Opening

A 27B community fine-tune with 1.5M downloads and 2.6k likes on Hugging Face, where the author hands you one-line commands for llama.cpp and Ollama right there on the model card. That's the first thing we're talking about today: DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NEO-CODER-MAX-MTP-GGUF. The other two picks: a 0.8B model where only two tensors were changed and no training happened at all, and a collection of video-generation turbo LoRAs for ComfyUI users. As for how that "no training, yet dramatically different behavior" story actually works, I'll get to that shortly.

Let me be upfront: this show did not actually run any of these models. Every capability, performance, and hardware claim below is quoted from the official model cards and pages.

This Issue's Picks

1. DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NEO-CODER-MAX-MTP-GGUF: a 27B GGUF packaged for consumer hardware

• What it is: The name is long enough to sound like a national ID number, but break it apart and it's actually pretty clear. This is a multi-stage community fine-tune built on Qwen/Qwen3.8-27B. The author describes it on the model card as built "for consumer hardware" and says it was trained with Unsloth on consumer hardware. The repo ships GGUF directly, 27B parameters, apache-2.0 license. The model card also notes this is a de-censored (heretic) version, and the Features list separately includes Vision.

• Where you'd use it:

  • Running a local coding agent on your own workstation: You've already got a decent card and want to spin up an OpenAI-compatible server with llama.cpp, then hook up a CLI agent like Pi or OpenClaw to work on your own repo. The TOOL CALLING section of the model card is very specific: it recommends at least q4km, q5ks/q5km is better, Q6 is the top pick, and it explicitly warns that below q4km tool calling may break. That kind of "here's the quant level where things fall apart" information is actually rare on most model cards.
  • People who can't stand reasoning models spinning their wheels: The author writes that this version is called TURBO because thinking tokens are compressed to between 1/2 and 1/10 of the original, and this applies across all three reasoning modes (xhigh/medium/low). If your workflow is multi-turn local refinement, tweak a prompt a bit, tweak it again, having to wade through a huge thinking block every round genuinely wears you down. This approach is worth a look.
  • Writing or story drafts without refusals: The model card says this is a de-censored version, and the page includes five creative writing demos. Novel drafts, branching game narratives, the kind of material generation that keeps getting blocked by safety policies: that's the positioning the author gives it himself. And since Features lists Vision, image Q&A is also within its claimed capabilities.

• Can you run it?: No specific VRAM numbers are given. The model card only says it's built for consumer hardware and was constructed with Unsloth on consumer hardware. What it does give is quantization guidance: minimum q4km for tool calling, Q6 recommended; for sampling, Temp 0.6 or 0.7, Rep pen set to 1 (i.e., off). My own take is that for a 27B at the q4km level, you basically need enough room to "fit the model file itself into memory." That's a common-sense estimate derived from the parameter count and quant format, not an official figure, so go by the actual file size you download.

• Buzz this issue: HF trending score 250, 1.5M downloads, 2.6k likes, the highest-profile pick this issue. There are 5 quantized or repackaged versions from the same family on Hugging Face this issue.

• Versus the alternatives: All of the following are the author's own claims. He says this model beats base Qwen 3.8 27B on all 7 core benchmarks, and also beats Qwen3.6-35B-A3B, Qwen 3.6 27B, and Qwen 3.5 27B. He claims it's the first fine-tune at this size to hit 735 on 8-bit ARC-C and 719 on 4-bit. He also notes that if your hardware can't handle it, Qwen3.5-9B-The-Defiant-Fable from the same series is a smaller alternative, though not as strong as this one. The page asks readers to check the community tab for third-party benchmarks, which I think is pretty honest: go dig around there before you trust the numbers.

• How to get it: The Local Apps section of the HF page gives you commands directly. For llama.cpp, use llama serve -hf DavidAU/Qwen3.8-27B-TURBO-...-MTP-GGUF:Q4_K_M to start a local OpenAI-compatible server, or ollama run hf.co/DavidAU/Qwen3.8-27B-TURBO-...-MTP-GGUF:Q4_K_M. The same repo includes both regular and MTP NEO-CODER MAX dual-imatrix quants.

• Link: https://huggingface.co/DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NEO-CODER-MAX-MTP-GGUF

2. ML-Intern-lab/Qwen3.5-0.8B-weight-space-programming: two tensors changed, zero training, a perfect control group

• What it is: 0.8B parameters, BF16 safetensors, Apache-2.0. The point isn't the scores, it's the subject matter. The model card says this is the previously unreleased checkpoint from experiment 8 of "Sparks of Weight Space Programming," in which a frontier LLM directly hand-edited 131,072 weights with no training whatsoever. More importantly, the author notes that relative to the public base (Qwen/Qwen3.5-0.8B, with a specified commit), only two tensors differ: the mlp.down_proj.weight of layers.15 and layers.19.

• Where you'd use it:

  • Teaching material for a reading group or internal talk: You want to explain to your team what actually separates "weight editing" from "fine-tuning," and talking about it alone is abstract. The model card says the base model loops its output until it hits the token limit on nearly every math problem while the edited one doesn't, and it includes a Space demo where you can watch base and edited versions tackle the same problem side by side. Dropping that demo into your slides lands harder than ten minutes of explanation.
  • Interpretability or model repair research: With only two tensors different, you get an extremely clean control group. You can diff the weights directly and study which parameters are actually responsible for pathologies like repetition or looping. Material with only two variables is genuinely hard to come by.
  • Reproducing an eval pipeline on a small model: The model card spells out the eval protocol thoroughly: vLLM, thinking enabled, temperature 0.6, top_p 0.95, top_k 20, presence_penalty 1.5, up to 32,768 new tokens, seed 0, with the final \boxed{} graded by math-verify 0.9.0. Raw scores and per-question results live in the repo's eval/ directory. If you want to practice building your own eval pipeline, re-running theirs is a very low-cost starting point.

• Can you run it?: No consumer hardware threshold is stated. The model card only mentions the validation environment was an NVIDIA H200 with vLLM 0.33.0 and transformers 5.20.2, and notes that weight reconstruction and hash verification were done on CPU. On the file side it's 0.9B parameters, BF16 safetensors. Based on that parameter count and precision, the weights should land around the 2GB mark, which a typical laptop or consumer GPU can handle. One caveat though: that's my own estimate from parameter count and format, the official page gives no consumer hardware numbers at all.

• Buzz this issue: HF trending score 64, 538 downloads, 64 likes. Small numbers, but for an experimental checkpoint like this, the likes-to-downloads ratio is actually pretty interesting.

• Versus the alternatives: Compared with the base Qwen3.5-0.8B, the model card's own table shows MATH-500 accuracy of 5.0% (25-question smoke test, the original text says 7.9) against 51.0% for this checkpoint (original text says 48.1), with the truncation rate dropping from 95% to 31.6%. The author also explains that this checkpoint was reconstructed using the original author's apply_edit.py with sha256 comparison on every tensor. All of this comes from the model card.

• How to get it: The Usage section of the model card gives a vLLM example directly, LLM(model=ckpt, dtype="bfloat16", max_model_len=34816), paired with the tokenizer's chat template (enable_thinking=True). The HF page also provides a transformers pipeline and a one-line vllm serve command.

• Link: https://huggingface.co/ML-Intern-lab/Qwen3.5-0.8B-weight-space-programming

3. Minimax H3 Turbo LoRas: turbo LoRAs from various sources, repacked so ComfyUI can actually load them

• What it is: The only adapter pick this issue, hosted on Civitai. The base model is MiniMax H3, format is BF16 SafeTensor, 591.55 MB per file, plus an optional 23.97 KB MiniMax_H3.json file. Worth clarifying the positioning first: the uploader explicitly states they did not train or distill the original turbo weights. What this page offers is a modified compatibility version, intended to let LoRAs that previously wouldn't load be read by ComfyUI's built-in LoRA loader when using pruned or curve-form models.

• Where you'd use it:

  • Cutting the step count for MiniMax H3 video generation: The page lists recommended parameters for each version. For example, larryvrh_v4_600_ema is 6 to 8 steps, sampler er_sde or res_multistep, scheduler simple, Strength 1.0; the dmad version uses er_sde or euler, with Video/Audio Shift at 12/2. If you already have a working workflow and the only problem is how long each clip takes, this kind of turbo LoRA is the classic fix.
  • Comparing turbo LoRAs from different sources side by side: This single page collects lightx2v, larryvrh, DMAD, taomate_fl2va, pika and more, 7 variants in total. For anyone serious about tuning quality, being able to swap LoRAs on the same base and compare visuals and audio saves you from hunting down files all over the place.
  • People stuck on compatibility issues: If you've ever downloaded a turbo LoRA, dropped it into ComfyUI, and watched it fail to load, solving exactly that is why this page exists. The page credits drbaph and kijai for the ComfyUI-compatible versions.

• Can you run it?: No VRAM or GPU threshold is stated. The page only gives file specs plus step count and sampler recommendations. Licensing deserves a careful look: the page states that the MiniMax H3 Community License Agreement's scope excludes the EU, UK, South Korea, and the US. On the Civitai terms side, derivatives are allowed and commercial use is marked as Image/RentCivit/Rent/Sell/SellMerge. If your project has cross-border or commercial considerations, go read the original text on the page yourself, don't make the call based on my summary.

• Buzz this issue: 54.2k downloads on Civitai's weekly chart, rated Overwhelmingly Positive across 969 reviews.

• Versus the alternatives: The uploader's own assessment is that dmad_minimax_h3_4step_full_critic_comfyui matches larryvrh v4 on imagery, but they personally find the audio quality better. The original weights come from larryvrh, the lightx2v team, TaoLiveAIGC, Bytedance, the DMAD project by Texas A&M students, and Pika-Labs. To say it again: this page is about integration and compatibility, not original training.

• How to get it: Pick a version on the Civitai page and download the .safetensors, 7 variants, with the default best match being minimax_h3_turbo_v4_step600_pruned_comfyui.safetensors. Drop it into your ComfyUI LoRA directory and load it with the built-in LoRA loader. The page notes all demo images were generated at strength 1.0.

• Link: https://civitai.com/models/2837571

Closing

What these three have in common is that each author puts "how to wire this up on your own machine" right up front: the 27B gives you a one-line llama.cpp command and an explicit quantization floor, the 0.8B gives you a complete eval protocol and an eval directory, the LoRA gives you steps and samplers for every version. That's genuinely not the norm in the open source ecosystem, and it's worth noticing. If you only want to touch one thing today, I'd say grab the 0.8B weight-space-programming checkpoint first. It's the lightest and fastest to download, and for that "two tensors, zero training" story, seeing the demo once yourself hits far harder than hearing me describe it. See you next episode.