Skip to main content

Software Sushi JournalWashi Ledger · desk research

The Best Coding Model for the Mac Studio M5 Ultra 512 GB (Nov 2026)

Best pick on today's published evidence: GLM-5.3-Flash at 8-bit with oMLX. Front-end or 3D-heavy work? Pick DeepSeek V4.1 Flash. Full GLM-5.3 when quality matters more than speed.

16 min readby David Bujosa
315GB peak of 512tick: ≈384 GB default limit≈69 GB GPU headroom
Vals SWE-bench Verified
92.0%
High
8-bit decode, measured on this machine
≈40 tok/s
Medium
Prefill, 4K–128K context
≈2,150 tok/s
Medium
MLX peak memory
≈315 GB
Medium
Unified memory
512 GB
High

Software Sushi · the video companion

Watch the evidence unfold.

The recommendation, benchmarks, memory limits and real costs — in an animated, narrated walkthrough.

YouTube only loads when you press play. Watch on YouTube

You have a Mac Studio with the M5 Ultra and 512 GB of unified memory. Which coding model should you run on it?

GLM-5.3-Flash at 8-bit is the best pick on today's published evidence. It scores 92.0% on SWE-bench Verified in Vals AI's harness, ships under a plain MIT license, and has now been measured on this exact machine: about 40 tokens per second at 8-bit, with around 2,150 tokens per second of prompt processing. It fits with room to spare, peaking at about 315 GB.

That is the answer, not the whole story. That 92.0% is measured against the older DeepSeek V4 Flash 0731 (88.8%); DeepSeek V4.1 Flash has no Vals SWE-bench score yet. If your work is front-end or 3D-heavy, pick DeepSeek V4.1 Flash: it wins 15 of 20 shared coding boards and Vals Terminal-Bench 2.1 (74.5% vs 62.9%), and the six-board coding composite is a tie (56.6 vs 56.2). The full GLM-5.3 trades speed for quality, and a faster 4-bit build with multi-token prediction can more than double decode speed. Below: the benchmarks, the setup we would use, and what the $22,299 box really costs next to an API. This is desk research, not a Software Sushi hardware test. Download the full report (PDF) for every figure, source and confidence label.

Why the M5 Ultra's 512 GB changes the answer

The machine

Mac Studio M5 Ultra · 36-core CPU · 512 GB · $22,299 with 1 TB

unified memory installed
512GB
High
memory bandwidth, decides decode
1.2TB/s
High
compute, decides prefill
80GPU cores
High
default GPU wired limit (75%)
≈384GB
Estimate
Apple specs and the configuration report (High). The 384 GB line is our arithmetic: 75% of 512 GB.

Most local coding advice starts with what fits on a 24 or 32 GB gaming card. Half a terabyte changes that. The Mac Studio M5 Ultra has 1.2 TB/s of memory bandwidth, an 80-core GPU, and enough unified memory for 8-bit versions of large mixture-of-experts models. The 512 GB configuration requires the 36-core CPU / 80-core GPU chip, became orderable on October 28, 2026, and costs $22,299 with 1 TB of storage. See Apple's specs and the configuration and price report.

Three numbers decide the experience. Memory decides what fits. An MoE model holds every expert in memory, even though it reads only the active ones per token. Bandwidth decides decode: how fast it writes, token by token. Compute decides prefill: how fast it reads your repository before it answers. Coding agents re-send tens of thousands of tokens on every turn, so prefill is the speed you feel.

One practical detail: macOS lets the GPU wire about 75% of memory by default, about 384 GB on this machine. That is enough for GLM-5.3-Flash at 8-bit. The full GLM-5.3, at a 406 GB measured peak, needs a raised wired-memory limit.

How we rate the evidence

  • HighOfficial spec, a public reproducible leaderboard, or two independent sources that agree.
  • MediumOne independent measurement with its engine, quantization and context published.
  • VendorThe maker's own claim. Shown only when labeled, never used to rank.
  • EstimateOur arithmetic from cited inputs.

Every key number in the report carries a label. High means an official spec, a public reproducible leaderboard, or at least two independent sources that agree. Medium means one independent measurement with its engine, quantization and context published. Vendor means the maker's own claim, shown only when labeled and never used to rank. Estimate means our arithmetic from cited inputs.

We compare models only within one leaderboard. For example, Z.ai reports GLM-5.3-Flash at 84.3 on Terminal-Bench 2.1 using its own harness; Vals AI's independent run scores it 62.9%. Speculative decoding matters just as much for speed: MTP took two DGX Sparks from 14.6 to 26.5 tok/s. Every speculative run below is marked.

The pick: GLM-5.3-Flash at 8-bit

Z.ai released GLM-5.3-Flash on August 26, 2026. It has 320B total parameters, 18B active per token, a one-million-token context window and a hybrid of linear and sparse attention. Maximum output length is less settled: 128K per Z.ai's docs, 163,840 per the Hugging Face card, so we rate that figure Medium. The LICENSE file is plain MIT. Do not confuse it with the full 754B GLM-5.3, which uses a custom license.

On SWE-bench Verified in Vals AI's harness, where the model fixes real GitHub issues, it scores 92.0%. That beats DeepSeek V4 Flash 0731 (88.8%) and Qwen3.8-27B (86.0%). Full GLM-5.3 scores 95.4%, and closed Claude Opus 5 leads at 97.0%. Those are all results from the same harness. DeepSeek V4.1 Flash has no Vals SWE-bench result yet, so the 88.8% row is the older V4 Flash, not V4.1.

SWE-bench Verified · Vals AI harness

Within one board, Flash sits 5 points behind the closed leader

High
  1. Claude Opus 5closed97.0%
  2. GLM-5.3open · 754B95.4%
  3. GLM-5.3-Flashopen · MIT · the pick92.0%
  4. DeepSeek V4 Flash 0731open · not V4.188.8%
  5. Qwen3.8-27Bopen · Apache-2.086.0%
  6. Kimi K2.7 Codeopen78.2%
Read November 16, 2026. Scale starts at 70%. DeepSeek V4.1 Flash has no result on this board yet; the 88.8% row is V4 Flash 0731.

Why 8-bit? You bought 512 GB, so there is no reason to make the default model smaller than it needs to be. The pipenetwork 8-bit MLX conversion is 334.1 GB on disk. In the M5 Ultra community benchmark, MLX peak memory was 312.6–315.4 GB from 1K to 128K context, with a 322.9 GB system wired peak.

Quantization is not free. Unsloth's KLD table reports 92.2% top-1 agreement with BF16 at 4-bit, 86.3% at 3-bit and 78.3% at 2-bit. Those are token-agreement figures, not coding benchmark scores. The MLX builder measured +3.2% perplexity for its mixed 4/8-bit build versus 8-bit, but only +0.1% for 6-bit. There is no public SWE-bench run comparing these quants.

What actually fits in half a terabyte

Memory · what fits in half a terabyte

8-bit Flash peaks at 315 GB, under the ≈384 GB default GPU limit

Medium
  1. Qwen3.8-27B4-bit16.1 GBFits under the default limit
  2. GLM-5.3-Flash oQ4e + MTPmeasured peak179.2 GBFits under the default limit
  3. GLM-5.3-Flash 8-bit MLXmeasured peak315.4 GBThe pick · fits under the default limit
  4. GLM-5.3-Flash 8-bit MLXon disk334.1 GBThe pick · fits under the default limit
  5. Kimi K2.7-Code 1TUD-Q2339 GBFits under the default limit
  6. GLM-5.3 754BMLX 4/8-bit, peak406 GBFits after raising the wired limit
  7. GLM-5.3-FlashBF16641.6 GBDoes not fit one machine
  8. DeepSeek V4 Pro 1.6Tnative862 GBDoes not fit one machine
  9. Kimi K3 2.8Tnative1,560 GBDoes not fit one machine
Measured MLX peaks on 512 GB machines (Medium); disk sizes from model cards and Unsloth tables. The 384 GB line is an Estimate (75% of 512 GB): it leaves ≈69 GB of GPU headroom over the 315.4 GB peak (≈61 GB over the 322.9 GB system wired peak), so a second large model or a very large KV cache needs a raised iogpu.wired_limit_mb.

The interesting options are GLM-5.3-Flash at 8-bit (315 GB measured peak), the full GLM-5.3 at mixed 4/8-bit (406 GB peak), and DeepSeek V4.1 Flash at oQ4e. Qwen3.8-27B is tiny by comparison, about 16 GB at 4-bit, and makes a useful fast helper while the larger model is busy.

Mind the headroom. The 8-bit build peaked at 315.4 GB, under the ≈384 GB default GPU limit, and peak memory rose only 2.8 GB from a 1K to a 128K prompt. That leaves ≈197 GB of the 512 GB installed, but only ≈69 GB of GPU-usable headroom at the default limit (≈61 GB against the run's 322.9 GB system wired peak). Running a second large model or a very large KV cache next to it means raising iogpu.wired_limit_mb. (Our arithmetic, Estimate.)

Some models still do not fit. GLM-5.3-Flash at BF16 is about 642 GB. DeepSeek V4 Pro is about 862 GB natively; Kimi K3 is about 1.56 TB. Kimi K2.7-Code can fit at 2-bit (339 GB), but has no measured M5 Ultra run and scores below Flash on Vals SWE-bench. Capacity alone is not a reason to choose it.

Measured on the M5 Ultra 512 GB: about 40 tok/s at 8-bit

Our earlier draft found no public 8-bit M5 Ultra run. Re-checking the community database found one. On November 14, an oMLX community submission ran GLM-5.3-Flash-MLX-8bit on the M5 Ultra 80-core / 512 GB, oMLX v0.8.0rc1, macOS 27.2, code context, with MTP, DFlash and speculative prefill all disabled.

GLM-5.3-Flash 8-bit, M5 Ultra 512 GB (one community session, Medium confidence)
ContextDecode (tok/s)Prefill (tok/s)Peak memory
1K39.91,240312.6 GB
4K38.52,138313.4 GB
8K38.52,180314.1 GB
16K40.22,180314.2 GB
32K40.12,176314.4 GB
64K37.72,142314.8 GB
128K36.62,105315.4 GB
Measured · GLM-5.3-Flash 8-bit · M5 Ultra 512 GB

Flat from 1K to 128K: decode holds near 40 tok/s, prefill near 2,150

Medium

Peak memory across the run: 312.6–315.4 GB. The full numbers are in the table above.

oMLX community benchmark, November 14, 2026, one submitter. No MTP, DFlash or speculative prefill. 128K time to first token: 62.3 s. Low 1K prefill reflects warm-up.

A 128K-token prompt took 62.3 seconds to first token. Decode slowed by only about 9% from the 16K result to 128K. A second session measured 38.3–40.4 tok/s at 1K–16K. This is still one community submitter on one platform, not two independent labs, so we rate it Medium. It has not been independently reproduced by Software Sushi.

The old 41–62 tok/s range is real, but it is a different build: mixed 4/8-bit on the 256 GB M5 Ultra. MacStories measured 41 at 16K and 28 at a 61K prompt (including reasoning tokens), while a short oMLX run measured 61.7 at 1K. Do not advertise 62 tok/s as the speed of the 8-bit model.

The alternatives: quality, front-end, or speed

DefaultMedium

GLM-5.3-Flash · 8-bit MLX

36.6–40.4tok/s, 1K–128K, plain

  • 92.0% Vals SWE-bench
  • ≈315 GB peak
  • MIT license
Front-end & 3DMedium

DeepSeek V4.1 Flash · oQ4e

66–82tok/s with MTP

Speculative decoding (MTP)
  • Wins 15 of 20 shared boards
  • Vibe Code 84.7% vs 30.8%
  • Composite 56.6 vs 56.2
Maximum qualityMedium

GLM-5.3 754B · 4/8-bit

18.9–24.9tok/s, 1K–4K

  • 95.4% Vals SWE-bench
  • 406 GB peak
  • Needs a raised wired limit
SpeedMedium

GLM-5.3-Flash · oQ4e + MTP

61.5–140tok/s with MTP

Speculative decoding (MTP)
  • ≈177–179 GB peak
  • Prefill ≈2,150–2,300 tok/s
  • Not the 8-bit speed
Small helperHigh

Qwen3.8-27B · dense

48tok/s (MacStories)

  • 86.0% Vals SWE-bench
  • ≈16 GB at 4-bit
  • Apache-2.0

DeepSeek V4.1 Flash for front-end and 3D

The head-to-head across 20 shared coding boards is not a GLM sweep: DeepSeek wins 15, and a six-board composite is effectively tied (56.6 vs 56.2). On Vibe Code Bench it leads 84.7% to 30.8%; on Design Arena 3D, 1339 to 1293 Elo. On Vals Terminal-Bench 2.1 it leads 74.5% to 62.9%. GLM leads Terminal-Bench 4.0 (Artificial Analysis), 32.8% to 26.8%. For repository agents, GLM on today's evidence; if your work is front-end or 3D-heavy, pick DeepSeek. Its M5 Ultra 512 GB oQ4e run measured 66–82 tok/s with MTP and 1,239–1,308 tok/s prefill.

Full GLM-5.3 when quality matters more than speed

The full 754B GLM-5.3 scores 95.4% on Vals SWE-bench, closer to closed Opus 5's 97.0%, and fits at 406 GB peak. It runs at 24.9 tok/s at 1K and 18.9 at 4K, with about 591 tok/s prefill. It needs a raised GPU wired limit and uses a custom license. Useful for difficult tasks; slower for everyday edits.

4-bit with MTP when speed matters

A newer oQ4e-MTP run on the same 512 GB machine measured 138.9 tok/s at 16K, 77.5 at 32K, 76.2 at 64K, 130.5 at 128K and 61.5 at 195K. Prefill held at 2,166–2,299 tok/s and peak memory stayed below 180 GB. A second session reached 140.0 at 16K. That spread is why we quote a range, label MTP and do not claim 140 tok/s as the 8-bit model's speed. Both were submitted on November 15 at 8:30 and 9:04 PM ET; the site shows the GMT date, November 16.

Qwen3.8-27B as a small helper

Qwen's dense 27B is Apache-2.0 licensed, scores 86.0% on Vals SWE-bench and leads Flash on Vals LiveCodeBench (84.0% vs 80.5%). MacStories measured 48 tok/s on the M5 Ultra and 1,701 tok/s prefill. It fits easily next to Flash for quick edits and lower-cost background tasks.

M5 Ultra versus M3 Ultra: the biggest win is prefill

Same build, same prompts · MacStories

Decode +54–58%. Prefill 2.5–4.1×. Reading the repo is the real win.

High
  1. Qwen3.8-Flash-Next, 16K×2.53 · prefill tok/sM3 Ultra1,143M5 Ultra2,887
  2. GLM-5.3-Flash, 16K×2.59 · prefill tok/sM3 Ultra428M5 Ultra1,107
  3. Qwen3.8-27B, 6K×4.11 · prefill tok/sM3 Ultra414M5 Ultra1,701
  1. Qwen3.8-Flash-Next (MTP)×1.54 · decode tok/sM3 Ultra70M5 Ultra108
  2. GLM-5.3-Flash 4/8-bit×1.58 · decode tok/sM3 Ultra26M5 Ultra41
  3. Qwen3.8-27B (dense)×1.55 · decode tok/sM3 Ultra31M5 Ultra48
MacStories ran the same oMLX build, model files and prompts on an M3 Ultra 512 GB and an M5 Ultra 256 GB (same chip and bandwidth as the 512 GB configuration). Tokens per second.

MacStories ran the same oMLX build on both generations. GLM-5.3-Flash decode rose from 26 to 41 tok/s; prefill at 16K rose from 428 to 1,107. Qwen3.8-27B prefill rose from 414 to 1,701. A 61K-token GLM prompt took 140 seconds on the M3 Ultra and 62.5 on the M5.

The community 8-bit runs point the same way: 33.7 to about 40 tok/s decode, and 542.5 to 1,457–2,138 prefill at 4K. Those use different oMLX versions, so they show direction, not a controlled test. If you already own an M3 Ultra 512 GB, it remains capable. The M5's gain is mainly reading the repository, not just streaming the answer.

Recommended setup: oMLX, 8-bit MLX, Claude Code

Use oMLX on the Studio. It serves both the OpenAI-compatible and Anthropic Messages APIs, so Claude Code can connect without a translation proxy. The following comes from the oMLX install and tuning guide and its Claude Code launcher; these commands are documented, not a Software Sushi M5 Ultra measurement.

zsh · mac-studio · documented commands, not a Software Sushi run
# Install oMLX
brew tap jundot/omlx https://github.com/jundot/omlx
brew install jundot/omlx/omlx
# Download an 8-bit build (334.1 GB on disk)
hf download pipenetwork/GLM-5.3-Flash-MLX-8bit --local-dir ~/models/GLM-5.3-Flash-MLX-8bit
# Serve on localhost:8000 (OpenAI + Anthropic APIs)
omlx serve --model-dir ~/models
# In another terminal, connect Claude Code
omlx launch claude

The measured run is labeled GLM-5.3-Flash-MLX-8bit, but does not identify a repository revision. We cannot confirm it is byte-identical to the pipenetwork build. The pipenetwork card also reports numerical fixes to mlx-vlm's early glm5_next port (SwiGLU clamp and mHC dtype). Use a runtime that includes them; we could not verify which oMLX release includes every fix.

  • Context: give coding agents 64K or more. The Claude Code launcher enforces at least 48K; Ollama's Claude Code guide recommends 64K+.
  • Reasoning effort: GLM defaults to max. Use low or high for quick edits to avoid spending time on reasoning you do not need.
  • Memory: Flash 8-bit fits under the default GPU limit. Full GLM-5.3 needs iogpu.wired_limit_mb raised; the setting resets on reboot. Leave memory for macOS, cache and other apps.
  • From a laptop: ssh -L 8000:localhost:8000 user@mac-studio, then use 127.0.0.1:8000 as the local endpoint.

Aider, Cline and OpenHands can use the OpenAI-compatible endpoint at http://localhost:8000/v1. Cline recommends Compact Prompt; OpenHands needs at least 22K context. Cursor needs a public HTTPS endpoint and routes through Cursor's servers, so that path is not private or offline. Also note: GLM-5.3-Flash is cloud-only in Ollama. Do not run an Ollama cloud tag and call it a local test. For MLX use oMLX or mlx-vlm, not mlx-lm (which lacks this architecture).

For a small team, batching helps. The conservative 8-bit session measured 85.9 tok/s aggregate at eight simultaneous requests, 2.15× the single-stream rate. Another shorter session scaled much better; we use the conservative figure until reproduced. There is still no published minutes-per-real-task measurement for a Claude Code agent on this machine.

How it compares with GPUs and DGX Spark

GLM-5.3-Flash · single stream · tok/s

The Mac is not the fastest box. It is the one that fits 8-bit.

  1. 2× RTX PRO 6000SGLang NVFP4 + EAGLE · speculative153.5Medium
  2. M5 Ultra 256 GB4/8-bit plain, 1K61.7High
  3. M5 Ultra 512 GB8-bit plain, 1K–128K36.6–40.2Medium
  4. M3 Ultra 512 GB8-bit plain, 1K33.7Medium
  5. 2× ASUS GX10 (GB10)vLLM EXL3 4-bit32.9Medium
  6. 2× DGX SparkvLLM NVFP4 + MTP · speculative26.5–29.7High
  7. 2× DGX Sparkno speculation14.6High
  8. 1× RTX 5090 + RAMexperts offloaded9–10Medium
Different quantizations are not interchangeable: NVIDIA rows use NVFP4, Mac rows MLX. This is serving experience, not a controlled chip benchmark. Each row carries its own label.

Two RTX PRO 6000 cards reached 153.5 tok/s with SGLang NVFP4 and EAGLE, and about 388 aggregate at four streams. Two DGX Sparks reached 14.6 plain and 26.5–29.7 with MTP, reproduced by three groups. The highest two-node figure we found is 32.9 tok/s: Context Studios' own run on 2× ASUS Ascent GX10 (the same GB10 chip) with an EXL3 4-bit build in vLLM, single stream. One RTX 5090 with experts offloaded to system RAM managed 9–10 tok/s. No complete RTX PRO workstation price was available, so we leave it out of the cost table rather than inventing one.

The money question: buy for privacy, not savings

At $0.50 per million output tokens, $22,299 buys 44.6 billion GLM-5.3-Flash API output tokens. That is far more than one local stream can generate in a reasonable ownership period.

Cost per million output tokens · three years

On output tokens alone, the API is roughly 8–50× cheaper

Estimate
Cost per million output tokens over three years (our estimate)
OptionAssumptionsRelative costPer million output tokens
GLM-5.3-Flash API$0.15 input ($0.03 cached) / $0.50 output$0.50High
M5 Ultra 512 GB, 24/7$22,299 + about $613/yr power, 40–62 tok/s≈ $4.10–6.40Estimate
M5 Ultra 512 GB, workday8 h/day, 250 days/yr≈ $17–26Estimate
Claude Opus 5 APIClosed, 97.0% Vals SWE-bench; $5 input / $25 output$25High

$22,299buys 44.6 billion GLM-5.3-Flash API output tokens

352 hof 8× H200 on AWS at $63.296/hr, for the same money

85.9tok/s aggregate at eight simultaneous requests (2.15×)

Our estimate, output tokens only; the worked agent session below counts input too (about 2–22×). Power uses Apple's 385 W maximum and EIA's 18.19¢/kWh; real inference power is lower. One stream, reasoning tokens included. API prices are list prices (High).

The power estimate uses Apple's 385 W maximum and EIA's average residential price of 18.19¢/kWh; real inference power is lower. Token counts include reasoning tokens and assume one stream. The top of that 24/7 range is the measured 8-bit rate of about 40 tok/s. Counting output tokens only, the GLM API is roughly 8–50× cheaper per token. Against Claude Opus 5 at $25 per million output tokens (the model behind the 97.0% score), a heavily used box does cost less per token. The case for the box is code that cannot leave the building, offline work, predictable cost, no provider rate limits, or a team sharing one machine. Renting 8× H200 on AWS at $63.296/hr only makes sense for sustained high-utilization serving; $22,299 is about 352 instance-hours.

A worked agent session, input included

The 8–50× figure counts output tokens only. Coding agents resend the repository context on every turn, so input tokens dominate the API bill. Our assumptions (Estimate): 40 turns, an average 50K tokens of context resent per turn (2.0M input tokens) and 1,500 tokens written per turn (60K output). Local speed comes from the measured 8-bit run (≈2,150 tok/s prefill, ≈40 tok/s decode), with no prompt-cache reuse on the Mac.

One 40-turn agent session (our estimate)
Per sessionGLM-5.3-Flash APIM5 Ultra 512 GB, workday useM5 Ultra 512 GB, 24/7
Cost basis$0.15 in / $0.50 out per million; cached input $0.03$3.79 per hour (3-year amortization, 8 h × 250 days, 385 W)$0.92 per hour (3-year amortization, 24/7, 385 W)
Time—15.5 min prefill + 25 min decode = 40.5 min40.5 min
Cost, no caching$0.30 input + $0.03 output = $0.33 (input is 91%)$2.56 (7.7× the API)$0.62 (1.9× the API)
Cost, 90% of input cached$0.1122× the API5.4× the API

Counting input narrows the gap from 8–50× (output only) to about 2–8× without API caching, or about 5–22× with it. The Mac's fast prefill is why: it reads resent context at ≈2,150 tok/s instead of paying per token. Local prefix caching would shorten the Mac's time further; we have not measured it. Bottom line: the GLM-5.3-Flash API is cheaper either way, by 8–50× on output tokens alone and by about 2–22× once input is counted.

If you do not have an M5 Ultra 512 GB

On the 256 GB M5 Ultra, run GLM-5.3-Flash at 4-bit (about 176 GB measured peak; 41–62 tok/s in plain runs). On an M3 Ultra 512 GB, use 8-bit at about 34 tok/s. On a 128 GB M5 Max, gpt-oss-120b runs at about 79 tok/s. On 16–32 GB GPUs or laptops, choose Qwen3.8-27B or gpt-oss-20b rather than trying to offload Flash. Phones and small edge boxes are better clients than coding-agent hosts.

A small sidebar on a viral comparison

A September post from @MiaAI_lab said the M5 Ultra was slower than two DGX Sparks, 28 vs 39 tok/s. The 28 was MacStories' 61K-prompt number; at 16K it was 41. No independent two-Spark run reaches 39: published runs span 14.6–32.9, where 29.7 is the best NVFP4 + MTP result and 32.9 is Context Studios' 2× GX10 run (same GB10 chip, EXL3 4-bit, vLLM). The comparison was misleading, not because the Mac number was fabricated, but because the setup and context were different. The complete claim-by-claim audit is in our earlier report; it is not the reason for our recommendation.

What we still need to test

The biggest gap is now independent reproduction, not the absence of an 8-bit M5 Ultra result. One community submitter published two sessions with settings; no reviewer or second lab has repeated them. We also lack a measured Claude Code task end to end on this machine, a controlled quant-quality comparison on coding tasks, and clear versioning for the MLX runtime correctness fixes.

Software Sushi can fill that gap: film an M5 Ultra 512 GB running the 8-bit build, repeat the 1K–128K benchmark, then run a real repository task with the timer, throughput and memory on screen. Publish the recipe and raw logs. Until then, no simulated terminal output and no claim that we tested it ourselves.

Frequently asked questions

What is the best coding model for the M5 Ultra 512 GB?

On today's published evidence, GLM-5.3-Flash at 8-bit for agentic coding: 92.0% Vals SWE-bench, MIT, about 40 tok/s measured. DeepSeek V4.1 Flash has not been scored on that board yet, so this could change. If your work is front-end or 3D-heavy, pick DeepSeek V4.1 Flash; for quality over speed, full GLM-5.3.

Does GLM-5.3-Flash 8-bit fit?

Yes. The build is 334.1 GB on disk; the M5 Ultra run peaked at 315.4 GB in MLX (322.9 GB system wired peak). That fits below the default GPU wired limit of about 384 GB, leaving ≈61–69 GB of GPU headroom; raise iogpu.wired_limit_mb for a second large model or a very large KV cache.

Is 41–62 tok/s the 8-bit speed?

No. That range comes from mixed 4/8-bit runs on a 256 GB M5 Ultra. The published 8-bit 512 GB run is about 37–40 tok/s at 1K–128K context. A 4-bit build with MTP is faster again (61.5–140).

Can I use it with Claude Code?

Yes, through oMLX's Anthropic-compatible endpoint and omlx launch claude. These are documented commands, not a measured Software Sushi agent run. GLM-5.3-Flash is cloud-only in Ollama, so use a local MLX runtime.

Will the Mac save money versus the API?

Not at GLM-5.3-Flash API prices: the API is 8–50× cheaper on output tokens and about 2–22× cheaper once input tokens are counted. Buy for privacy, offline work or a shared team box, not for token savings.

The verdict for the M5 Ultra 512 GB

Best pick on today's published evidence: GLM-5.3-Flash at 8-bit with oMLX. If your work is front-end or 3D-heavy, pick DeepSeek V4.1 Flash, and keep full GLM-5.3 for the difficult tasks where waiting longer is worth it. The hardware's real advantage is enough memory to avoid aggressive quantization and enough prefill speed to feed a coding agent a repository. The API is cheaper; privacy is the reason to own the box.

Download the full report (PDF) for every source and confidence label. To run private coding agents on your own AWS account or hardware, see Software Sushi on AWS Marketplace, explore our AI/ML consulting services, or book a discovery call.

Sources and verification52 references, numbered as in the PDF

All pages accessed November 14–16, 2026 (America/New_York). The benchmark, price and oMLX sources behind the main figures were re-checked on November 16, 2026; oMLX submission times are converted to America/New_York.

  1. Apple — Mac Studio technical specifications
  2. MacRumors — Mac Studio with M5 Ultra now available with 512GB for $22,299 (Oct 28, 2026)
  3. Apple Support — Mac Studio power consumption and thermal output
  4. oMLX Community Benchmarks — DeepSeek-V4.1-Flash oQ4e-mtp on M5 Ultra 512GB (Nov 9, 2026)
  5. oMLX Benchmark — GLM-5.3-MLX-mixed-4_8bit-mtp on M5 Ultra 512GB
  6. senn-tech — M5 Ultra: first measured local LLM numbers
  7. llamaperf — M5 Ultra Mac Studio for local LLMs: early measurements (cites MacStories)
  8. Pinggy — Self-Hosting LLMs on the 512GB M5 Ultra Mac Studio: What Actually Fits (Sep 1, 2026)
  9. Dr. Chaos — The 512GB M5 Ultra Mac Studio for Local AI and Security (Nov 6, 2026)
  10. llama.cpp Apple Silicon thread (discussion 4167), as summarised by llamaperf: same llama-bench test on M3/M5 Ultra
  11. MacStories — M5 Ultra Mac Studio Review: The Dream Mac for Local AI Agents (Sep 21, 2026)
  12. LLMBenchmark.ai — SWE-bench Verified (Vals AI) leaderboard, read Nov 16, 2026
  13. Helix — GLM-5.3-Flash on 2, 4 or 8 RTX PRO 6000 GPUs (Sep 14, 2026)
  14. llamaperf — M5 Max 128GB local LLM performance (community reports)
  15. CloudCostDB — What a GPU actually costs on EC2, and when Bedrock tokens win (Nov 6, 2026)
  16. Z.ai — GLM-5.3-Flash: Frontier Intelligence, Flash Cost (launch post, Aug 26, 2026)
  17. Hugging Face — zai-org/GLM-5.3-Flash LICENSE file (MIT)
  18. Z.ai developer docs — GLM-5.3-Flash/FlashX
  19. Digital Applied — GLM-5.3's weights are out. The licence is not MIT (Aug 28, 2026)
  20. LLMBenchmark.ai — LiveCodeBench (Vals AI) leaderboard, read Nov 16, 2026
  21. LLMBenchmark.ai — DeepSeek V4.1 Flash vs GLM 5.3 Flash (87 shared benchmarks, read Nov 16, 2026)
  22. Artificial Analysis — GLM-5.3-Flash vs Qwen3.8 27B (prices, output speed, Terminal-Bench 4.0), read Nov 16, 2026
  23. Vibecoder Journal — Aider Polyglot leaderboard (updated Nov 15, 2026)
  24. measuredlocally.com — M5 Ultra Mac Studio vs four DGX Sparks: the same model (Oct 3, 2026)
  25. Context Studios — Mac Studio M5 Ultra for local AI, with own 2× GX10 measurements (updated Oct 12, 2026)
  26. oMLX Benchmark — GLM-5.3-Flash MLX 8-bit on M3 Ultra 512GB
  27. LLMCheck — GLM-5.3-Flash on M5 Ultra (lists oMLX plain run, 61.7 tok/s at 1K)
  28. Unsloth — GLM-5.3-Flash: how to run locally (quantization KLD table)
  29. DevelopersIO (Classmethod) — GLM-5.3-Flash NVFP4 on 2× DGX Spark with vLLM (Sep 2026)
  30. LimeChain — GLM-5.3-Flash on 2× DGX Spark, vLLM path (GitHub)
  31. Graphometer — GLM-5.3-Flash on one desktop (RTX 5090 + CPU offload)
  32. Hugging Face discussion — 2× DGX Spark vLLM config, 20–23 tok/s with MTP
  33. Apidog — What is GLM-5.3-Flash? (specs, Artificial Analysis speed, list price)
  34. Ollama docs — Claude Code integration (Anthropic-compatible API, 64K+ context)
  35. Ollama library — glm-5.3-flash (cloud tag only)
  36. Cursor forum — local LLM needs a public HTTPS endpoint (staff answer)
  37. ngrok docs — Use a local LLM in Cursor with Ollama
  38. Cline docs — Local models (Ollama, LM Studio, compact prompt)
  39. OpenHands docs — Run local LLMs (LM Studio, Ollama, vLLM, SGLang)
  40. Hugging Face — mlx-community/GLM-5.3-Flash-4bit (190 GiB; mlx-lm lacks glm5_next, use mlx-vlm)
  41. Hugging Face — zai-org/GLM-5.3-Flash model card (supported serving frameworks)
  42. U.S. EIA — Electric Power Monthly, Table 5.3 (average residential price, 2026)
  43. oMLX Community Benchmark — GLM-5.3-Flash-MLX-8bit on M5 Ultra (80c) 512 GB, 1K–128K session, oMLX v0.8.0rc1, no speculative decoding (Nov 14, 2026)
  44. oMLX Community Benchmark — GLM-5.3-Flash-MLX-8bit on M5 Ultra (80c) 512 GB, 1K–16K session with batching (Nov 14, 2026)
  45. oMLX Community Benchmark — GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c) 512 GB, 16K–195K, MTP on (submitted Nov 15, 2026, 8:30 PM ET / Nov 16, 01:30 GMT)
  46. oMLX Community Benchmark — GLM-5.3-Flash-oQ4e-mtp on M5 Ultra (80c) 512 GB, 8K–32K, MTP on (submitted Nov 15, 2026, 9:04 PM ET / Nov 16, 02:04 GMT)
  47. oMLX — product page (OpenAI and Anthropic /v1/messages endpoints)
  48. oMLX source — Claude Code integration (`omlx launch claude`)
  49. jacar.es — oMLX on Mac: install, tune and benchmark a local LLM server (Homebrew install, wired limit, SSH tunnel)
  50. Hugging Face — pipenetwork/GLM-5.3-Flash-MLX-8bit (334.1 GB on disk, perplexity vs 8-bit, runtime fixes)
  51. Hugging Face — orcarouter/GLM-5.3-Flash-MLX (2- to 6-bit OrcaSAQ builds; 6-bit ≈296 GB; runs under mlx-vlm)
  52. Anthropic — Claude pricing (API list prices; Claude Opus 5: $5 input / $25 output per million tokens, listed under legacy models), read Nov 16, 2026