OSS models MiniMax M3.1 Flash weights public (MiniMax Community License) · Tencent WeVisOne-VL · OlmoEarth v1.3 · Phonon
OSS models this window: MiniMax M3.1 Flash weights went public late Oct 22 ET, resolving the Watch item carried since Issue 42. Tencent posted WeVisOne-VL 2B/4B on HF (no newsroom post). AI2 shipped OlmoEarth v1.3. Gradium’s Phonon TTS checkpoints back Ollama 0.40.9.
Repo MiniMaxAI/MiniMax-M3.1-Flash created Oct 17 (private); single commit “Init” at 2026-10-23T03:37Z (≈11:37 PM ET Oct 22); not gated; ~252 GB, 72 files; 1,096 downloads / 186 likes at our check. Card: “native multimodal mixture-of-experts model for coding, agentic workflows, computer use, and cybersecurity”; text/image/video in; “Approximately 428B total; 23B activated per token”; “Up to 1,000,000 tokens”; MiniMax Sparse Attention throughout; quantization-aware training (routed experts NVFP4, attention/dense MXFP8, router FP32); built-in DSpark speculative-decoding draft. Benchmarks (vendor-reported; MiniMax harness notes list Claude Code / Codex / mini-SWE-agent versions): BankerToolBench 76.2, DeepSWE v1.1 73.7, OSWorld 2.0 61.7, Agents’ Last Exam 29.6, CyberGym 82.2, Terminal Bench 4.0 38.1 (vs 51.8 Opus 5, 39.9 GPT-5.6 Sol, 42.3 DeepSeek V4.1 Flash-1017 in MiniMax’s table). “Results are for the official weights; third-party quantizations may differ.” Recommended sampling temp 1.0 / top_p 0.95; reasoning_effort low → max. Serving: SGLang cookbook, vLLM recipe (TP4 on Blackwell B200/B300/GB200/GB300 or AMD MI355X; TP8 on H100/H200), TokenSpeed; Transformers / KTransformers / Unsloth / ATOM “in progress.” Contradictions: card 1M context vs Ollama cloud listing 976K; MiniMax release notes called it “open-weight” Oct 19 but weights were not public until Oct 22–23.
Grant: permission “to deal in the Software for non-commercial purposes.” Then: “If the Software (or any derivative works thereof) is used for any Commercial Use”: (1) “prominently display ‘Built with MiniMax M3.1-Flash’” on a website, UI, blog post, about page or docs; (2) “obtain a separate, prior written authorization from MiniMax by contacting api@minimax.io … if such products and services generate more than 20 million US dollars … in yearly revenue; otherwise, you only need to send a one-time notice to api@minimax.io.” “Commercial Use” includes offering paid products/services that rely on the model or derivatives and commercial use of APIs for it (hosted inference counts), and deploying derivatives such as fine-tunes. Prohibited uses include to “assist with, engage in or otherwise support any military purpose.” Not an OSI open-source license. Not legal advice.
unsloth/MiniMax-M3.1-Flash (≈252 GB mirror) and unsloth/MiniMax-M3.1-Flash-GGUF (created ~12:20 AM ET Oct 23; 0 files at check). tencent/WeVisOne-VL-2B / -4B (created Oct 22 ~5:48 AM ET; Apache 2.0; built on Qwen3-VL; detection, grounding, OCR, segmentation and matting with a pixel decoder; card targets research use). Tencent newsroom has no OSS post. AI2 OlmoEarth v1.3 (HF blog Oct 22 12:11 PM ET): Earth-observation embedding models. Gradium Phonon: gradium/phonon-31m-en (Apache 2.0, 31M params, 24 kHz, built on Kyutai Pocket TTS + Mimi codec; runtime xn-ptts, PyPI ptts 0.4.1 Oct 21); the 40M en/fr/de/es/pt/multilingual checkpoints are CC BY-NC 4.0.
RecursiveSI org now lists flint (BF16; created Oct 21 14:20Z), flint-int4, flint-int4-gguf, flint-nvfp4-gguf (created 13:51Z, updated through ~15:54Z). License apache-2.0 on the card; “Hosted API terms and data practices are separate from the weight license.” Issue 43 (8:40 AM ET Oct 21 check) saw “None public yet” — weights appeared about 1.5 hours later. Full specs and vendor claims on the Local page.
MiniMaxAI HF org newest model still MiniMax-Music3 (Aug 7). Ollama’s minimax-m3.1-flash:cloud free window ends Oct 22 11:59 PM UTC (7:59 PM ET). Tencent newsroom newest Oct 15 (consumer); tencent HF org: Hy-Memory-30B-A3B-GGUF card touched Oct 22. Granite Speech 5.0 details on the Local page.
Per NVIDIA’s Oct 21 GTC Berlin post: 40B-parameter MoE, 2.5B active, Apache 2.0, built with Recursive’s “Eureka Machine”; “over 220 tokens per second on a 64GB RTX Spark”; benchmark claims on SWE-Bench Pro v2, Terminal-Bench 2.1, SWE-bench Verified (vendor-reported; no model card to check). NeMo Automodel docs page and a NeMo RL flint-support branch are linked. huggingface.co/RecursiveSI: “None public yet” at ~8:40 AM ET Oct 21; HF API search found no Flint model from that org.
NVIDIA’s Oct 21 post: “MiniMax M3.1 Flash is a new open model built for NVIDIA DGX Station and DGX Spark.” MiniMaxAI HF org newest model still MiniMax-Music3 (Aug 7). Ollama: minimax-m3.1-flash:cloud “free until Oct 22, 2026 11:59 PM UTC” (7:59 PM ET). Tencent Hy-Memory-30B-A3B: GGUF (tencent, Oct 19) and mlx-community 4-bit/8-bit (Oct 20 ET). LiquidAI LFM2.5-24B-A2B-DSpark (+GGUF Oct 20) draft-model builds (title-level).
MiniMax model release notes entry dated Oct 19, 2026: “MiniMax M3.1 Flash — An ultra-efficient open-weight model for everyday work.” Ollama library minimax-m3.1-flash (cloud tag only): 428B parameters, 976K context, vision/tools/thinking; $0.15 in / $0.015 cached / $0.60 out per 1M tokens, free until Oct 22 7:59 PM ET; description claims “nearly three times faster” online inference and 70% lower usage cost (MiniMax-reported). Contradiction / watch: Hugging Face MiniMaxAI org newest model at our check is MiniMax-Music3 (Aug); no M3.1 Flash weights found. MiniMax’s model page URL for M3.1 currently serves the M2.5 page.
Tencent’s HF org added Hy-Memory-30B-A3B-GGUF Oct 19 07:52Z (≈3:52 AM ET), following the Apache 2.0 base and FP8 releases (Issue 41). Tencent newsroom still has no post about it.
Reflection blog newest is still “Beam is now available” (Oct 16); self-distillation recipes promised “in the coming days” not yet posted. Official Reflection HF repos unchanged since Oct 16; no mlx-community Beam build. Dory support PRs still open: vLLM #60149 (updated Oct 19), llama.cpp #30306 (Metal untested), mlx-lm #1988. NVIDIA’s Oct 15 blog dates Nemotron 3.5 Bolt E4B (4 GB) for Oct 23.
Weights public Fri Oct 16, 2026 (Reflection post “Beam is now available”; HF blog same day; AWS ML blog 11:23 AM ET). Announced Oct 5 with early access (“weights … later this month”); benchmark table updated Oct 8. Architecture (HF blog): 52 layers, hidden 5,120; 43 sliding-window (512) layers with RoPE + 9 global NoPE layers with power-law query scaling (1 + (pos+1)/2048) ** 0.1; GQA 48 Q / 8 KV heads, QK-RMSNorm, sigmoid output gate; first 3 layers dense SwiGLU, then 49 MoE layers × 256 experts (top-8, DeepSeek-V3-style group-limited sigmoid routing, no shared expert; experts ≈98.4% of params); scaled residuals 1/√k; logit tanh softcap ±30. KV cache ≈36 KiB/token BF16 (≈36 GiB at 1M per sequence vs 208 GiB if all layers were global). Pretrained on 23.8T tokens (Reflection). Checkpoints: BF16 (≈1.0 TB on HF), FP8 (Hopper/Ampere), NVFP4 (≈280 GB per HF blog; HF repo listing ≈294 GB incl. extras), DFlash2 drafter. Five reasoning levels (low/medium default/high/xhigh/max), preserve_thinking, tool-call grammar, response template in tokenizer (transformers>=5.20.1). Serving (HF): NVFP4 fits 2× B200 with little headroom, plan 4× B200; or 2× B300, 8× RTX Pro 6000; FP8 on 8× H200 (SGLang); BF16 vLLM TP=16 across two 8× H200 nodes with a “proposed” vLLM recipe — note vLLM v0.32.0 notes do not list Beam and pin transformers<5.20.0. LoRA SFT on one 8× B300 node ≈19,000 tok/s (TRL); GRPO via Miles (RadixArk). Harness: open-source Mirror CLI (Python frontend + Rust core); also tested with Pi, OpenCode, Hermes. Partners listed by Reflection include AWS (JumpStart us-east-1/us-east-2 + Trainium), Google Cloud, Microsoft Foundry, Databricks, Cloudflare, Ollama (cloud), LM Studio, OpenRouter, Together, Fireworks, Baseten; Dell and HPE on-prem. HF Inference Providers: Baseten, Fireworks, Together. Benchmarks Reflection-reported (BF16): SWE-bench Verified 81.6%, Terminal Bench v2.1 80.1%, SWE-Bench Pro v1 66.1%, GPQA Diamond 90.6%, AIME 2026 97.7%, MCP Atlas 80.8%; NVFP4 within ≈1–2 points on most rows. Reflection’s Oct 5 post says Kimi K3 “remain[s] ahead on raw capability” and claims 3–4× less inference compute than GLM-5.2 (its own FLOPs ≈ 2 × active params × generated tokens estimate, using Artificial Analysis/DataCurve data). License Apache 2.0; model card: Reflection “does not prohibit distillation, use of outputs to train other models, fine-tuning, parameter-efficient adaptation, quantization, merging”; Usage Policy still applies. Safety: reflection.ai/beam-safety-report. HF repos not gated at check time. Reflection offers $10K Hub credits for best Beam projects (HF blog).
Unsloth (a Reflection-named launch partner) GGUF repo, sizes summed from the HF file listing at check time: UD-IQ1_S ≈103 GB, UD-IQ1_M ≈124 GB, UD-IQ2_XXS ≈134 GB, UD-IQ2_M ≈165 GB, UD-Q2_K_XL ≈187 GB (5 shards), UD-IQ3_XXS ≈197 GB, UD-Q3_K_XL ≈242 GB, UD-Q4_K_XL ≈307 GB, Q8_0 ≈533 GB. Repo publishes KLD and top-1 agreement charts vs disk size, not task scores. Runtime: llama.cpp #30670 (Beam converter, graph, harmony-xml tool-call parser) merged Oct 16 22:07Z (6:07 PM ET), released as b11701 (7:20 PM ET); PR body reports real-weight conversion and generation tested on CUDA. No Apple Silicon numbers published by any primary. macOS gives Metal a recommended working set below physical RAM (SGLang #36780 measured 37.44 GiB on a 48 GB M4 Pro), so on 256 GB the 2-bit build is the realistic fit once KV is added. Ollama library beam: cloud tag only. No mlx-lm PR or mlx-community Beam quant as of 9 AM ET Oct 19. Other community quants: DevQuasar and AtomicChat GGUF, diffbot EXL3 3bpw for 2× RTX PRO 6000.
Release date Oct 16, 2026 (model card; HF repos created ≈12:25–2:30 PM ET). Looped (latent-recurrent) LM on a normalized-GPT backbone: Input 8 layers, shared Recurrent block 20 layers, Output 8 layers; input reinjected each loop; interleaved SWA + full attention. Loop count set at inference (n_recurrent_loops in vLLM --hf-overrides; dory.num_loops in llama.cpp); NVIDIA suggests 1 loop for minimum compute and 4 for agentic tasks. Last-loop KV caching keeps one cache per recurrent attention layer (no KV growth with loops). Pre-trained natively in NVFP4, then BF16. Formats: BF16, MXFP8, NVFP4, GGUF Q5_K_M (with SHA-256). License OpenMDW-1.1. Runtime: vLLM PR #60149 (install from PR), llama.cpp #30306 (card pins an NVIDIA fork), mlx-lm #1988 — all unmerged at check time. Benchmarks and token-reduction plots NVIDIA-reported; AA v4.3.2 figures are NVIDIA internal estimates.
HF repo created Oct 15 (FP8 variant Oct 16). ~3B-active MoE trained to be the memory writer inside external memory frameworks (Mem0, MemOS, Hy-Memory): decides what to store, how precisely, and how new evidence updates old entries. Two-stage training (on-policy distillation, then RL with chunk-level credit) inside the deployed harnesses. Card evaluates on fixed 20% subsamples of LoCoMo, LongMemEval, PersonaMem-v2, Travel/Shopping, Code Memory Bench Tencent-reported. Serve with vLLM (tested v0.31.0). Apache 2.0. Tech report arXiv 2610.22196. Not posted on the Tencent newsroom.
NVIDIA blog Oct 15, 2026. Super VL: 120B total / 12B active; interleaved Mamba-2, MoE, and attention with C-RADIO vision encoder; MTP head for native speculative decoding; up to 1M context; 100B-token Indic continued pre-training, 35B-token 1M-context extension, 67B tokens SWE mid-training at 256K. Repos: nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16, -NVFP4, -Base-BF16, -NVFP4-LiLiCorr. License OpenMDW-1.1; open datasets and recipes. Card lists deployment on 4× H100 or 2× GB200 and a vLLM nightly or vllm/vllm-openai:v0.32.0 image — no v0.32.0 GitHub release exists as of Oct 16 8:20 AM ET (latest v0.31.0). Claims “5x faster than the next best” on the Artificial Analysis index NVIDIA-reported / vendor-cited. Bolt 74B-A9B 24GB (native 2-bit, GGUF + LiLiCorr drafts) available; Bolt E4B 4GB due Oct 23. Also Oct 16: Nemotron-3.5 Embed 1B/8B and Rerank 8B repos posted (not yet announced). Security note: vLLM GHSA-mm38-5g96-7j6x (Oct 9, High) — Nemotron-VL processors disabled Pillow’s decompression-bomb guard; patched ≥0.25.0.
HF blog Oct 15, 2026 (repos created Oct 13). 5.6B image transformer; text encoder Qwen3-VL-4B-Instruct; separate 9B prompt upsampler black-forest-labs/flux-3-image-klein-pup (pipeline expects full scene captions). Checkpoints: flux-3-image-klein (timestep-distilled, 16 steps, no CFG) and -base (50 steps, CFG; for fine-tuning/LoRA). Up to 10 reference images; bounding-box layout control on a 0–1000 grid via JSON regions after the caption. HF timings (H100, distilled): 1024² text-to-image 1.93 s, 1 reference 3.87 s, peak 23.8 GiB; NVFP4 ≈13.4–15.3 GiB on RTX 6000; NVFP4/FP8 via from_single_file; Modular Diffusers pipeline; LoRA scripts. No Apple Silicon numbers published. License: FLUX Kommunity License v1.0 — non-commercial use, plus commercial output generation by Qualifying Users under US$5M gross annualized revenue (incl. affiliates); the card says this is “not blanket permission for commercial fine-tuning or distribution.” HF gate: auto.
HF blog Oct 15, 2026. Granite models compiled with activated-LoRA adapters (a general decision adapter, specialist granitelib adapters) plus a co-located Granite Speech ASR encoder, served from one endpoint; adapters reuse the base model’s KV cache. IBM claims its decision adapter beats Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions, and Clef-Flash on average on safety-related decision tasks IBM-reported. Adapter SDK “planned.”
Blog published 2026-10-08; model repos created Oct 6–7 (lightonai/LightOnOCR-3-0.8B, -1B, -4B). License Apache 2.0 (model card). The 1B keeps the LightOnOCR-2-1B architecture; 0.8B and 4B adopt the Qwen3.5 vision-language architecture (load with Qwen3_5ForConditionalGeneration; Transformers ≥5.2, saved with 5.5.4; thinking disabled by default). Two modes: empty prompt = transcription; prompt grounding = labeled bounding boxes (coordinates normalized 0–1000), image descriptions, and chart data as HTML tables. LightOn reports 9–14% fewer output tokens in grounding mode than Chandra-OCR-2 and Infinity-Parser2-Pro on 512 olmOCR-bench pages; training used SFT plus RL with manually verified data (LightOn-reported). Serving: Transformers and vllm serve; client/CLI/viewer and benchmark code in the LightOnOCR GitHub repo.
Repo created 2026-10-09T06:42Z ≈ 2:42 AM ET (last modified 09:43Z). Safetensors total ≈ 5.33B parameters; tags include custom_code, youtu_vita, document-parsing, OCR, audio, video. License: “other” (youtu-parsing LICENSE), not Apache; read it before commercial use. Inputs: document page, natural image, chart, flowchart, geometry figure, audio clip, audio-visual video; output family selected by --task prompt, covering perception (layout with bbox, text/LaTeX/OTSL tables, Mermaid flowcharts, ASR, acoustic events, camera motion) and cognition (captions, reports). Card News: “[2026-10] Technical report, model weights, vLLM plugin and inference examples released”; evaluation code “TBD”. Code at TencentCloudADP/youtu-parsing. Results on the card are Tencent-reported. Needs trust_remote_code, so pin a revision.
HF blog Oct 7 (16:54Z ≈ 12:54 PM ET). d1-3B from LFM2.5-VL-3B (decoder-only; text + image). d1-omni-600M (experimental, early research release) from LFM2.5-Encoder-350M with vision and audio encoders (text + image, or text + audio). Single forward pass, no generated tokens. Liquid-reported: Decision Index 0.2.1 d1-3B 48.57 vs Decider 35B-A3B 47.11; seven-dataset mean d1-3B 82.9, d1-omni-600M 78.4, Decider 4B 81.1, Decider 2B 77.1. Latency with NVIDIA (vendor-reported): Apple M5 Pro 30 ms one question / 41 ms three / 640 ms 3.4K-token state / 62 ms 384px image / 78 states/s packed; Jetson AGX Thor 16 ms; Orin Nano 50 ms; RTX 4090 8 ms; AMD MI325X 9 ms. No speed numbers for omni. Requires transformers>=5.14 and trust_remote_code=True; sample picks CUDA, MPS or CPU. HF license: LFM 1.0 (“other”), not Apache. Also on HF: d1-3B-w8a8 and d1-omni-600M-GGUF.
JetBrains AI blog Oct 8 (12:00Z ≈ 8 AM ET). Same architecture as Mellum2 (open-sourced in June); nearly all new work is post-training, mainly RL “with millions of sandboxed runs across thousands of environments”, plus filtered open and in-house RL tasks (math, competitive programming, science, tool use, SWE). Compared with Mellum2, Qwen3.5-9B and Gemma 4 E4B on one setup; largest gain in agentic coding. JetBrains-reported: “almost twice as many tokens as Qwen3.5-9B” under heavy load; MTP makes single requests “about 1.6 times faster”. HF repo: JetBrains/Mellum2.1-12B-A2.5B-Thinking (Apache 2.0). Discrepancy: the blog says GGUF builds for llama.cpp, Ollama and LM Studio and the vLLM MTP head are “coming soon”, but a JetBrains GGUF repo already exists on HF (created Oct 7 18:43Z). Treat it as early until JetBrains announces it.
HF blog Oct 7 (12:45Z ≈ 8:45 AM ET). IOI 2026: Nemotron-3-Ultra-CC (550B total / 55B active; SFT + GenCorrect generate-evaluate-refine) scored 535.4/600 vs a 361.12 gold threshold and a top human score of 498.27, in a live run under contest constraints that was unofficial and not in the official ranking. IMO 2026: Nemotron 3 Ultra general + SFT + RL checkpoints in a generate-verify-refine system scored 30/42 vs gold 29, graded by official IMO graders, natural language only. Nemotron-3-Nano-CC (30B / 3B active) on IOI 2025: 130 → 280 (SFT) → 291 (RL) → 468 (GenCorrect). One SFT epoch on Ultra beat fully post-trained Nano. Data: 22,000 curated competitive-programming problems; IMO SFT 414,890 examples over 15,818 proof problems; RL on 9,597 problems. Open: Nemotron Labs IMO 2026 collection (SFT/RL checkpoints, both datasets, Nemotron-IMO-Bench with 200 problems), Nemotron-3-Ultra-CC, NeMo-Skills pipelines. NVIDIA-reported.
Launch 2026-10-06T16:00Z ≈ 12 PM ET (blog.google; DeepMind blog 3:57 PM ET). 740M total: 130M transformer backbone + 140M embedder (the 270M text model), plus 170M vision and 300M audio encoders that load optionally. 24 layers, model dim 512, 5:1 local:global attention, 1,024-token sliding window, 262,144 vocab, mean pooling, 512→768 projection. Output 768-d, MRL truncation to 512 / 256 / 128. 8,192-token context (“up to 5.5 minutes of audio, 29 images, 58 video frames”). 100+ languages. Task prefixes such as task: search result | query: {q} and title: {t} | text: {c}. Card benches at 768-d (Google-reported): MTEB multilingual v2 61.36 (EG1 61.15); MTEB code v1 78.68 (EG1 68.76); MIEB lite 64.64; MMEB v2 image 57.28 / VisDoc 67.84 / video 50.67; MSEB retrieval 69.54; MAEB 49.39. At 256-d: MTEB multilingual 60.41, code 76.18. On-device: ~191 MB active RAM text-only and ~567 MB full multimodal, quantized, on a Pixel 11 Pro (Google-reported). Contradiction: the card and launch blog say MRL gives “up to 6x” storage reduction, while the Google AI Edge post says “up to 8x”. Day-0 integrations listed by Google: Transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, LM Studio, Qdrant, Unsloth; Kaggle; LiteRT community builds; Model Garden “coming soon”.
Repo created 2026-10-06T16:08Z, auto-converted. Files from the HF API: embeddinggemma-2-BF16.gguf ~558 MB, -Q8_0.gguf ~310 MB, mmproj-…-BF16.gguf ~982 MB, mmproj-…-Q8_0.gguf ~555 MB. Run with llama serve -hf ggml-org/embeddinggemma-2-GGUF; llama.cpp’s /v1/embeddings has accepted typed media since v0.6.0.
Blog Oct 6. Before Kolibri mid-training, Aleph Alpha scanned 8.3 billion mid-training documents against its whole eval suite and removed 1.5 million that quoted a benchmark item nearly verbatim. Many held benchmark training problems, not test items. Reinserting them in a proxy run inflated HumanEval by 45 points and MMLU by 28, “and the gain reached even the questions that never leaked.” The detector scored at most 60% of leaked MMLU questions and 13% of leaked TriviaQA questions above threshold; the rest were removed only because they shared files. Pre-training was not decontaminated, and the model still recited HumanEval solutions learned there. Author-reported.
Repo created 2026-10-03T16:24Z ≈ 12:24 PM ET Sat Oct 3. Card: 320B MoE with 18B active, 288 routed + 1 shared expert per layer, mHC + DSA, one MTP layer. Quants: Q4_K and Q2_K, with non-expert tensors kept at Q8_0; Q8_0 mmproj for vision; MTP sidecar files; no imatrix. Run with llama serve -hf ggml-org/GLM-5.3-Flash-GGUF (needs llama.cpp with #27773, in v0.6.0). The card does not list file sizes, so check the repo before pulling onto a 256 GB Studio. The Bedrock-hosted GLM 5.3 (753B) is a different, larger model; see Enterprise.
Aleph Alpha blog dated 03/10/2026 (Oct 3, “On the Day of German Reunification”); HF repo created 2026-10-02T05:54Z, card “Release Date: 3rd of October 2026.” German/English MoE reasoning model with tool calling. Card: 78,103,074,560 total / 3,457,573,120 active params; 50 layers, 4:1 SWA:GQA, 384 experts/layer (1 shared + 6 routed), Muon. Context: native 262,144; validated to 1,048,576; recommends ≤262k for serving. Weights float8_e4m3fn 128×128 blocks (~78 GB); BF16 repo also published. Min hardware listed: 2× A100 80 GB / 2× H100 / 1× H200 / B200 / B300. Pre-training 20T tokens (~62.5% EN, ~23.9% DE, ~13.6% code) on 768 B200s. Apache-2.0; signatory of EU GPAI Code of Practice. Pareto / benchmark claims vendor-reported. mlx-lm support PR open; llama.cpp support not in tip notes as of b11417.
HF blog 2026-10-02T15:19Z ≈ 11:19 AM ET Fri Oct 2. Open model that turns a research question + retrieved literature excerpts into a cited report. Base Qwen/Qwen3-8B; SFT checkpoint allenai/AstaBrief_8B_SFT (data AstaBrief_SFT_Mix, reports generated by Claude 3.5/3.7 Sonnet, o3, o4-mini, GPT-4.1 via Asta ScholarQA); final allenai/AstaBrief_8B. Both Apache-2.0. Powers Asta “Fast mode”; across the full Asta pipeline Fast mode averages 51.1 s vs 178.5 s for Claude-powered Thinking mode (~3.5×, Ai2-reported). Ai2 notes most training/eval ran in 2025 against the frontier of that time and was not rerun. Example workflow for local reports from your own PDFs released. Card recommends the published SFT prompt format.
Research post September 27, 2026; HF createdAt 2026-09-27T14:14:32Z ≈ Sep 27 10:14 AM ET. 309B total / 15.5B active MoE; native 1M context; hybrid SWA (128) + DSA (top-2048) with GQA4; built on MiMo-V2.5 base. MIT license. Card: FP8-capable NVIDIA GPUs; weights ~315 GB (+ headroom). Deploy example pins transformers[torch,kernels]>=5.17.0. NaiveRT claims (vendor): up to ~2,000 tok/s Ultrafast / ~50 tok/s Standard; open-source NaiveRT “available by Oct 12th” per research post — treat runtime as not yet GA. Not Apple Silicon / MLX.
HF API sample ~8:35 AM ET Sep 23: mlx-community/Qwen3.6-35B-A3B-DFlash2-4bit created ~6:06 AM ET Sep 23. Community upload for Apple Silicon try-this under current mlx-vlm 0.7.2 / mlx-lm — verify license + architecture support before recommending. Not Apple/HF library GA. Skip MiMo-V2.6 rehash (Issue 22).
