Model directory

Open-source LLMs the community actually runs: parameter size, quantization releases, and test status at a glance — models without on-site benchmarks are labeled as such.

57 models

Sorted by recently added (newest entries first)

"Quantization releases" counts community and official releases; "benchmark runs" counts only records verified by this site. The footer "Models" figure counts models with at least one published benchmark — a different scope from the directory total here.

Kimi-K3.1-Pro-Highspeed

Moonshot AI

The high-speed serving variant of Kimi K3.1 Pro — the same weights, with officially claimed **2-3x** output speed for more responsive interactions. Pricing is double the standard version: **$4.00** input / **$20.00** output per million tokens, with cached input at $0.20 and cache writes at $4.00 (5 min) / $8.00 (1 h). Context and capability figures match K3.1 Pro (2.8T total / 104B active, 1M context, native multimodality, Terminal-Bench 4.0 at 65.7%). Live now alongside the standard build as kimi-k3.1-pro-highspeed; weights are promised by Nov 4, so the site logs it under the RESTRICTED convention with zero on-site runs.

2800BRestricted

Arena Elo —

no indexed quantization releases yetnot benchmarked yet

View

Kimi-K3.1-Pro

Moonshot AI

Kimi K3.1 Pro, the first model in the Kimi K3 series from Moonshot AI, released 2026-10-23 — a post-trained version of K3 billed as the company's most capable 3T-class open-weight model to date: **2.8T total / 104B active** parameters, a **1,048,576-token (1M) context window** and native multimodality (text/image/video). Release evaluations: Terminal-Bench 4.0 **65.7%** (ahead of GPT-6 Astra at 58.2%, just behind Claude Opus 5.5 at 66.4%), Kimi Code Bench 3.0 Long-horizon 61.2% (K3 scored 55.95%), BenchCAD with tools 92.7% (Claude Opus 5.5 96.2%, GPT-6 Astra 95.9%), and a state-of-the-art 64.4 on OfficeQA Pro v2, with matching upgrades in finance and 3D creation. Official pricing per million tokens: input **$2.00**, output **$10.00**, cached input $0.20, cache write $2.00 for 5 minutes / $4.00 for 1 hour — roughly 33% below the previous K3 ($3.00 / $15.00). The API is live today as kimi-k3.1-pro, alongside a Highspeed variant (2-3x faster, $4.00 / $20.00). **The full weights are promised by Nov 4**, so the site logs it under the RESTRICTED convention with zero on-site runs.

2800BRestricted

Arena Elo —

no indexed quantization releases yetnot benchmarked yet

View

NVIDIA-Nemotron-3.5-Bolt-E4B-4GB

NVIDIA

NVIDIA's Nemotron 3.5 Bolt, released 2026-10-23 (Google Gemma-4-E4B-it on Apache-2.0, post-trained by NVIDIA, post-training data cutoff May 2026): 8B total parameters with 4B effective, aimed at local voice and device assistants, interactive personas and embedded edge interfaces. Its focus is multi-turn tool use, instruction following and input grounding; the chat template controls thinking mode and conversational mode, with a tunable thinking budget. The official Q4_K_M GGUF measures **6.18GB** (the "4GB" in the repo name refers to the 4B effective-parameter class), and NVIDIA ships a llama.cpp Docker recipe (ghcr.io/ggml-org/llama.cpp:server-cuda13-b11459, -c 65536, --jinja) plus a DFlash draft model for speculative decoding (--spec-type draft-dflash --spec-draft-n-max 7). Context is 64K tokens in and out, across 36 languages. Q4_K_M benchmarks (against Gemma-4-E4B-it Q4_K_M / Qwen3.5-4B Q4_K_M): MultiChallenge 55.60, RoleBench 71.21, RoleMRC 79.12, ToolSandbox 74.13, IHEval 82.40, IFBench 64.00, StructEval-T 79.99, BFCL-v3 67.31, Tau2-Bench 52.32, AgentIF 22.95, RULER (64K) 85.76, LongMemEval 69.92, LongBench v2 (48K) 48.67 — the clearest gains over the base model are in tool use and role play. License OpenMDW-1.1 (the Gemma 4 base remains under Apache-2.0).

8BOpen weights

Arena Elo —

no indexed quantization releases yet26 benchmark runs

View

NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4

NVIDIA

NVIDIA's quantized release of Nemotron 3.5 Super VL (2026-10-23, NVFP4 with quantization-aware distillation from the BF16 model): a 120B-total / 12B-active vision-language MoE. The architecture interleaves Mamba-2 and Transformer attention layers in a hybrid MoE, with a C-RADIO vision encoder and a vision-language connector; context runs up to **1,000,000 tokens**, and native Multi-Token Prediction (MTP) enables speculative decoding (official LiLiCorr drafter, seven speculative tokens per step). Reasoning level is switchable between off / on (medium) / on through the chat template; deployment targets are 1x DGX Spark (GB10), 1x RTX PRO 6000, 1x H200, 2x H100 or 1x GB200, running on vLLM (official nightly image with --speculative_config.method dflash, mamba-backend flashinfer, video-pruning-rate 0.5). The NVFP4 weights measure **75.0GB** in total across 88 safetensors shards. Benchmarks (NVFP4 / BF16): GPQA-Diamond 89.90 / 89.52, Terminal-Bench 2.1 62.78 / 60.67, SWE-bench Verified 71.07 / 71.10, SWE-bench Pro 51.30 / 51.78, tau2-bench retail 69.30 / 69.24, tau2-bench airline 75.75 / 76.50, Video-MME v2 32.54 / 32.85, OSWorld-Verified 52.3 / 51.14, AA-Omniscience index 5.3 / 4.03, GDPval-AA v2 1206 / 1211. License is OpenMDW-1.1 (commercial use permitted). **Cataloged with zero on-site runs**: under the site's conservative padded VRAM accounting a 120B-class NVFP4 model needs roughly 180GB+, so 96GB-class cards do not get a fit verdict.

120BOpen weights

Arena Elo —

no indexed quantization releases yetnot benchmarked yet

View

elephant-1

Baseten (Base Labs)

The 9B sibling of mammoth-1 (released 2026-10-21, MIT, a full fine-tune of Qwen3.5-9B) for fast, cheap decisions: same single shared-prefix prefill returning calibrated probabilities for every option. Scores: best model under 15B on the Jev Decision Index 0.4 (62.65, tied #8 of 208) and 65.2 on a replica of the Artificial Analysis Decision Index, level with d1-ultra (65). Division of labor with mammoth-1: reach for the 27B when broader general knowledge matters, the 9B when latency and cost drive routing and guardrail duty. Self-hosting requires the official setup.sh with the OneJev runtime (a single CUDA GPU).

9BOpen source

Arena Elo —

no indexed quantization releases yet39 benchmark runs

View

mammoth-1

Baseten (Base Labs)

Baseten's 27B decision model released 2026-10-21 (MIT, a full fine-tune of Qwen3.5-27B with the vision encoder frozen). Give it a state and a set of questions and it returns a calibrated probability for every option from one shared-prefix prefill — cheaper and steadier than asking a general LLM question-by-question for routing, guardrail and classification work. Scores: #1 of 208 models on the Jev Decision Index 0.4 (71.24, ahead of Perplexity Decider v1.2 at 69.42) and #1 on a replica of the Artificial Analysis Decision Index (65.9). Full-attention layers use a block-bidirectional mask; max sequence length 16,384 tokens. Note: pipeline is text-classification and self-hosting requires the official setup.sh with the OneJev runtime (a single CUDA GPU) — this is not a general chat model.

27BOpen source

Arena Elo —

no indexed quantization releases yet26 benchmark runs

View

Raon-1-mini-preview

KRAFTON

The first open model in KRAFTON AI's Raon family, trained from scratch (preview release 2026-10-22, Apache-2.0, official size designation 30B-A4B): 28.8B total parameters, 4.3B active (6 of 128 routed experts plus 1 shared). Hybrid attention is the headline — 18 of 25 layers use Raon Delta Attention (linear attention that decouples the erase and write gates of Kimi Delta Attention, with a fixed-size state) and 7 use Gated MLA (KV cache of only ~8KiB per token in BF16) in a 3:1 layout; 262,144-token context. Multilingual Korean/English/Japanese/Chinese/Spanish/French/Thai with native tool calling (Hermes Agent, Pi and OpenCode already integrated). Official quantizations are MXFP8 and NVFP4, and **NVFP4 runs on a single NVIDIA RTX 5090 (32GB)**; a DFlash2 draft model ships alongside (up to 2.18x higher output throughput measured on a B300). Weight repos are request-to-access with automatic approval; official GGUFs: Q4_K_M 19.7GB, Q6_K 25.8GB, Q8_0 30.7GB. Note: preview release — specs and results may change before the final version.

28.8BOpen source

Arena Elo —

no indexed quantization releases yet24 benchmark runs

View

Flint-40B-A2.5B

Recursive

Recursive's coding-agent MoE released 2026-10-21 (Apache-2.0): 39.8B total parameters, about 2.5B active per token (including the output head; the body alone is ~2.2B — the official size designation is 40B-A2.5B). Architecture: 49 layers (2 dense + 47 MoE), 256 routed experts with 6 selected per token plus an always-active shared expert; attention is 35 sliding-window layers (513-token window) plus 14 full-attention layers, RoPE on sliding-window layers and no positional encoding on full-attention ones, with per-head QK RMSNorm and attention sinks. 262,144-token context, 150,144 vocabulary, knowledge cutoff June 2026. Release evaluations: Terminal-Bench 2.1 54.9%, SWE-bench Pro v2 89.7%, SWE-bench Verified 74.2% (Recursive-measured, temperature 1.0, average pass@1 over at least five attempts). BF16 weights are about 80GB; community GGUFs (unsloth) run UD-Q4_K_XL at 25.1GB, Q8_0 at 42.3GB and UD-IQ1_M at 11.4GB. TechCrunch and WIRED report the model was largely built by AI in under four months.

39.8BOpen source

Arena Elo —

no indexed quantization releases yet22 benchmark runs

View

MiniMax-M3.1-Flash

MiniMax

MiniMax's new Flash-series flagship (Oct 19, API-first with weights to follow). More aggressive sparsity than the predecessor: official claims ~3x faster online inference and 70% lower task cost ($0.15 input / $0.60 output / $0.015 cached per 1M tokens), with 200+ tok/s output (P50 170+). Benchmarks: BankerToolBench 76.2, DeepSWE v1.1 73.7%, OSWorld V2 61.7, Agents Last Exam 29.6 (first among Chinese models on both), CyberGym 82%. Ran anonymously as "Space Bunny" processing 200 trillion tokens. Available now via MiniMax API, M Plan and MiniMax Code; model weights promised later — parameter count unpublished, local runs await the weights.

Restricted

Arena Elo —

no indexed quantization releases yetnot benchmarked yet

View

Hy-Memory-30B-A3B

Tencent Hunyuan

Tencent Hunyuan long-term-memory writer (Oct 15, Apache 2.0), 30B total / ~3B active. Not a general chat model: it slots into memory frameworks like Mem0 / MemOS / Hy-Memory to extract, merge and update long-term memory entries over conversation streams. Trained with OPD + RL; paper arXiv 2610.22196. FP8 official weights and community GGUF conversions are out.

30BOpen source

Arena Elo —

no indexed quantization releases yet31 benchmark runs

View

Beam-1.0-501B-A23B

Reflection

Reflection's 501B-total / 23B-active MoE reasoning model (Apache 2.0) built around a simple claim: about 1/3 the inference-time reasoning tokens of comparable models. 1M context, BF16 / FP8 / NVFP4 weights; community GGUF quants and speculative-decoding backends like DFlash2 have been landing since Oct 16.

501BOpen source

Arena Elo —

no indexed quantization releases yetnot benchmarked yet

View

DeepSeek-V4.1-Flash-1017

DeepSeek

Checkpoint refresh of DeepSeek-V4.1-Flash released Oct 17 (MIT). Broad benchmark gains over the base model: HLE w/o tools 36.8→42.6, Terminal-Bench 31.2→42.3, DeepSWE 74.2→77.0, Agents Last Exam 31.8→35.4; shorter outputs mean faster effective speed and lower cost. 1M context, 37B active parameters, same architecture and deployment path as the original.

763.2BOpen source

Arena Elo —

no indexed quantization releases yetnot benchmarked yet

View

57 models