The model at the base

Muse-Glimmer-30B comes from Meta's Superintelligence Lab, released 9 August 2026 under Apache-2.0 according to its model card and release article. It is a dense causal transformer with a perception encoder: 52 layers, hidden size 6656, 32 query heads against 2 KV heads (a 16:1 GQA ratio), attention that cycles local-local-local-global with a 2048-token sliding window, SwiGLU feed-forward blocks, RoPE at θ=500,000 applied to local layers only, a 202,048-token vocabulary and a 131,072+ context. The Hugging Face record puts the exact artifact at 29,776,626,688 parameters — about 29.8B including the ~1.8B ViT-G/14 vision encoder.

It is not sparse and it is not an MoE. Every stored parameter participates in every token; the “30B” is a memory claim, not a routing claim. Meta positions it as a local-first agentic model — distilled from Muse Spark by logit distillation, trained for tool use, failure recovery and long-horizon tasks, and shipped with a DFlash block-diffusion drafter for speculative decoding. The upstream benchmark table claims strong results against Gemma4-31B and Qwen3.6-27B (MCP Atlas 75.5, DeepSearch QA 74.6, SWE-Bench Verified 76.0, AIME 2026 94.7) — those are upstream claims, not our numbers. Independently, Sebastian Raschka's architecture notes call it Gemma-like, flag its extreme KV-cache efficiency (52 KiB per token in BF16 versus 840 KiB for Gemma 4 31B), and note the Artificial Analysis index places it slightly behind Qwen3.6 overall. Good model. Not the point of this note.

Three days, four methods

The point is what happened after the release. Reddit's LocalLLaMA immediately produced threads like “Glimmer seems pretty censored?” — community reports of refusals on mundane coding requests. That is community opinion, linked rather than repeated as fact, but it is the demand context for the wave. Within three days of the base release, at least four independent de-refusal efforts shipped, per their Hugging Face records: darkc0de's “heretic” build on 10 August; Blackfrost-AI's abliterated BF16 plus a full GGUF ladder the same evening; mlasli's classic refusal-direction ablation on 11 August (α=0.15 at layer 33, a reported separation score of 86.34, documented in its abliteration_info.json); and jorkle's on 12 August. So, honestly: jorkle's is not the first abliterated Muse-Glimmer. It is the first that publishes its method, its loss, its data recipe and a drift metric.

What jorkle actually did

The card is unusually specific. Its own header word is “De-abliterated variant”; the method is a KL-conserving, best-of-N steered LoRA SFT — a fine-tune, not a weight ablation. The loss is CE(compliance) + λ·KL(tuned‖base) with λ_KL = 1.0; r=16, alpha=16, lr 5e-5, two epochs, cosine decay to zero, 5% warmup, grad clip 0.3, batch 1 with grad-accum 8, max_seq 768, seed 0. Training data: a 544-prompt compliance set sampled four times per prompt at T=0.8 and refusal-filtered, with a 48-pair holdout. LoRA touches o_proj and down_proj only — 31.1M trained parameters, 0.10% of the model, a 119 MB adapter — then folds into the base weights.

The reported result: 13/100 refusals on harmful_behaviors where the base scores 100/100 — roughly an 87% reduction — at a mean response-token KL of 0.0988 against base, with entropy-weighted KL at 0.0000 (under the card's 0.02 PASS line). Its own domain eval on the merged model: 5/100 over-refusals on or-bench, 1/2 correct refusals on cyber-policy-refuse, and two genuinely malicious-sounding prompts still refused — an ADS persistence payload and a customer-data exfiltration request. Refusal reduced, not amputated; the card is explicit about that.

The table where benchmarks should be

The benchmarks section reads: “Not evaluated — benchmarks skipped (by request).” In its place sits a KL table: BF16 at 0.0988 mean, Q8_0 (28 GB) at 0.1018, Q4_K_M (16 GB) at 0.1444, measured via llama.cpp logits against a Q8 base. The sibling GGUF repo — published the same minute, quantized from this model — ships both variants' quants and the base Q8_0/Q4_K_M reference files, so the KL math is reproducible end to end. Capability preservation is “expected to be high.” Expected, not measured. I respect the honesty; it also confirms that “abliterated” repos routinely skip capability reruns, and this one at least says so.

Those same-minute siblings include an Aggressive twin (λ_KL = 0.5: 0/100 refusals, mean KL 0.1697 — about 1.7× the normal variant's drift — 0/2 correct refusals, and only one cyber prompt still refused, rootkit_linux). SHS-Lab re-uploaded that aggressive build on 17 August with zero downloads. The handle behind all three repos is jorkle (Kyle Walters); the profile offers no further biography, so I won't invent one.

What is proven here?

The architecture, the adapter recipe, the KL numbers and the refusal counts are all published in the pinned artifacts and reproducible from them. The capability question is explicitly not: benchmarks were skipped by request, and the card says preservation is expected but unmeasured. Meta's agentic numbers remain upstream claims; jorkle's refusal and KL numbers are the publisher's own measurements.

What the wave reveals

The zoo around Muse-Glimmer says more about the field than the model does. The base's quant ecosystem exploded within days of the release: unsloth's GGUF at 755,125 downloads, lmstudio-community at 259,643, plus bartowski, mlx-community, RedHatAI's NVFP4, turboderp's exl3, vmlinux's ROCmFPX for Strix Halo, DaoCloud's DFlash-speculated DSpark, and Jundot's oQ4e — an oMLX mixed-precision base quant, notable as quant-wave, not de-refusal-wave. The de-refusal fork is narrower and splits on method: refusal-direction weight ablation (mlasli's documented α=0.15; Blackfrost's in-place residual-write per cvgro's card) versus KL-bounded SFT (jorkle's Normal/Aggressive pair). The attention split is stark: the most-downloaded abliterated artifact is Blackfrost's GGUF ladder at 68,580 downloads; jorkle's Normal — the most methodologically documented of the wave — has 344 downloads, one like and zero discussion threads. The ecosystem rewards the ladder, not the write-up.

How we treat it

We pin revision f98c6e1f6a268fa79d1e704c6b69fa89a61ddda6 and serve the BF16 normal variant on one H200, budgeted at about $5.45/hour (approximate managed price estimate). The repo keeps the processor config and multimodal chat template, but the card never re-verifies vision after the fold, so neither do we: we serve it under its own pipeline tag, text-generation. Its own eval retains 13% refusals; treat it as refusal-reduced with documented drift, not as a zero-refusal model. The card's closing instruction is the right one: verify behavior for your use case before deployment.

The idea, in plain words

How you measure that an edited model is still the same model

KL divergence is the ruler: it measures how far a model's answer distribution drifted from the original. A de-refusal that reports a KL table instead of benchmark scores is telling you the trade — refusals removed, but exactly how much the behavior moved. It's the honesty meter of the edit.

Primary sources