoning; Decision Index 0.3 61.2 (off) / 64.1 (no limit) vs Jev 58.0; Blind Earth 71.9%→84.7% at 1s.
Source: Respan (opens in a new tab) - 11Open Models16 Oct 2026
Jais 3: six open Arabic agent models up to 375B
IFM and Inception42 released six open Arabic-centric models from 0.9B to a 375B MoE (23B active), 512K context, built on K2 Horizon. Their eval: 375B leads open models at 84.0 Arabic average; 7B leads its size class on agentic tasks. Apache 2.0 on Hugging Face.
Source: IFM / MBZUAI (opens in a new tab) - 12Open Models16 Oct 2026
Ultramarine 1 Flash: Fable-class work at $0.13/task
Paper Instruments RL-trained open GLM-5.3 Flash for finance, law, consulting, and ops: +4.1 pts avg on six knowledge-work benches, 6.5% fewer tokens. Their PI suite scores 59.3 at $0.13/task vs Claude Fable 5 at $8.19. JobBench 62.9. Weights coming to Hugging Face.
Source: Paper Instruments (opens in a new tab) - 13Open Models16 Oct 2026
IBM Granite Switch 4.1: chat, decisions, ASR in one endpoint
IBM's Granite Switch 4.1 packs chat, a Decision aLoRA, a Guardian safety adapter, and Granite-Speech-5.0-Turbo-CTC (fastest on OpenASR per IBM) into one endpoint at ~the same VRAM as the base. A 3B Doom demo hits 60–100 decisions/sec vs Jev's reported 10.
Source: IBM Granite / Hugging Face (opens in a new tab) - 14Open Models16 Oct 2026
Hugging Face: BERT tokenization ~50× faster than 0.23
Hugging Face's tokenizers work made BERT (WordPiece) encoding about 50× faster than v0.23 (English chart: 50.8×). Normalization was ~88% of encode time, now ~13%; token ids unchanged. WordPiece still drives ~627M Hub downloads/month. Still a PR at blog time.
Source: Hugging Face (opens in a new tab) - 15Open Models14 Oct 2026
Google DeepMind's fine-tuned Gemma spots scams on-device
Built with Singapore's GovTech: Gemma fine-tuned on local scam conversations flags manipulation on the phone, so encrypted chats aren't scanned in the cloud. A lightweight classifier screens messages, then Gemma checks the wider conversation. Training excluded private/PII data.
Source: Google DeepMind (opens in a new tab) - 16Open Models13 Oct 2026
Tiny open-source decision model beats Opus 5 and GPT-5.6
FLock.io THIS/THAT Model 1.2 (built on Qwen3.5-2B, 1.88B params, MIT) scores 0.878 accuracy on 1,710 public questions in 31.4 ms with zero output tokens — ahead of Claude Opus 5 (0.834) and GPT-5.6 (0.816). Runs on laptops and phones.
Source: FLock (opens in a new tab) - 17Open Models13 Oct 2026
Anthropic launches free OSS Scanner for open source
Opt-in free vuln-finding informed by Project Glasswing. CyberGym: LLMs from under 20% to over 85% vulns found. 29,000+ candidates, ~6,000 manually reviewed; validation 97 critical/high across 48 projects → 85 (88%) met CVD bar. Enroll via PR.
Source: Anthropic (opens in a new tab) - 18Open Models13 Oct 2026
Mistral Ministral 3 is running on a satellite in orbit
Orbitworks and Loft Orbital say the first satellite of a planned 50-satellite AI constellation is in orbit after an Oct 1 SpaceX rideshare. It flies two NVIDIA Jetson AGX Orin modules and runs Mistral Ministral 3 (3B and 8B) onboard to summarize acquisitions and answer operator queries — alongside nearly 20 other models. Alert in seconds instead of dumping raw imagery hours later; first bird of a $1B program.
Source: Loft Orbital (opens in a new tab) - 19Open Models13 Oct 2026
NVIDIA open-sources the full Run:ai platform
Full GPU orchestration stack joins NVIDIA DSX AI Factory as open source: KAI Scheduler (Apache 2.0, CNCF Sandbox), CUDA 13.4-native GPU fractions, and more — no paid Community Product subscription.
Source: NVIDIA (opens in a new tab)