primitive-ai/Muse-Glimmer-30B-mixed-NVFP4

Tied with BF16 on accuracy. Two and a half times the throughput. A third of the size.

Mixed-precision quantization of meta-models/Muse-Glimmer-30B to NVFP4,
GPTQ-calibrated: 86.6 against BF16's 86.4 while running
2.5× faster at concurrency 32 and 2.7× smaller, level with the fastest quant of this model and smaller than all of them.

---

Why this quant

• 🎯 Accuracy tied with BF16. 86.6 against BF16's 86.4, pooled over both suites. All five builds span 86.4–86.9, which is one tie at this suite's ±1-point spread.
• ⚡ 2.5× BF16 at concurrency 32, 2.81× single-stream. 1,176 tok/s against 473, and 71.4 against 25.4 in a separate single-stream pass. Per-token latency 27.2 ms against BF16's 67.7 ms.
• 🥇 Smallest of the field, and level with the fastest. RedHat's NVFP4 measured +1.1% against this build here and −5.0% in an earlier sitting; our own repeat of this checkpoint moved 4.6%, so read those two as tied and take the 1.4 GiB. Both are clear of AWQ-INT4 and FP8-block.
• 🔬 Calibrated, not round-to-nearest. Rounding is chosen by GPTQ against a text-only calibration set, and it costs nothing at serving time.
• 👁️ The vision tower is untouched. Vision, embeddings and norms stay BF16. A VLM loses its visual grounding before its words.
• 🧩 Stock vLLM, no patches. Plain compressed-tensors nvfp4-pack-quantized.

---

Serve it

vllm serve primitive-ai/Muse-Glimmer-30B-mixed-NVFP4

Text and image inputs both work. This vLLM build registers no Muse-Glimmer reasoning parser, so if thinking is enabled it arrives inline in content rather than split into reasoning_content.

---

Measured

1,370 items across fourteen public benchmarks. A 1,170-item knowledge suite and a 200-item tool-calling suite, under one fixed protocol for every row: temperature 0.6 / top_p 0.95 / top_k 20, thinking forced on, a 16,384-token budget, concurrency 32, all rows back to back in one sitting on one RTX PRO 6000 Blackwell.

build	size	overall	knowledge	call	abstain	runs k/a	finished	out/answer	tok/s @ 32	per-token latency
BF16 (reference)	55.5 G	86.4	88.8	75.6	57.5	1/1	99.9%	714	473	67.7 ms
RedHat FP8-block	32.1 G	86.5	88.7	78.1	55.0	1/1	99.7%	760	787	40.6 ms
cyankiwi AWQ-INT4	22.4 G	86.9	89.3	76.9	55.0	1/1	99.9%	697	1024	31.3 ms
RedHat NVFP4	21.8 G	86.5	88.9	77.5	52.5	1/1	99.7%	755	1189	26.9 ms
this repo	20.4 G	86.6	88.8	78.6	54.5	3/5	99.7%	800	1176	27.2 ms
this repo, an earlier sitting	—	—	89.2	—	—	1/1	99.7%	800	1176	27.2 ms

overall is one number over both suites: the 1,170 knowledge and 200 tool-calling items pooled as 1,370, weighted 85.4% and 14.6% by item count. Repeat runs of one checkpoint moved it by about a point, so treat a gap below 1.0 as a tie.

Where the tool-calling half goes: call vs abstain

build	agentic	call	abstain
BF16 (reference)	72.0	75.6	57.5
RedHat FP8-block	73.5	78.1	55.0
cyankiwi AWQ-INT4	72.5	76.9	55.0
RedHat NVFP4	72.5	77.5	52.5
this repo	73.8	78.6	54.5

call is accuracy on the 160 rows that require a call; abstain is the 40 whose correct action is to call nothing. Benchmarks: BFCL v4 (30, including irrelevance), xLAM/APIGen (45), ToolACE (45), Glaive v2 (40), nvidia When2Call (40, the abstention rows). Tool schemas go in the system prompt and the model answers with a JSON array of calls, the same way for every model. An item counts only if every expected call is present with acceptable arguments and no call is invented.

Abstention is the weak axis for every model we have measured (52–82%), so a build can look strong on overall and still over-call.

How to read this. overall spans 86.4–86.9 across the five builds, a 0.5-point range on a suite whose repeat spread is about a point, so that column is a tie rather than a ranking. What is outside noise is size: this build is the smallest of the five at 2.49× BF16 aggregate throughput. RedHat's NVFP4 posts 1,189 tok/s against this build's 1,176, but the same checkpoint re-measured in an earlier sitting posted 1,233, a 4.6% swing on identical weights, so read those two as tied on speed and separated by 1.4 GiB on disk.

Tool calling holds up under quantization on this model. Every quant here matches or beats its BF16 on the tool-calling half, and the spread across all five is inside the ±1.5-point repeat band for that suite, so the pooled column is carrying no hidden regression.

---

Comparable with our other models

Accuracy numbers move for reasons that have nothing to do with the model: a shorter token budget, a different temperature, or whether the model was allowed to reason at all. So every number in this table, on this card and on our other cards, comes from one fixed protocol.

The same 1,370 items: a 1,170-item knowledge suite (MMLU-Pro, ARC-Challenge, HellaSwag, WinoGrande, CommonsenseQA, BoolQ, OpenBookQA, GSM8K, MATH-500) and a 200-item tool-calling suite (BFCL v4, xLAM/APIGen, ToolACE, Glaive v2, nvidia When2Call). temperature 0.6, top_p 0.95, top_k 20, thinking forced on, a 16,384-token budget, no reasoning parser, scoring the last ANSWER: in the reply. Concurrency 32 on one RTX PRO 6000 Blackwell, each model's rows in one sitting. Auto-scored, no LLM judge. Both halves are means of at least three runs per build.

model	shape	size	overall	knowledge	call	abstain	finished	out	tok/s @ 32
Laguna-XS-2.1	31 B MoE	19.3 GiB	81.7	83.8	68.4	73.5	98.9%	1097 tok	1523
Nemotron-3.5-Lightning-30B-A3B	30 B MoE+Mamba	19.2 GiB	87.1	87.9	85.4	70.5	97.9%	1429 tok	2204
Ornith-1.5-35B-A3B	35 B MoE	22.6 GiB	88.7	91.7	74.4	60.0	99.3%	760 tok	1469
Muse-Glimmer-30B (this repo)	30 B MoE	20.4 GiB	86.6	88.8	78.6	54.5	99.7%	800 tok	1176
Qwen3.8-27B	27 B dense	20.7 GiB	88.8	90.4	85.5	54.5	99.7%	651 tok	908
Laguna-S-2.1	110 B MoE	64.0 GiB	84.3	87.1	64.6	81.0	97.3%	995 tok	670
Qwen3.8-Flash-Next	180 B MoE (6 B active)	183.7 GB	90.3	92.2	84.8	56.7	99.5%	686 tok	816

Read overall with finished. overall scores an answer that overran the token budget as wrong, but it cannot say whether the model needed the room or failed to stop; finished and out separate those. A gap under 1.0 is a tie. The tok/s column comes from each model's own sitting and drifts a few percent between sittings, so read it as a bracket.

call and abstain are the tool-calling suite's two halves, reported separately. call is accuracy on the 160 items that require a tool call; abstain is the 40 whose correct action is to call nothing. They used to be pooled into one agentic number, and the pooling misled: a model with ordinary call accuracy and unusual abstention discipline outscored models that are better at actually making calls. Weight them by your own workload's mix.

Per-benchmark detail, both halves

benchmark	Laguna-XS-2.1	Nemotron-3.5-Lightning-30B-A3B	Ornith-1.5-35B-A3B	Muse-Glimmer-30B	Qwen3.8-27B	Laguna-S-2.1	Qwen3.8-Flash-Next
knowledge
mmlu_pro	79.0	82.0	89.5	89.0	89.5	81.0	89.0
math_500	76.0	80.0	88.0	74.0	82.0	83.0	86.0
gsm8k	97.0	96.0	99.0	98.0	98.0	99.0	97.0
arc_challenge	94.0	97.3	98.7	98.0	98.0	96.0	98.7
hellaswag	70.7	78.0	86.7	83.3	82.0	76.7	88.7
winogrande	84.7	89.3	89.3	90.7	89.3	86.0	95.3
commonsenseqa	80.8	83.3	86.7	82.5	84.2	85.8	86.7
boolq	90.0	92.0	92.0	90.0	91.0	92.0	94.0
openbookqa	94.0	99.0	98.0	97.0	97.0	94.0	96.0
tool calling
bfcl_v4	83.3	96.7	76.7	86.7	96.7	90.0	96.7
xlam	64.4	73.3	55.6	66.7	77.8	48.9	82.2
toolace	37.8	71.1	71.1	64.4	68.9	48.9	73.3
glaive	87.5	97.5	92.5	100.0	100.0	85.0	100.0
when2call	75.0	77.5	52.5	55.0	55.0	75.0	55.0

What's quantized to what

tensors	format
language-model MLP and attention projections, lm_head	NVFP4, group 16
vision tower, embeddings, norms, router	BF16

Weights-only; activations are quantized dynamically at runtime by vLLM. Rounding is chosen by GPTQ against a text-only calibration set. The vision tower is not quantized.

---

_{primitive · more models · inference economics for production LLM systems}