DeepSeek 92GB Pruned Model Requires 128GB Mac, Multilingual Accuracy Only 65.4%

1. Separating Released Models from Announcements

On November 5, 2024, MLX-Serve author David Dalcu (@ddalcu) posted a tweet (3,229 views) using phrases like Coming soon, Upcoming, and hoping for — none of which had materialized by November 6. However, a model was indeed published on Hugging Face the same day:

ddalcu/DeepSeek-V4.1-Flash-1017-MLX-Serve-iQ-MLX-Pruned192-2.25bpw

(v1.0, 2026-11-05, MIT). Its specification: retains 192 of 384 routed experts per layer (still activating 6 experts per token), pruned experts quantized to 2-bit GPTQ, attention backbone and shared experts kept in MXFP8; includes 1M context, DSpark speculative decoding, vision tower; ~92GB resident / 296GB download , requires mlx-serve 26.11.2+, explicitly labeled "for 128 GB Mac".

Of the three publicly available tiers, the smallest is 92GB resident. The tweet mentions a "more compact iQ-MLX package that will run on 96GB M5 Ultra", but as of November 6 no such repository has been found — it remains a promise.

2. How 92GB Fits into 128GB: Pruning + SSD Offloading

Two techniques combine, neither sufficient alone:

Pruning ≠ quantization. Quantization keeps all parameters but lowers storage precision; pruning deletes half the experts, halving the router's candidate pool. ISTA-DASLab's GSQ-RCO Coder GGUF uses the same idea: 384→192 experts, Transformer weights drop from 138.2GB to 72.12GB, retaining 94.8% on SWE-bench Verified.

Engram tables stay on SSD. ~203GB of the model are two lookup tables; each token reads only a few rows, so they need not reside in RAM. Hence 296GB download but only 92GB resident. The trade-off: the model card's own benchmark shows placing Engram on an external disk costs ~ 27% prefill throughput , so internal SSD is required.

An honest detail: the lossless reference package (288GB resident / 477GB download) carries a note: "We haven't been able to load it ourselves on a 512GB machine." Even the reference tier hasn't been validated on hardware.

3. The 96GB Math Doesn't Add Up

macOS does not give the GPU all unified memory by default; the Metal driver's wired limit is ~75% of total — a 96GB machine defaults to ~72GB usable. Raising it requires commands like sudo sysctl iogpu.wired_limit_mb=81920 (80GB), but this resets on reboot and must be lowered if memory pressure turns yellow.

The tweet suggests setting ioGPU to 90 (90GB limit, leaving only 6GB for macOS). Some real-world tests recommend 80GB (leaving 12–16GB for system), others up to 88GB, but all agree system headroom is mandatory. The tweet's 90GB is 10GB above the conservative 80GB.

Harder constraint: 92GB resident > 90GB limit. Even at the tweet's 90GB limit, the current package doesn't fit. The promised "more compact" package must cut resident memory further — estimated ≤85GB for 90GB limit, ≤75GB for safe 80GB limit (both approximate, still need a few GB for KV cache and runtime). Cutting another 7–17GB from 92GB; quality impact unknown, no numbers provided.

Pricing: Mac Studio M5 Ultra 96GB/1TB starts at ¥46,999 (1.2TB/s bandwidth); upgrading to 256GB adds +¥30,000 (~¥76,999 total). Paying tens of thousands more for an unreleased package with degraded Chinese performance is a bad equation.

4. Quality Metrics Matter More Than Memory

The model card states plainly: "Pruning isn't free. Half the experts gone: code, reasoning, tool use hold up; chatter degrades more, non-English text degrades severely ." It reports top-1 token agreement with the original greedy decoding:

Overall: 85.82% (vs 92.64% for the 256GB 3.0bpw package)

Tool use: 92.0%, Reasoning: 91.9%, Code: 88.2% — acceptable

Chitchat: 79.6%, Multilingual: only 65.4%

This package was calibrated on "code / agent / tool use" data; Chinese falls on the sacrificed side. Therefore the tweet's claim "96–128GB should be everyone's new default" contradicts the model card's own advice: "If your Mac has 256GB, use the 3.0bpw package." The author builds the engine and promotes his format; readers should look at the quality table.

5. Benchmark Breakdown: Format Diff 3.5%, Speculative Decoding 60%

The tweet's chart compares three packaging formats on the same machine ("same machine, one run per bar"), with a note that the Sushi row came from an independent plugin run that morning, unaffected by same-day changes — not a strict same-version control ; machine model unspecified, so not a 96GB preview.

Numbers laid out: with speculative decoding off, three formats yield 70.3 / 71.6 / 69.2 tok/s — spread only 3.5% . With DSpark on: 96.2 / 111.1 / 111.5 tok/s, relative to own baselines +37% / +55% / +61% (approx), spread widens to 15.9%. Prefill at 8.8K prompt: 3,779 / 4,093 / 3,623 tok/s.

All three bars are 192-expert pruned/low-bit versions (GSQ-RCO IQ2_XXS GGUF vs MLX vs Sushi) — comparing packaging formats, not model generations (inferred from chart title). Conclusion reversed: base decoding barely differs; real speed determinant is DSpark speculative decoding implementation and acceptance rate — not choice between GGUF and MLX.

6. Should Your 96GB Mac Follow?

Already have 96GB M5 Ultra: Don't wait for this package. MLX-Serve official tier already runs Qwen3.8-Flash-Next (125B / ~6B active) on 96GB; official speed table shows 175 tok/s with speculative decoding on M5 Ultra 256GB (different machine, don't extrapolate directly). If tinkering, first set ioGPU to 80GB, verify internal SSD speed.

Already have 128GB Mac: Pruned192 can be downloaded, but only recommended for code, agent, tool use — not Chinese long-form or multi-turn chitchat.

256GB / 512GB: Go straight to 3.0bpw (209GB resident), higher quality tier; lossless package unvalidated by official, test small tasks first.

Planning to buy for this: Wait for three things: 96GB package actually released, third-party 96GB tok/s benchmarks, independent Chinese-task evaluations. The tweet author himself wrote "once Qwen releases, this post is obsolete"; last public check (Nov 2) shows Qwen4 still no model card, no API identifier, no repo; official stance only promises 27B tier first, Max tier no weights released in over a year.

One sentence: Fitting in memory ≠ worth being default; the GBs saved by pruning were taken from Chinese and multi-turn dialogue. Will your 96GB wait for a package that hasn't shipped?

Sources (all public, verifiable)

HF: iQ-MLX Pruned192 2.25bpw model card (92GB / quality slices)

HF: Lossless reference package (288GB / 512GB Mac / Engram external disk loses 27% prefill)

Apple China: Mac Studio M5 Ultra 96GB/1TB pricing

MLX-Serve GitHub: internal tier list and speed table

macOS iogpu.wired_limit_mb settings and headroom recommendations

ISTA-DASLab: GSQ-RCO expert pruning method description

Qwen open-weight promise actual boundaries