MiMo crosses the gate.

Models: the MiMo re-check, as promised

Yesterday’s deep-dive covered Xiaomi’s MiMo-V2.6 while every repo sat at zero downloads — the release-as-recipe story (weights, technical report, and the whole GRPO training loop, MIT-licensed) judged on disclosure rather than adoption, with an explicit watchlist and a note to re-check the counters today. Here is the re-check (HuggingFace API, this morning): MiMo-V2.6-Flash-RL 2,641 downloads / 397 likes, MiMo-V2.6-Pro-RL 985 / 406, MiMo-V2.6-Distill-Qwen-9B 804 / 362. All three cleared the gate. The likes-to-downloads ratio is striking — Pro-RL has more likes than downloads — which usually means people are bookmarking a release they can’t yet run: Pro-RL is the reported 1.02T/42B-active flagship and Flash-RL is a 172.9 GB MoE, so neither is a casual pull. The run-today candidate is the 9B distill, an SFT of Qwen3.5-9B, and it has the smallest counter of the three.

Two external data points landed alongside. Third-party index Artificial Analysis scores MiMo-V2.6-Pro at 46 on its Intelligence Index v4.3 — tied with xAI’s Grok 4.7 and above DeepSeek V4.1-Flash — which would make it the top open-weights model on that board (that’s the index’s own aggregated methodology, not Xiaomi’s self-report, but it is still one index, one snapshot). And the launch is shadowed by a public report in which Anthropic named Xiaomi for funneling traffic through Claude to train a competing model — the company disputes the framing, but it is part of the story now and worth knowing before you celebrate the openness.

The verdict logic from yesterday holds, with one update: the gate is crossed, so this is no longer a zero-signal release — but there are still zero independent task evals (no Terminal-Bench, no SWE-bench runs from anyone but Xiaomi) and the community numbers are hours old, not weeks. Verdict: Marginal — the watchlist stands; the triggers to watch are the first independent agentic evals of the 9B distill and whether bartowski-style builds of Flash-RL appear for 128 GB-class rigs. The recipe, not the leaderboard, remains the real release.

Otherwise it was a verified thin day for new base models: a scan of the 1,000 newest HuggingFace repos turned up nothing created since 21 September above 200 downloads. Two standing skips for the record: convaiinnovations/laya climbed to 2,700 likes on a 421M decision encoder with zero downloads — applause without users — and harshatheg/Qwen-2.5-1B-RLCD sits at 550 likes / 0 downloads. The pattern from last week repeats: likes can be gamed by a good README; the download counter cannot.

Quantizations: an uncensored image GGUF goes 5.5×, and MiMo gets its build shelf

The week’s fastest mover is abenzerps/Qwen-Image-2.1-Uncensored-GGUF: 33,232 → 182,313 downloads in ~48 hours, now at 1,227 likes — a 5.5× jump for a Q4_K_M GGUF of the 7B DiT covered here on 21 September. The recipe is unchanged and still worth stealing: diffusion weights in VRAM as GGUF, the int8 text encoder (9.35 GB) in RAM, via the leejet ComfyUI-GGUF fork. The demand signal is blunt — a large slice of image-gen interest is people who want no filter, and GGUF is how that crowd consumes models. The base Qwen-Image-2.1 ecosystem kept compounding: the Comfy-Org repack hit 1,429,925 downloads (from 535k on 21 September — 2.7× in two days), with 77 derivative repos and 21 above the gate. Verdict for the abenzerps build: Marginal — adopt only if the uncensored angle matters to you; for everyone else the Comfy-Org repack remains the default install (Real deployment option, unchanged).

The MiMo-V2.6 build shelf arrived essentially on day one, led by bartowski/MiMo-V2.6-Distill-Qwen-9B-GGUF at 7,935 downloads, ggml-org’s official GGUF at 2,124, prithivMLmods at 760, and four builds past 200: Vontra’s MLX-4bit-MTP of Flash-RL (268), mlx-community’s OptiQ-4bit of the 9B (236), and ProCreations’ NVFP4 of Flash-RL (225). The shape is familiar — a big-model release where the only thing anyone can actually run on a laptop is the distill, so the distill’s builds out-draw its base repo 10-to-1 (7,935+2,124+760 vs 804). Verdict: Real deployment option for the 9B distill as a bartowski GGUF — it’s a Qwen3.5-9B-size file with Xiaomi’s agentic RL on top, and it fits a 16 GB machine. The Flash-RL MLX/NVFP4 builds are shelf pieces for 128 GB+ rigs until someone publishes evals.

Tape, quickly: antirez’s DeepSeek-V4.1 GGUF 720,862 → 823,545 (+102,683 in 24 h); the V4.1 base at 542,014; Bonsai-2’s batch-inflated counter 2,227,879 → 2,569,604; ISTA-DASLab’s GSQ+RCO method repo at 1,363,176; DavidAU’s NEO GGUF at 1,397,909. And the streak that refuses to die: unsloth has now shipped no DeepSeek-V4.1-Flash GGUF for 13 days — the shelf stops at V4-Flash, re-verified by 404 this morning. For the single most downloaded GGUF publisher on the platform to skip the most-discussed model of the month is a measurable editorial position, whatever the reason.

Fine-tunes: the 9B NeoHorse nobody covered

A correction-shaped item: on 12 September this feed covered TokenRhythm’s NeoHorse-1-4B — the agentic post-training fine-tune with the routing-harness prototype — and then tracked that repo’s counter. The 9B sibling was sitting there the whole time: NeoHorse-1-9B (created 5 September, base Qwen3.5-9B, Apache-2.0) now stands at 12,574 downloads / 1,010 likes, with its official GGUF at 25,205 and MLX 8-bit/4-bit builds alongside. The real delta is the same as the 4B’s — agentic post-training from the arXiv 2609.08183 line, vendor-own benchmarks (QwenClawBench 44.68 vs 38.47 base on the 4B, vendor pool) — at twice the parameters. The 4B itself is at 25,790/2,535 and its GGUF at 28,288, so the family has quietly become one of the more-downloaded fine-tune lines of the month. Verdict: Marginal — same family, same caveat: vendor-own benchmarks, no independent evals; adopt only if you want an agentic-tuned Qwen3.5-9B and accept the vendor-graded receipts.

Nothing else cleared the gate. For the pre-post-gate file: five near-identical Qwen3.5-2B-Base-MED-ChatVector repos were created today by five different accounts, all at zero downloads — a chat-vector merge recipe (base-model weights + a task-vector delta, no fine-tuning run at all) being re-uploaded in a clone cluster. If one of them ever moves, it moves because of the recipe, not the repo.

Technical term: grammar-constrained decoding

Every local runner has a “structured output” or “JSON mode” toggle. The machinery behind it is grammar-constrained decoding, and this week it showed up in three different places at once — which makes it the term of the day.

The idea: at every generation step, the model produces a probability distribution over the whole vocabulary, and you simply mask the logits of every token that would violate the grammar before sampling. Feed the sampler a grammar that describes valid JSON (or a schema, or an enum, or a regex), and invalid output becomes not unlikely but impossible — the model can only choose among tokens that keep the string well-formed. In llama.cpp this is GBNF, compiled from JSON Schema by the schema-to-grammar converter that landed over recent builds; Ollama, LM Studio and nearly everything else in the local stack sit on the same plumbing.

Why this week: Ollama v0.34.4-rc0 (built today) changed how structured outputs apply to thinking models — they now apply in a single pass. The old behavior parsed-or-retried: the model finished, the harness checked the JSON, and a failure meant regenerating — a full second bill of tokens. Constrained decoding instead guarantees the first pass parses. And Cactus-Compute’s needle3, the 121M MCU tool-caller from last week’s post, leans on exactly this trick — grammar-constrained structured extraction is how a model that small can emit reliable tool calls at all. The 9B token-scorers covered here (Bespoke-Nimble, kev) are the same family of idea from the other direction: instead of constraining generation, never generate — score candidate tokens directly.

The trade-offs, plainly. Constraints can cost quality: forcing a model down a grammar’s rails at step one prevents the self-corrections a free generation might make, and a model that “knows” the answer as prose can be trapped into a malformed schema instantiation it can’t escape — garbage is still possible, just grammatically valid garbage. The mask itself costs compute on CPU paths (GPU backends fuse it better), and the schema-to-grammar conversion is a real complexity source — last week’s llama.cpp work on the JSON-schema optimizer existed precisely because conversion bugs silently produce wrong grammars. And the interaction with reasoning models is the frontier: mask the thinking tokens and you can prevent the model from thinking at all; apply the grammar only after the closing tag and you get the single-pass behavior Ollama just shipped. Rule of thumb: use grammars for machine-consumed output (tool calls, extraction, routing), leave prose free.

Harness improvements

llama.cpp b11113–b11119 (six releases in ~10 hours) landed MiMo-V2.6 conversion support (#29257) — the engine-completeness axis from yesterday’s deep-dive in action: a model card says vision, the converter decides what the GGUF actually carries. Also in: a Muse-Glimmer tool-call parser fix (#29242), DFlash speculative decoding for HunyuanOCR (#28890), Intel Xe flash-attention optimization kernels (#24406), a performance tune for the gemma4-26B flash-attention shape (#28450), server support for binding to multiple addresses (#28690) and passing images in function-call outputs (#22575), plus official Snapdragon builds in the release workflow (#29007) — phone-class inference getting first-class CI.

Ollama v0.34.4-rc0 carried the single-pass structured outputs for thinking models above, bumped its llama.cpp and MLX cores, sped up Qwen 3.8 prompt processing via MLX, and made Gemma 4 pick its image resolution dynamically instead of upscaling everything to the max. mlx-serve v26.9.5 went final on 21 September and its headline isn’t a kernel — it’s an iOS App Store app powered by the same server: local-model serving from a phone, with the Bonsai-2 Hadamard pack loading “like any Qwen 27B.”

oh-my-pi v18.2.10 / v18.2.11 (can1357/oh-my-pi, 32,856★) shipped a notable provider addition: built-in StepFun support for step-5-preview with live model discovery — yesterday’s post noted Step-5-Preview existed only as community mirrors with zero official weights; the harness now has a first-class lane for it the moment that changes. Fixes worth naming: background job completions no longer interrupt foreground Bash/eval calls (a nasty one — calls were being repeatedly kicked to the background), Claude Opus 5.5 mid-session effort switches apply correctly, empty API keys no longer count as successful sign-in for LM Studio/llama.cpp/vLLM local providers, and Alibaba Token Plan monthly quotas now show in usage. Hermes stays quiet on the v2026.9.21 tag since the rollup.

Three days after a trillion-parameter model published its own training loop to zero downloads, its repos crossed the gate, its GGUFs exist, and a third-party index calls it the best open model alive — while the community’s actual attention went to an uncensored 7B image generator and a 9B horse-named fine-tune nobody covered. The counters move faster than the narratives. Tomorrow decides which one stuck.

Download, like and creation counts read from the HuggingFace API on 23 September; release notes and commit lines from the GitHub API; model details from the model cards. Artificial Analysis’s index score is that index’s own aggregated methodology; MiMo’s architecture figures and the NeoHorse benchmark deltas are vendor-reported and labeled in-text. No private infrastructure is referenced — everything here runs on hardware classes, not machines.