Local models on Apple Silicon, benchmarked rather than quoted.
Running a capable model on your own hardware is mostly an argument about memory — how much you actually have, given that 96GB of marketing is 77.76 GiB you can spend, and what you are willing to give up to fit inside it. The benchmarks here are run, not quoted: MLX against llama.cpp on one machine and one set of weights, MXFP8 against Q8 on the same model, speculative decoding measured instead of assumed. Every number arrives with the command that produced it and the hardware it ran on, because a tokens-per-second figure missing either one is not a measurement. Results that contradicted what I expected are left in. Those tend to be the useful ones.
Measured runs on real hardware: runtimes, quantisation formats and the memory ceiling nobody quotes.
We tried to benchmark Nanbeige4.3-33B on a Mac Studio with MLX and it never started: our engine dropped its tool calls, and the 96 GB Studio served at most 75,925 tokens of the 322,851 our Gauntlet needs. We can give you measured speed and memory, a fit table by Mac size, and Particle-1 measured against two DeepSeek V4.1 Flash snapshots.
Our own 4-bit DWQ build of the official Qwen3.8-Flash-Next weights ran on one M5 Max Mac Studio in our MLX engine and scored 0.6932 on our Gauntlet sb-7.2 coding benchmark. Qwen3.8 Flash on OpenRouter has 0.6846 on the same board, and I explain why that comparison isn't settled. This is what we built to get there, with the numbers, and what went wrong on the way.
We split Xiaomi's MiMo-V2.6-Flash across two Mac Studios over Thunderbolt 5 RDMA, with vision and a 256k context, and ran our agentic coding benchmark on it. It scored 0.6605 against 0.7909 for the same model in Xiaomi's cloud. Before the 3D ceiling it earned 0.8309 to the cloud's 0.838. I go through where the rest went, and the bugs we hit splitting one model over two machines.
Elzi is the little voxel robot who narrates our YouTube videos. She was born yesterday, she's still waiting for a salary, and her first video came down two hours after it went up. Here's how the two of us built her in five days on a local TTS model, what's still wrong with her, and why we want your ears.
Qwen3.8 Flash beat Qwen3.8 Max Prime on an Atlassian Forge app built with no internet, 0.7961 to 0.6953, and lost to it on a payments web app, 0.6846 to 0.8364. Every check behind both results, the one manifest line that decided Forge, a probe in which neither model showed it knew the rule until it was handed the docs sentence, and what fine-tuning our own 27B taught us about forgetting.
Claude Opus 5.5 vs Sonnet 5.5: Anthropic's prices, context window, effort levels and release dates side by side, then our two coding benchmarks. Sonnet edged Opus on a payments app, 0.7984 to 0.7926, a margin I read as level. On an Atlassian Forge app Opus won, 0.9767 to 0.8649 after a 15 October re-score. GPT-6.1 Sol and Sol Pro ran on both boards, and the full bill for fine-tuning our own 27B model is at the end.
v0.5 of the Qwen3.8-27B we fine-tune for Forge, Jira, Confluence and JSM is out. It beat v0.4 on every gate we measure, with no waiver needed this time, using a training method that fixed the one thing capacity kept breaking. GGUF versions for llama.cpp and Ollama, Q8_0 and Q6_K, are now published alongside the existing MLX release, parity proven across three separate measured links plus a served-model probe.
Converting a fine-tuned MLX model to GGUF for llama.cpp is one merge script and one convert_hf_to_gguf.py away. Nothing in that path tells you the GGUF actually matches the model you fine-tuned. This is the three-link method we used to prove it for a 27B model, with the exact commands, the real KL-divergence numbers, and the flag-collision bug that silently broke the measurement the first time we ran it.
We're going to Atlassian Team '26 Europe in Amsterdam on 6-8 October, no sponsor booth, no speaking slot, just the two of us. Here's what we'd actually talk about: 60,000+ users migrated at a 100% delivery rate, six open-source migration toolkits anyone can take from GitHub, and thirteen published model weights on Hugging Face from training we've written up start to finish, adapters and all.
Qwen3.8-27B and Qwen3.5-9B were taught Forge, Jira and Confluence on a single Mac Studio with 96 GB. This is the whole procedure with the seven scripts that do it, printed in full: a pinned MLX environment with two patches, an 8-bit base with its speculative-decoding head kept as a sidecar, a chat-JSONL dataset built from a folder of Markdown, a segmented trainer with full-state checkpoints, an adapter expander, a per-module merge into the 8-bit shards. The kit was run end to end for this page on the 9B, in 21 minutes, and its real output follows each step, including the one where the validation loss turns the wrong way.
Three days after v0.3, the Qwen3.8-27B we teach Forge went from 12 of 25 compiling apps to 21 of 25. On the way our own release rule rejected it, correctly by its own text, over a 40-character window repeated 13 times at 32k context. The window was a line of box-drawing characters in an ASCII diagram. Here is what changed, what the metric got wrong, and what we changed so it cannot happen the same way again.
Qwen3.8-27B knows almost nothing about Atlassian Forge: ask it for an app and it invents a manifest format and imports packages that do not exist. Over four days on one Mac Studio I trained it until 14 of 25 manifests passed Atlassian's own validator and 12 of 25 apps compiled, from zero. The weights are public. Here is what it is, what it is not, and why I think it is a first.
Our port reclaim logged a warning every single time it ran in the shipped app, and nobody noticed for weeks — because the warning branch was the only branch that could execute there. macOS keeps lsof in /usr/sbin, and a packaged app's PATH does not have it.
MCP went stateless on 2026-07-28. I probed ten real servers: none implement it. Then I measured the thing that actually costs you — 145 tool definitions, 40,784 tokens, and the 30% that is one server repeating itself.
Metal will not give you the RAM on the box, and the number it does give is not the 75% everyone repeats. I measured the three ceilings on a 96 GB Mac Studio, then measured what modern hybrid-attention models actually spend against them — including a Gemma 4 cache that quietly holds three times its own sliding window.
Multi-token prediction is merged in llama.cpp and still an open PR in mlx-lm. I measured both on an M3 Ultra with the same model. Every default MTP setting was slower than no MTP at all, and the runtime that deletes the MTP head outright was still the fastest thing I ran.
Ant Group's 124B/5.1B-active hybrid-linear MoE hit Hugging Face on 2 August. The memory arithmetic fits a 96 GB Mac with room to spare, and mlx-lm still refuses it. I counted exactly which tensors block it — 385 of 62,237 — then measured what that sparsity actually buys on the models that do run.
I quantised real Qwen3-Coder weights both ways on an M3 Ultra. MXFP8 reconstructs them about 10x worse than 8-bit affine at identical size — and then costs only 1% perplexity end to end. Both numbers are true, and the gap between them is the interesting part.
We built a harness that makes local coding agents produce real software, grades it by running it, and ran the same model as GGUF and MLX. Here's the harness, its modes and archetypes — and which build to run.
A deep dive into Qwen3-Coder-Next, the 80B MoE model that brings powerful local development within reach. Learn why this 2026 release matters for developers who value control, predictability, and consistent results.