Three model stories landed within days of each other this month, and they answer different questions. One is about how small a usable agent model can get. One is about what "efficient" means when the checkpoint is 178 GB. One is a promise with a deadline attached.
Saluki squeezes Qwen3.8-27B into 7.89 GB
A community quantization called Saluki 27B puts Qwen3.8-27B into a 7.89 GB GGUF file aimed at local tool calling. The model card says the main file runs in stock llama.cpp, with text generation kept separate from an optional vision add-on. The repository lists the main model plus two downloadable vision files, so you can verify the packaging yourself rather than trusting a benchmark table.
Hugging Face dates the repository creation to October 8, with the last model card edit on October 10. A trend snapshot taken October 12 showed 108,900 cumulative downloads and 342 likes. Those counters do not tell you how many distinct people are running it or how fast it is growing, and I would not read them as a quality signal. What they do show is that a lot of people are at least curious.
The pitch is narrow and testable: tool calling on local hardware. A 27B model in under 8 GB implies aggressive quantization, which usually costs you something on reasoning-heavy tasks. If your workload is "parse this JSON, call this function, return a result," that trade may be fine. If it involves multi-step planning over long documents, test before you commit to it.
I like this release for what it does not claim. No leaderboard sweep, no "frontier" language, just a file you can download and a llama.cpp build you probably already have on disk.
MiMo-V2.6-Flash and the naming mess
Xiaomi's MiMo-V2.6-Flash is a sparse mixture-of-experts model: 309 billion parameters in total, 15 billion active per token, 1M-token context, 128K maximum output, MIT license, no gating. The Hub's safetensors metadata reports 310,756,322,688 parameters, which lines up with the card. Earlier coverage, including ours, carried a 159B figure read off the same repository panel. That number is stale, and the current one is the one to use.
The repository you actually want is XiaomiMiMo/MiMo-V2.6-Flash-RL, created September 21, 90 files, roughly 178 GB of shards in FP8 (e4m3) with a dynamic activation scheme. Requesting the bare name XiaomiMiMo/MiMo-V2.6-Flash returns an authorization error because that repository does not exist. The -RL suffix marks the post-training lineage, not an adapter bolted onto some other base model. The file listing is the evidence: full backbone, vision encoder, audio stack, speculative drafter. An adapter is megabytes. This is 178 GB.
That naming has a real cost. People search for mimo-v2.6-flash-rl and then try to pass it to the hosted API, where the identifier is lowercase and unsuffixed: mimo-v2.6-flash. Type the suffix into a request and it will not resolve. Xiaomi's documentation says so plainly, which is more than most vendors bother to do.
Pricing on the hosted endpoint runs $0.0028 per million input tokens on a cache hit, $0.14 on a cache miss, and $0.28 for output. The base URLs are https://api.xiaomimimo.com/v1 for the OpenAI-compatible surface and https://api.xiaomimimo.com/anthropic for the Anthropic-compatible one. Authentication accepts either an api-key header or a Bearer token.
One discrepancy is worth flagging before you size drafter memory. The card describes a five-layer drafter predicting seven tokens. config.json in the same repository sets num_nextn_predict_layers to 3. Both cannot describe the same artifact, and the config is what the runtime reads. Size from the config and verify after loading.
The 1M context claim does check out cleanly: max_position_embeddings is 1,048,576.
A MOPD variant was published September 27. Its own card calls it the upgrade of the RL checkpoint rather than a different model. Same lineage, same scale, six days apart. Do not pull both expecting different memory footprints.
On independent evaluation there is exactly one score I have seen: 38 on an index where the larger MiMo-V2.6-Pro scores 46. Xiaomi's own benchmark tables have not been reproduced outside Xiaomi.
Mistral Large 4 and the trillion-parameter deadline
Mistral opened a public preview of Mistral Large 4 on October 6. One trillion parameters total, 52 billion active, natively multimodal, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centers. Internally it is ML4; the announcement also calls it "le Chonk."
The weights are promised by the end of October. Until then, Mistral says it is red-teaming with cybersecurity partners and state authorities on a version with reduced moderation and expanded cyber capabilities. The published figures are 82% on a vulnerability reproduce-and-patch test and 93% on Cybench. Nobody outside Mistral has verified either number, and the license terms for the weights have not been released.
A trillion-parameter open-weight model is mostly a question about serving. At FP8 you are looking at roughly a terabyte of weights before you count KV cache, activations, or tensor-parallel overhead. Most people reading this will hit ML4 through the preview API, not on their own rack. What open weights actually buy you here is inspection and fine-tuning access, which is a different value proposition than cost per token.
What to actually do with this
If you want a local tool-calling model on consumer hardware, Saluki is worth a download. 7.89 GB and stock llama.cpp is a low bar to clear, and the vision files are optional.
If you want a hosted MoE with long context, mimo-v2.6-flash at $0.14 per million input tokens on a cache miss is cheap enough to test without a budget conversation. Get the identifier casing right and read config.json before you trust the model card on drafter shape.
If you want Mistral Large 4 weights, wait for the release and read the license. Open weights and permissive licensing are separate things, and Mistral has not published the terms yet. Until it does, the ML4 numbers are a preview, not a reference.