MiaAI_lab reported updates to GLM 5.3 Flash deployed in EXL3 quantization on a configuration of 2x DGX Sparks hardware.

A public Hugging Face repository identified as openbmb/MiniCPM5-2B-DSpark-GGUF provides a derivative GGUF conversion of MiniCPM5-2B-DSpark, with a listed weight file named MiniCPM5-2.6B-DSpark.gguf. The repository is classified as a derivative and carries a declared Apache-2.0 license.

The repository was created on 2026-09-09.

A public Hugging Face repository for openbmb/MiniCPM5-2B-GPTQ has been created, listing safetensors weight files and carrying a declared Apache-2.0 license. The repository is classified as a derivative, is not gated, and was created on 2026-09-07.

Repository creation is not a release date, and the file listing does not verify a benchmark, license compliance, successful download, or runtime support. Download access may require accepting publisher terms.

A public Hugging Face repository at ornith-ai/Ornith-1.5-35B-A3B-GGUF lists GGUF weight files in BF16, Q4_K_M, Q5_K_M, Q6_K, and Q8_0 quantizations, along with a multimodal projector file named mmproj-Ornith-1.5-35B-BF16.gguf. The repository declares its license as MIT.

TraffAlex reports that Qwen3.8-27B scored 52 on the Artificial Analysis index, placing it at the same level as DeepSeek V4-Pro and GPT-5.6 Luna, and describes it as the first local model to hit frontier-class scores. According to the same post, Unsloth's dynamic quants fit the model in 17GB, and every tier from a 24GB 3090 to a 32GB 5090 runs it.

The post reports that a 24GB 3090 runs Qwen3.8-27B 8-bit at 115+ tok/s, while a 32GB 5090 runs it at 132K context at approximately 60 tok/s. These figures are attributed to TraffAlex's reporting and have not been independently verified.

sudoingX reports that Ling-3.0-flash-int4 runs at 42 tok/s and notes that the vision-language variant Ling-3.0-flash-VL-int4 fits on one Spark at 76GB. A speed benchmark for the VL-int4 variant is planned but has not yet been completed.

TraffAlex reports that DeepSeek-V4-Flash-0731, described as a 304B mixture-of-experts model, runs on 2x DGX Sparks and achieves near-Opus quality at 75–89 tokens per second with 1M+ verified context. The post characterizes the model as one that "owns the 128GB tier."

The same post notes that V4.1-Flash just rolled out on September 11 and is trending again, though the reported performance figures are attributed specifically to the V4-Flash-0731 variant.

TraffAlex reports that Qwen3.8-2.4T-A95B compresses from 4.9TB to 397GB at 1-bit quantization, with the post describing the result as rivaling GPT-5.6 Sol in quality.

TraffAlex reports that Gemma 4 12B processes text, image, and audio modalities while fitting within 8GB VRAM, positioning the model in the small and edge tier alongside the 26B-A4B and 31B variants.

TraffAlex reports that Kimi K3, which launched July 22, can run on a MacBook Pro when quantized to 1-bit. The same post states that the model's launch crashed Moonshot's servers badly enough that the company paused signups for 48 hours.

The repository inclusionAI/Ling-3.0-flash-int4 is listed as public on Hugging Face with a declared MIT license and a canonical repository class. Hugging Face metadata for the repository reports these access and licensing details.

The repository's listed weight files consist of 24 safetensors files, spanning model-00001-of-00024.safetensors through model-00024-of-00024.safetensors.

TeksEdge reports that Saanora took Qwen3.5-9B and trained it to decide when an answer would be better as a chart, an interactive 3D scene, a self-grading quiz, flashcards, a simulation, an explorable concept map, or interactive steps, and then generate the structured data needed to render it. The post reports 84% valid output on the first try, 87% correct selection of plain prose when a UI artifact is not useful, and 12.3 out of 14 multi-turn artifact edits. The model is described as a 9B dense model with 262K context, open weights, and standard Qwen3.5 architecture, with a community llama.cpp GGUF quantization at Q4_K_S approximately 5.5GB.

Reported reasoning benchmarks for the post-trained model are MMLU-Pro 80.6, GPQA Diamond 79.4, and AIME 2026 83.3. TeksEdge reports that on several of those benchmarks it beats the published numbers for Gemma 4 12B, although other small models still lead individual benchmarks.

MichaelGannotti reports that a two-stage Sol-H3 recipe on one NVIDIA DGX Spark reduced the wall time for generating 5 seconds of 768p MiniMax-H3 video from 423 seconds to 68 seconds, a 6.2× speedup on the matched cell at 3× the pixels. He notes that NVIDIA's reported 56-second target was not reached and that the generated clips remain internal.

sudoingX reports that on a DGX Spark 128GB, Qwen3.8-Flash-Next (180B with 11B active) decoded at 44.4 tok/s, holding its speed at depth with a 35% drop at 63K context, and its 1M context fits in memory, measured to 128K. Ling-3.0-flash (124B with 5.1B active) hit 42.1 tok/s on the official int4 under MIT license, described as the fastest vendor-official release on the board, but it collapses as context fills, dropping 83% at 51K with a 256K context. Laguna-S-2.1 (117.6B with 8.5B active) ran at approximately 35 tok/s on code with dflash, staying flat at any depth across a 262K context, and is described as coding native.

Qwen3.5-122B (122B with 10B active) achieved 35.3 tok/s with MTP, losing 15% speed from 90K to 231K context across a 262K context, and is called the reliable all-rounder. DeepSeek-V4-Flash fits at 3bit on one box but decoded only 16.5 tok/s with no drafter, which sudoingX reports is too slow to build with.

MichaelGannotti reported that Ling 3.0 Flash is running for free on OpenRouter. The model's Hugging Face repository, inclusionAI/Ling-3.0-flash, is listed as public and ungated with a declared MIT license, and was created on August 2, 2026.

Elliot Arledge reports benchmarking DeepSeek V4.1 Flash on KernelBench-Mega using a custom Kimi-Linear Decode kernel targeting RTX PRO 6000, achieving 17.1x the optimized PyTorch baseline overall. The kernel runs one cooperative launch with 19 grid barriers, fusing int4 dequant GEMVs with the zero point folded into the FMAs, the short conv, the gated-delta state update, absorbed MLA over the latent cache with a cross-block softmax merge, and a router plus eight routed experts. Speedup was 15.8x at 2K context and 17.8x at 16K context, with the run taking 3.2 hours on an unlimited budget and exiting on its own.

Arledge notes a persistent atomicAdd accumulator that is never zeroed, causing every layer from the second KDA layer onward to inherit the previous layer's activations. Output cosine similarity is 0.9997 on the first token and 0.98 on the next, which passes the 0.98 gate. He states that zeroing the accumulator costs nothing and describes the result as interesting, not clean. Arledge adds that he has not used the model much beyond this benchmark, though it has been a decent general task delegator.

teortaxesTex reports that GLM-5.3-Flash performs substantially worse than V4.1 Flash on kernel engineering tasks, describing the gap as significant. The post characterizes V4.1 Flash as showing a "really massive improvement" in this domain, while stating that GLM-5.3-Flash is "nowhere close."

Snixtp reports running GLM-5.3-Flash as a main orchestrator on 2x PRO 6000 GPUs, with Qwen3.8-27B deployed as subagents on 4x RTX 3090 GPUs, describing the multi-model setup as working "surprisingly well." Both models are available as public, ungated Hugging Face repositories: GLM-5.3-Flash under zai-org with a declared MIT license, and Qwen3.8-27B under Qwen with a declared Apache-2.0 license.

analogalok reports that Qwen3-TTS-12Hz-1.7B-VoiceDesign, a discrete multi-codebook LM using the Qwen3-TTS-Tokenizer-12Hz, generates speech from natural language conditioning alone. As the post describes it, "No reference audio. No fine tuning. Just raw natural language conditioning." The user describes age, gender, accent, microphone proximity, and emotional state in plain text, and the model synthesizes a new voice identity from scratch. The Hugging Face repository for Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign is public, ungated, and lists Apache 2.0 as the declared license, though its metadata notes that repository creation is not a release date and a file listing does not verify successful download or runtime support.

analogalok reports peak VRAM of approximately 6 GB at BF16 and that the model "Runs smoothly on a free Google Colab T4 GPU." The post characterizes the result as open weights achieving "steerable, cinematic voice generation on 6 GB VRAM." These runtime figures are a single user's report and were not independently verified.

MiaAI_lab reported the availability of official FP8 versions of DeepSeek-V4.1-Flash, GLM-5.3, and GLM-5.3-Flash. HuggingFace metadata confirms that the deepseek-ai/DeepSeek-V4.1-Flash repository is public, ungated, classified as canonical, and carries a declared MIT license, with a repository creation date of 2026-09-10. The zai-org/GLM-5.3 and zai-org/GLM-5.3-Flash repositories are also public, ungated, and canonical, both created on 2026-08-25; GLM-5.3 carries a declared license of "other" while GLM-5.3-Flash carries MIT.

A public Hugging Face repository named poolside/Laguna-S-2.1-NVFP4 provides an NVFP4 quantization of Laguna-S-2.1, declaring poolside/Laguna-S-2.1 as its base model. The repository is classified as canonical and is published under the openmdw-1.1 license.

The repository listing includes 49 safetensors weight files, with names ranging from model-00001-of-00049.safetensors through model-00030-of-00049.safetensors visible in the metadata.

A free Google Colab notebook for the Qwen3-TTS-12Hz-1.7B-VoiceDesign voice design demo has been announced by analogalok. The model repository Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign is listed as public on Hugging Face with a declared license of Apache-2.0 and a canonical repository class.

MiaAI_lab announced an EXL3 quantization of Qwen3.8-27B designed to run on 16 GB Nvidia GPUs, offered through a one-click installer. The post describes the configuration as ready for action but does not provide benchmark results or further technical details about the quantization.

The base model repository Qwen/Qwen3.8-27B is publicly available on Hugging Face as a canonical repository, created on 2026-08-05, with a declared Apache-2.0 license and no gating. The metadata does not describe the EXL3 quantization artifact itself.

An FP8 quantization of GLM-5 is publicly available in the Hugging Face repository zai-org/GLM-5-FP8, with metadata showing it is not gated and carries a declared MIT license. The repository is classified as canonical and its listed weight files include 142 safetensors shards, with names ranging from model-00001-of-00142.safetensors through model-00030-of-00142.safetensors in the sampled listing.

yume_arasaki reports a deployment gotcha for Qwen3.8-27B-DFlash2 on SGLang: the stock launcher ships without reasoning and tool-call parsers, so qwen3 and qwen3_coder parsers must be added or thinking returns as raw XML and tool_calls stays null. The reported harness configuration uses SGLang with a pinned digest, DFlash2 K=10, KV bf16, and 262144 context, with needle tests run at 3 repeats.

Hugging Face metadata lists the repository z-lab/Qwen3.8-27B-DFlash2 as public and canonical with a declared Apache-2.0 license and Qwen/Qwen3.8-27B as the declared base model. The metadata notes that repository creation is not a release date and that a file listing does not verify a benchmark, license compliance, successful download, or runtime support.

A public Hugging Face repository from ggml-org offers GGUF-format conversions of Qwen3-TTS-12Hz-1.7B-Base, including Q4_K_M, Q8_0, and bf16 weight files alongside corresponding mmproj files for Q8_0 and bf16. The repository is classified as a derivative, is not gated, and has no declared license.

A public Hugging Face repository for inclusionAI/Ling-3.0-flash-VL-fp8 provides an FP8 quantization variant of the Ling-3.0-flash-VL model. The repository is listed as canonical and carries a declared MIT license.

The model weights are distributed across 64 safetensors files, with the listed weight files including model-00000-of-00064.safetensors through model-00029-of-00064.safetensors among others.

A public Hugging Face repository named Qwen/Qwen3.8-Flash-Next-FP8 lists an FP8 quantization variant of Qwen3.8-Flash-Next with the declared base model Qwen/Qwen3.8-Flash-Next. The repository is classified as canonical, is not gated, and declares its license as "other." Hugging Face metadata shows the repository was created on 2026-08-24.

The repository listing includes 131 safetensors weight files, from model-00001-of-00131.safetensors through at least model-00030-of-00131.safetensors. Hugging Face notes that repository creation is not a release date and that a file listing does not verify a benchmark, license compliance, successful download, or runtime support; download access may require accepting publisher terms.

TheZachMueller reported that NVIDIA has produced an NVFP4 quantization of GLM-5.3-Flash. The canonical GLM-5.3-Flash repository on Hugging Face, hosted by zai-org, was created on 2026-08-25 and is ungated under an MIT license.

0xSero reports that Ternary-Bonsai-27B-gguf takes Qwen3.6-27B and quantizes the weights to Q2, targeting systems with 8GB VRAM while providing a 262K token context window and vision support. Hugging Face metadata for the prism-ml/Ternary-Bonsai-27B-gguf repository confirms it as a derivative of Qwen/Qwen3.6-27B, with listed weight files including Q2_0, PQ2_0, Q2_g64, dspark-Q4_1, F16, and dspark-bf16 variants, plus mmproj vision projector files in BF16 and Q8_0 formats.

A public Hugging Face repository called r0b0tlab/Qwen3.8-27B-NVFP4-MTP-sm121 hosts an NVFP4-quantized build of Qwen3.8-27B with Multi-Token Prediction, distributed across four safetensors weight files. The repository is listed as public, ungated, and declared under the Apache-2.0 license, with a repository creation date of 2026-08-16.

TeksEdge reports running GLM-5.3-Flash, a 320B-parameter Mixture-of-Experts model with 18B active parameters, at Q4_K_XL quantization (186 GiB) using 96GB of total VRAM and approximately 208GB of host RAM for pinned experts. The demonstrated machine paired 2× RTX 5090 and 2× RTX 5080 with 512GB DDR5, achieving 24.5 tok/s decode at 8K context and 159 tok/s prefill at 27K context.

TeksEdge cautions that the LayerStoRm inference engine is very experimental, RTX 50-series only today, with PCIe 5.0 recommended, and not production-ready. The GLM-5.3-Flash repository on Hugging Face, created on 2026-08-25, is listed as canonical with an MIT declared license and is not gated.

A public GGUF conversion of Ternary-Bonsai-27B is available in the prism-ml/Ternary-Bonsai-27B-gguf repository, listed as a derivative of Qwen/Qwen3.6-27B under the declared Apache-2.0 license. The repository's listed weight files include six model quantization variants: F16, PQ2_0, Q2_0, Q2_g64, dspark-Q4_1, and dspark-bf16.

The repository also includes two multimodal projector files, mmproj-BF16 and mmproj-Q8_0, alongside the model weight files.

MiaAI_lab announced on September 8, 2026 that an EXL3 quantization of GLM-5.3-Flash is receiving an update, described as one that will "upgrade your experience." The post does not specify what changed in the quantization or when updated artifacts would be available.

The canonical GLM-5.3-Flash repository on Hugging Face, listed under zai-org/GLM-5.3-Flash, is public and ungated with a declared MIT license. The repository was created on August 25, 2026 and its weight files are distributed across 62 safetensors files.