Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

GLM 4.7 Flash is supported in mlx-lm 0.30.3 (h/t Ivan Fioravanti) The 4-bit runs fast (43 tok/s generation, ~800 tok/s prefill) on a base M5 32GB laptop.

Awni Hannun

45,174 subscribers

141,299 views • 9 months ago •via X (Twitter)

Education News & Politics Science & Technology

Anya Rossi• Live Now

Private livecam show

22 Comments

Filip Strand9 months ago

@ivanfioravanti I suddenly have a lot of trouble to find room on my disk for all the shiny new things 😅

Chriscoveries9 months ago

@ivanfioravanti ANE on GPU on m5 still does nothing. Is it just a waste of silicon? Thanks for the test

mzba9 months ago

@ivanfioravanti If it's good at the tool call, it will be a beast for the local agent.

Vox9 months ago

@ivanfioravanti 800 tok/s prefill is wild. does it handle long context well or does it start to choke past 8k tokens?

Jason9 months ago

@ivanfioravanti As a Hugging Face Partner Provider, @novita_labs has already deployed GLM-4.7-Flash, available pay-as-you-go with no monthly subscription required, making it easy for developers to test, build, and scale instantly.

TechMD9 months ago

@ivanfioravanti That’s awesome

Kyle Hessling9 months ago

@ivanfioravanti Getting 55 tokens/second on my M4 Pro 24GB, not much room for context, though. Wishing I had opted for the 36 GB of RAM version...

Thibault Lb9 months ago

@ivanfioravanti Wow nice

Lucas Lain9 months ago

@ivanfioravanti @grok how much you estimate on a 4090 ?

jake ciggy9 months ago

@ivanfioravanti Use btop

Cody Avant9 months ago

@ivanfioravanti Ang hope for 1.58b? 16gb M4 mini.

Kris Krakowiak9 months ago

@ivanfioravanti What do they compare it to gpt-oss-20b instead of gpt-oss-120b?

michael jackson pós-colostomia ➜🤠9 months ago

@ivanfioravanti Can you tell us about the set context size? 800tok/s sounds like a dream, I just wonder if it is at an usable context size

HedgeQuantX9 months ago

@ivanfioravanti On a dedicated server would be much faster 😇

索南9 months ago

@ivanfioravanti I got ~50 TPS with reasoning off. ~33 TPS reasoning on.

Shreya9 months ago

@ivanfioravanti Nice

Mito Rektfarmer9 months ago

@ivanfioravanti Will this work max mini then?

Jithin Palepu9 months ago

@ivanfioravanti that's pretty impressive for a base M5 laptop. 43 tok/s for generation is no joke. can't wait to see what people do with this!

Anees Iqbal9 months ago

@ivanfioravanti Whats the time to first token?

Ramon 🎹9 months ago

@ivanfioravanti thanks for this! do you think it’s useable on a m5 24gb ?

Sebastian Buzdugan9 months ago

@ivanfioravanti those prefill numbers on m5 are surprisingly strong

Erik Beltran 🧙‍♂️9 months ago

@ivanfioravanti rtx 5060ti ,16 gb vram, offload (10 layers gpu) q4 gguf , 25 tks prompt, 8tks

Related Videos

Running GLM 4.7 Flash (8-bit) with Tensor Parallel / RDMA on 2 M4 Pro Mac Minis at 60 tok/sec. mlx-lm 0.30.5 features huge speedups for GLM 4.7 Flash for long context (h/t N8 Programs & Awni Hannun). M5 Pro (~28 Jan) will have ~4x faster prefill and ~1.3x faster decode.

Running GLM 4.7 Flash (8-bit) with Tensor Parallel / RDMA on 2 M4 Pro Mac Minis at 60 tok/sec. mlx-lm 0.30.5 features huge speedups for GLM 4.7 Flash for long context (h/t N8 Programs & Awni Hannun). M5 Pro (~28 Jan) will have ~4x faster prefill and ~1.3x faster decode.

Alex Cheema

56,741 views • 9 months ago

Running GLM-4.7-Flash with OpenCode locally on M4 Max MacBook Pro. 4-bit model runs at 82 tok/sec. Prefill will get ~4x faster with M5 Max MacBook Pro (~28 Jan). EXO will also support disaggregating prefill and decode across devices, e.g. DGX Spark.

Running GLM-4.7-Flash with OpenCode locally on M4 Max MacBook Pro. 4-bit model runs at 82 tok/sec. Prefill will get ~4x faster with M5 Max MacBook Pro (~28 Jan). EXO will also support disaggregating prefill and decode across devices, e.g. DGX Spark.

Alex Cheema

128,404 views • 9 months ago

For even higher throughput and lower latency: batch generation + tensor parallel with mlx-lm + and mlx.distributed. Here it's generating at 63 tok/sec (throughput) with GLM 4.7 in 6-bit and batch size 4 on 4 M3 Ultras:

For even higher throughput and lower latency: batch generation + tensor parallel with mlx-lm + and mlx.distributed. Here it's generating at 63 tok/sec (throughput) with GLM 4.7 in 6-bit and batch size 4 on 4 M3 Ultras:

Awni Hannun

22,721 views • 10 months ago

GLM-4.7 runs quite well on an M3 Ultra with mlx-lm, even at a near lossless precision (6-bit here). It generated the best space invaders game I've seen yet for a local model (even included sound effects!). Generated 6600 tokens and ran at 16 tok/s.

GLM-4.7 runs quite well on an M3 Ultra with mlx-lm, even at a near lossless precision (6-bit here). It generated the best space invaders game I've seen yet for a local model (even included sound effects!). Generated 6600 tokens and ran at 16 tok/s.

Awni Hannun

237,191 views • 10 months ago

Running GLM-4.7-Flash on 4 x M4 Pro Mac Minis using EXO Labs. Uses tensor parallelism with RDMA over Thunderbolt & MLX backend (h/t Awni Hannun). Runs at 100 tok/sec. We're working on optimizing this at EXO Labs. Aiming to hit ~200 tok/sec on this setup soon.

Running GLM-4.7-Flash on 4 x M4 Pro Mac Minis using EXO Labs. Uses tensor parallelism with RDMA over Thunderbolt & MLX backend (h/t Awni Hannun). Runs at 100 tok/sec. We're working on optimizing this at EXO Labs. Aiming to hit ~200 tok/sec on this setup soon.

Alex Cheema

62,320 views • 9 months ago

DeepSeek-V4-Flash powering 4 parallel agents on Pi (by Mario Zechner) 🚀 Running on M3 Ultra at ~30-34 tok/s and 160-187GB peak URAM using MLX-LM. Special shoutout to clandestine.eth 🦇🔊, Pedro Cuenca, Tarjei Mandt, Ivan Fioravanti and others for helping optimize and shape this PR. PR:

DeepSeek-V4-Flash powering 4 parallel agents on Pi (by Mario Zechner) 🚀 Running on M3 Ultra at ~30-34 tok/s and 160-187GB peak URAM using MLX-LM. Special shoutout to clandestine.eth 🦇🔊, Pedro Cuenca, Tarjei Mandt, Ivan Fioravanti and others for helping optimize and shape this PR. PR:

Prince Canuma

110,674 views • 6 months ago

Meet Wally. Our inference stack for open frontier models, built to be the fastest place to run them. Performance snapshot: GLM-5.3 Flash: 380 tok/s. GLM-5.3 Max: 790 tok/s Qwen3.8-27B: 485 tok/s DeepSeek-V4.1 Flash: 615 tok/s 1/7

Meet Wally. Our inference stack for open frontier models, built to be the fastest place to run them. Performance snapshot: GLM-5.3 Flash: 380 tok/s. GLM-5.3 Max: 790 tok/s Qwen3.8-27B: 485 tok/s DeepSeek-V4.1 Flash: 615 tok/s 1/7

RunAnywhere

66,857 views • 1 month ago

dflash-mlx 0.1.8 running on M5 Max and M3 Ultra with Qwen3.6-27B-4bit. Video is normal speed! 🔥 It's fast! And NAX kernel seems using less tokens, it did this on 4/5 tests. M5 Max 2086 tokens | 56.3 tok/s | 64.5% acceptance M3 Ultra 2387 tokens | 64.0 tok/s | 69.2% acceptance

dflash-mlx 0.1.8 running on M5 Max and M3 Ultra with Qwen3.6-27B-4bit. Video is normal speed! 🔥 It's fast! And NAX kernel seems using less tokens, it did this on 4/5 tests. M5 Max 2086 tokens | 56.3 tok/s | 64.5% acceptance M3 Ultra 2387 tokens | 64.0 tok/s | 69.2% acceptance

Ivan Fioravanti

15,823 views • 5 months ago

GLM-5.3-Flash running at up to 300 tok/s single stream, and 1000 tok/s with 8 concurrent sessions. On average it's working at 150 tok/s per stream. Phenomenal model, really something special. On 2x Sparks we get 60 tok/s and 4x 100 tok/s 300B params is the sweet spot.

GLM-5.3-Flash running at up to 300 tok/s single stream, and 1000 tok/s with 8 concurrent sessions. On average it's working at 150 tok/s per stream. Phenomenal model, really something special. On 2x Sparks we get 60 tok/s and 4x 100 tok/s 300B params is the sweet spot.

0xSero

30,672 views • 2 months ago

GLM-5 runs with mlx-lm on a single 512GB M3 Ultra in Q4. It's quite good in my initial testing and pretty fast as well. It generated a highly functional space invaders game using 7.1k tokens at 15.4 tok/s and 419GB memory. Thanks to Gökdeniz Gülmez and Tarjei Mandt for the port.

GLM-5 runs with mlx-lm on a single 512GB M3 Ultra in Q4. It's quite good in my initial testing and pretty fast as well. It generated a highly functional space invaders game using 7.1k tokens at 15.4 tok/s and 419GB memory. Thanks to Gökdeniz Gülmez and Tarjei Mandt for the port.

Awni Hannun

60,694 views • 9 months ago

Apple Silicon + Gemma 4 fans: this is for you. Pico AI Server now supports continuous batching with MLX-Swift. 43 tok/s on 1 stream. 26 tok/s per stream on 2 concurrent streams. That’s 52 tok/s total. a 21% throughput gain on a six-year-old MacBook Pro M1 Max!

Apple Silicon + Gemma 4 fans: this is for you. Pico AI Server now supports continuous batching with MLX-Swift. 43 tok/s on 1 stream. 26 tok/s per stream on 2 concurrent streams. That’s 52 tok/s total. a 21% throughput gain on a six-year-old MacBook Pro M1 Max!

Ronald Mannak

54,999 views • 7 months ago

WIP: First attempt to speed up prefill for Flash-MoE. Original repo did token-by-token without streamed experts. Added: Batched linear attention + batched full attention (Flash Attention style) with custom Metal kernels. Without experts: 6.2x faster prefill (11 -> 68 tok/s) With experts at full-attn layers only: 1.9x faster (11 -> 20.5 tok/s) — same output quality Qwen3.5-397B, 4-bit, 209GB, M5 Max 128GB 1/3

WIP: First attempt to speed up prefill for Flash-MoE. Original repo did token-by-token without streamed experts. Added: Batched linear attention + batched full attention (Flash Attention style) with custom Metal kernels. Without experts: 6.2x faster prefill (11 -> 68 tok/s) With experts at full-attn layers only: 1.9x faster (11 -> 20.5 tok/s) — same output quality Qwen3.5-397B, 4-bit, 209GB, M5 Max 128GB 1/3

Anemll

19,572 views • 7 months ago

Qwen3-Coder-Flash runs quite fast on an M4 Max with mlx-lm. Running the 4-bit here, generated 4,467 tokens at >107 tokens/sec:

Qwen3-Coder-Flash runs quite fast on an M4 Max with mlx-lm. Running the 4-bit here, generated 4,467 tokens at >107 tokens/sec:

Awni Hannun

196,606 views • 1 year ago

GLM 5.3 Flash on 2x DGX Sparks Look at the tok/s at the top :) This is prose. Releasing soon

GLM 5.3 Flash on 2x DGX Sparks Look at the tok/s at the top :) This is prose. Releasing soon

Mia

16,704 views • 2 months ago

The top question I got was does Halo support local LLM running on a Mac. Here's an example of the app using gemma-4-26b-a4b-qat via LM Studio running on a base M5/32GB. It's giving an average of 40 tok/s and in my brief testing is able to execute all tool calls correctly. Very usable with this app.

The top question I got was does Halo support local LLM running on a Mac. Here's an example of the app using gemma-4-26b-a4b-qat via LM Studio running on a base M5/32GB. It's giving an average of 40 tok/s and in my brief testing is able to execute all tool calls correctly. Very usable with this app.

Tanmay

11,682 views • 4 months ago

A perfect coding model for MLX on Apple silicon.. Qwen delivered again. Runs quite fast on an M3 Ultra. Running the 4-bit quantized with mlx-lm:

A perfect coding model for MLX on Apple silicon.. Qwen delivered again. Runs quite fast on an M3 Ultra. Running the 4-bit quantized with mlx-lm:

Awni Hannun

186,831 views • 1 year ago

MiniMax M2.1 in 4-bit cruises on an M3 Ultra with mlx-lm. Generated a space invaders game using 5098 tokens at 47.2 tok/sec:

MiniMax M2.1 in 4-bit cruises on an M3 Ultra with mlx-lm. Generated a space invaders game using 5098 tokens at 47.2 tok/sec:

Awni Hannun

94,165 views • 10 months ago

Same Mac. Same model. Same words. My 512 GB M3 Ultra ran GLM-5.3-Flash on MLX last week: 27 tok/s, first token in 0.46 s. Today it runs TensorFold 0.6.2: 61 tok/s, first token in 0.20 s. In the clip it laps the reply twice while the MLX side is still typing. Same reply, byte for byte. Then eight requests at once, 78 tok/s combined, every answer identical to its solo run. The old engine took one at a time. Context stays at 1,048,576 on the Mac. Every number measured on my own machine with the same probe, one week apart. Nothing estimated. Ash Hart built it. Apache 2.0.

Same Mac. Same model. Same words. My 512 GB M3 Ultra ran GLM-5.3-Flash on MLX last week: 27 tok/s, first token in 0.46 s. Today it runs TensorFold 0.6.2: 61 tok/s, first token in 0.20 s. In the clip it laps the reply twice while the MLX side is still typing. Same reply, byte for byte. Then eight requests at once, 78 tok/s combined, every answer identical to its solo run. The old engine took one at a time. Context stays at 1,048,576 on the Mac. Every number measured on my own machine with the same probe, one week apart. Nothing estimated. Ash Hart built it. Apache 2.0.

Chad Hurley is Volatile Markets

28,183 views • 1 month ago

oMLX 0.7.1.dev1 is out! This dev release adds decision models (Clef, OpenJev), batched prefill for concurrent requests, a refreshed web UI, and faster Qwen3.8, Qwen3.6, GLM-5.3-Flash and DeepSeek V4.1. Speedups in this release - Qwen3.6-35B-A3B decode on M5 Max: 156 -> 205 tok/s (+31%). - GLM-5.3-Flash decode on M3 Ultra: 31.3 -> 40.6 tok/s (+30%). The fused decode kernels now run on M3 and M4, not only M5. - Qwen3.8-Flash-Next with expert offload, 32K prefill on M3 Ultra: 396 -> 636 tok/s (+61%). - Qwen3.8-Flash-Next, 8 prompts arriving 0.4 s apart on M5 Max: time to first token 7.41 -> 5.75 s (-22%) with batched prefill. What's new - Decision models. Clef, Clef-Flash and OpenJev answer typed questions with a probability for each option through the new /v1/systemone endpoint. - Batched prefill. Prompts that prefill at the same time now share one forward pass, so MoE models read their expert weights once. - A refreshed web UI with a log table, a settings rail and a customizable dashboard. - Native video input for Qwen3.5, 3.6 and 3.8. - More accurate oQe quantization, with 11-15% lower KL to the original model. - Fewer memory guard rejections in long sessions. Most of the speedups came from contributors again. Thanks to everyone who sent PRs and tested the dev builds!

oMLX 0.7.1.dev1 is out! This dev release adds decision models (Clef, OpenJev), batched prefill for concurrent requests, a refreshed web UI, and faster Qwen3.8, Qwen3.6, GLM-5.3-Flash and DeepSeek V4.1. Speedups in this release - Qwen3.6-35B-A3B decode on M5 Max: 156 -> 205 tok/s (+31%). - GLM-5.3-Flash decode on M3 Ultra: 31.3 -> 40.6 tok/s (+30%). The fused decode kernels now run on M3 and M4, not only M5. - Qwen3.8-Flash-Next with expert offload, 32K prefill on M3 Ultra: 396 -> 636 tok/s (+61%). - Qwen3.8-Flash-Next, 8 prompts arriving 0.4 s apart on M5 Max: time to first token 7.41 -> 5.75 s (-22%) with batched prefill. What's new - Decision models. Clef, Clef-Flash and OpenJev answer typed questions with a probability for each option through the new /v1/systemone endpoint. - Batched prefill. Prompts that prefill at the same time now share one forward pass, so MoE models read their expert weights once. - A refreshed web UI with a log table, a settings rail and a customizable dashboard. - Native video input for Qwen3.5, 3.6 and 3.8. - More accurate oQe quantization, with 11-15% lower KL to the original model. - Fewer memory guard rejections in long sessions. Most of the speedups came from contributors again. Thanks to everyone who sent PRs and tested the dev builds!

Jun Kim

25,291 views • 1 month ago

I didn't expect DeepSeek v4 PRO (not Flash) to run well on the Mac Studio M3 Ultra with 512GB of RAM. This is 2 bit quantized with the same DwarfStar recipe used for Flash. 433GB GGUF file. 130 t/s prefill, 13 t/s generation. Prefill in the video is low because small prompt.

I didn't expect DeepSeek v4 PRO (not Flash) to run well on the Mac Studio M3 Ultra with 512GB of RAM. This is 2 bit quantized with the same DwarfStar recipe used for Flash. 433GB GGUF file. 130 t/s prefill, 13 t/s generation. Prefill in the video is low because small prompt.

antirez

173,075 views • 6 months ago