New open-weight LLMs, September to November 2026: 21 releases and the hardware each needs

Between 1 September and 7 November 2026, 21 open-weight language models that I could verify on Hugging Face were released, from Liquid's 1.7B LFM3 to Moonshot's 2.8-trillion-parameter Kimi K3.1 Pro (weights uploaded 2 November). Only five have a build small enough for a 32 GB GPU, two more (Nemotron 3.5 Super VL and Ripple-65B-Preview) fit one 128 GB DGX Spark at 4 bits, and the rest need two Sparks or a multi-GPU server. Thinking Machines' Inkling 2 Small (announced 5 November) still has closed weights, and I found no Hugging Face repo for Meituan's LongCat-2.5. Everything below was checked on 7 November 2026.

All rows checked on Hugging Face on 7 November 2026. "Weights public" is the repo creation date from the Hugging Face API, except where the public release came later: Muse Spark Open's date is Meta's 15 October announcement and license date (the gated repo was created in September), Mistral Large 4's weight files were committed on 29 October (the repo dates from 30 September), MiniMax-M3.1-Flash's on 23 October after a 19 October announcement, and Kimi K3.1 Pro's on 2 November. Parameter counts and context are from each model card. Sizes are sums of the weight files in each repo's listing, in GB (10⁹ bytes). "Estimate" rows have no 4-bit build in the repos I checked, so I multiplied total parameters by 4.5 bits (4-bit weights plus scales) and divided by 8. GPUwerk has not run any of these models.

How did I decide what runs where?

I compared each model's weight files with three memory budgets, leaving room for the KV cache and the runtime. A Spark has 128 GB of unified memory, of which about 121.6 GiB (130.6 GB) is addressable for a model, as the guide to what fits in 128 GB explains. My working lines:

• 32 GB GPU (RTX 5090): weights up to about 28 GB, so 4 GB stays free for context.
• One Spark: weights up to about 120 GB, which is 130.6 GB minus 10 GB for KV cache and CUDA context.
• Two Sparks: up to about 250 GB across both nodes, with a runtime that splits the model.

The estimate formula for models without a 4-bit build is: total parameters (billions) × 4.5 ÷ 8 = GB. For K2-Horizon that is 375 × 4.5 ÷ 8 = 211 GB. Real GGUF files land close to this: Unsloth's UD-Q4_K_XL of Muse Spark Open is 505.5 GB against 459 GB from the formula, because dynamic quants keep some layers at higher precision. If you want to check a different context length, the DGX Spark memory calculator adds the KV cache for you.

Which new open models fit a 32 GB GPU?

Five releases have a build under 28 GB. JetBrains' Mellum2.1 is a 12B MoE with 2.5B active and a 131,072-token context; its 24.3 GB BF16 checkpoint fits as is. Nanbeige4.3-33B is a dense 33B with a 524,288-token context, and the community Q4_K_M file is 20.8 GB. Arete-27B is Qwen3.8-27B after supervised fine-tuning on about 52.8 million tokens, which the card says was aimed at removing state-propaganda framing; its Q4_K_M is 16.5 GB, the same as the base model.

NVIDIA's Nemotron 3.5 Bolt is the unusual one: a 74B model with 9B active that NVIDIA ships in a "24GB" build with native 2-bit quantization, which the card says is sized for 24 GB of memory. At 24.1 GB it leaves about 8 GB on a 32 GB card. The LFM3 family from Liquid AI is tiny by comparison: LFM3-1.7B is 3.5 GB in BF16, and the card puts it at about 1 GB at 4 bits.

The rest are out of reach for one consumer card, even with heavy quantization.

Which new releases fit one 128 GB DGX Spark?

Nemotron 3.5 Super VL and Ripple-65B-Preview are the two notable new models that fit one Spark at 4 bits with context to spare. They will run at very different speeds, because decode speed on a Spark is set by how many bytes each token reads from 273 GB/s memory.

My calculation: ceiling = 273 GB/s ÷ bytes read per token. Real decode lands below this ceiling, and speculative decoding can push past it on predictable text. These are not measurements.

The arithmetic is blunt. Ripple's card names the DGX Spark as hardware the model fits on, and it does, but a dense 65B model reads all of its weights for every token, so a single user should expect under 5 tok/s before speculative decoding (Ricursive also publishes a DFlash2 draft model for that). Nemotron 3.5 Super VL reads about a tenth as much per token. The site's Nemotron 3.5 Super VL page has the file sizes and NVIDIA's one-Spark vLLM settings.

Two late-August releases sit just outside this roundup and are covered on their own pages: Qwen3.8-27B fits a Spark at every precision, and GLM-5.3-Flash fits only below 2 bits.

What needs two Sparks or more?

Every new model above 300B total parameters needs at least two Sparks at 2 to 3 bits, and at 4 bits about half of them need the memory of four Sparks or more. The mid-size group is where a two-node cluster helps:

• MiMo-V2.6-Flash (309B, 15B active): the ggml-org MXFP4 GGUF is 167.4 GB, comfortable across two Sparks. Its Q2_K file (126.4 GB) technically loads on one but leaves almost nothing for context.
• MiniMax-M3.1-Flash (428B, 23B active): Unsloth's UD-Q3_K_XL is 194.0 GB. The 4-bit UD-Q4_K_XL at 249.4 GB sits right on my two-Spark line, so I would not plan on it.
• Beam 1.0 (501B, 23B active), Reflection AI's Apache 2.0 model: two Sparks only at UD-Q2_K_XL (186.9 GB).
• K2-Horizon, Intern-S2 and ERNIE-5.5-Flash-Base: no 4-bit build in the repos I checked, but my estimates (211, 223 and 158 GB) put them in two-Spark range once someone quantizes them.

The frontier-sized releases are a different class. DeepSeek-V4.1-Flash is 429.0 GB even in NVIDIA's NVFP4 build. Muse Spark Open's NVFP4 is 533.0 GB, Mistral Large 4's is 610.0 GB, and Kimi K3.1 Pro ships natively in MXFP4 (quantization-aware trained, per the card) at 1,561.3 GB. Sub-2-bit GGUFs bring DeepSeek, Muse and Mistral down to 175 to 233 GB, which loads on two Sparks with a real quality cost. How two nodes split a model, and why it does not make one user faster, is in the two-node cluster guide; the planned guide on large MoE models on multiple Sparks goes further.

Which reported releases are not open weights yet?

Three names circulating in release roundups did not check out as open-weight LLMs from this window on 7 November 2026.

• Inkling 2 Small (Thinking Machines, announced 5 November): the launch post says it is available through Tinker "with closed weights" and that the lab intends to publish Apache 2.0 weights the following week, pending safety evaluations. The post gives 594B total and 17B active parameters. Artificial Analysis lists it as not open weights. The earlier Inkling and Inkling-Small (July 2026, Apache 2.0) are open.
• LongCat-2.5 (Meituan): no repository in the meituan-longcat Hugging Face organization. The newest LongCat language model there is LongCat-2.0 from July.
• Qwen3.6-27B: open and Apache 2.0, but its repo dates from 21 April 2026. The current dense 27B from Qwen is Qwen3.8-27B (August).

Two others were announced before their weights arrived. MiniMax's 19 October post said the weights would follow later, and the weight files landed on 23 October. Moonshot's Kimi K3.1 Pro blog promised weights by 4 November; they arrived on 2 November.

Which licenses come with the new weights?

Twelve of the 21 releases use MIT or Apache 2.0 and two use the permissive OpenMDW-1.1; the other seven add conditions, and one excludes Europe. The full comparison is in the planned article on open-weight model licenses; the points that change a deployment decision:

• ERNIE-5.5-Flash-Base: the license grants rights only in a territory that excludes the European Union, the United Kingdom and South Korea, and that covers outputs as well. For an EU company, this model is off the table.
• LFM3: the LFM Open License v1.0 allows commercial use only for entities with under US$10 million in annual revenue.
• MiniMax-M3.1-Flash: commercial use needs a "Built with MiniMax M3.1-Flash" notice, and above US$20 million in yearly revenue, prior written permission.
• Kimi K3.1 Pro, Muse Spark Open, Mistral Large 4, Step-5-Preview: all add clauses aimed at companies that resell the model as an API (Model-as-a-Service), with revenue triggers from none (Mistral) to US$20 million (Kimi, Muse, StepFun).
• Nemotron 3.5: OpenMDW-1.1, a Linux Foundation license with no use restrictions and no conditions on outputs.

How good are the new releases?

Kimi K3.1 Pro is the strongest open model by a wide margin on the independent index I checked, and nothing that fits one Spark comes close to it. On the Artificial Analysis Intelligence Index v5.0.1, read 7 November 2026, Kimi K3.1 Pro scores 46, DeepSeek V4.1 Flash 1017 33, MiMo-V2.6-Pro 30, Ling 3.1 Flash (1019 update) 30, MiniMax-M3.1-Flash 29, Step 5 Preview 29, Mistral Large 4 26 and Muse Spark Open 25. Nemotron 3.5 Super VL, the best new model that fits one Spark at 4 bits, scores 14. The best closed model, Google's Gemini 4 Carbon, scores 60. Rankings, arena scores and the gap to closed models are in the open-weight leaderboard snapshot.

Vendor tables on the cards use newer agentic benchmarks (Terminal-Bench 4.0, DeepSWE v1.1, Automation-Bench) and are self-reported. DeepSeek's own card puts the 1017 update at 42.3 on Terminal-Bench 4.0 against 31.2 for the September release, for example, but no independent run of either was on the Terminal-Bench leaderboard when I read it.

Which new release should you try first?

It depends on the box you have, and for most people the answer is a mid-size model, not the headline one.

• 32 GB GPU: Nanbeige4.3-33B at Q4_K_M if you want a general dense model, Mellum2.1 for code completion and small agent jobs.
• One Spark: Nemotron 3.5 Super VL in NVFP4. It is multimodal, has a 1M context, uses the permissive OpenMDW-1.1 license, and its 12B active parameters keep decode usable. Ripple-65B is interesting but too slow without a draft model.
• Two Sparks: MiMo-V2.6-Flash at MXFP4. It is MIT-licensed, scores 27 on the AA index, and fits with room for context.
• A GPU server: Kimi K3.1 Pro if the license works for you, DeepSeek-V4.1-Flash-1017 if you need MIT.

Before you download 200 GB, run the five checks for a new model on GB10: file size, arm64 engine image, the parsers the card names, a short-context first start, and a real request. On GPUwerk, a Spark rents at $0.79/hour and two cabled Sparks at $1.79/hour, both billed per minute (see pricing).

FAQ

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

• moonshotai/Kimi-K3.1-Pro model card, Hugging Face, 2026-11-05
• Kimi K3.1 Pro tech blog, Moonshot AI
• deepseek-ai/DeepSeek-V4.1-Flash model card, Hugging Face, 2026-10-16
• deepseek-ai/DeepSeek-V4.1-Flash-1017 model card, Hugging Face, 2026-10-17
• Build with Muse Spark Open, Meta, 2026-10-15
• MiniMax M3.1 Flash, MiniMax, 2026-10-19
• MiniMaxAI/MiniMax-M3.1-Flash model card, Hugging Face, 2026-11-05
• mistralai/Mistral-Large-4-1T-A52B model card, Hugging Face, 2026-11-02
• ricursiveAI/Ripple-65B-Preview model card, Hugging Face, 2026-11-06
• NousResearch/Arete-27B model card, Hugging Face, 2026-11-06
• Introducing Inkling 2 Small, Thinking Machines Lab, 2026-11-05
• LFM3 release blog, Liquid AI, 2026-10-27
• nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16 model card, Hugging Face, 2026-10-23
• reflection/Beam-1.0-501B-A23B model card, Hugging Face, 2026-11-03
• baidu/ERNIE-5.5-Flash-Base model card and license, Hugging Face, 2026-11-05
• LLM Leaderboard (Intelligence Index v5.0.1), Artificial Analysis