Tijs Teulings 🦑
@tijs.org
iOS developer at DPG Media · making indie apps like @kilowattcharging.com, @kipclip.com · https://openpds.tijs.org · https://obscure.cat · based in The Hague, NL · support my work on https://atprotofans.com/support/did:plc:aq7owa5y7ndc2hzjz37wy7ma
@cocore.dev this also means cocore on mlx is likely viable again 😛 I think most of these optimizations would translate fine, though Mei is single user and I guess cocore is not so that might need some tinkering still.
With the release of Mei 0.5.0 it has become much easier to run things. You can brew install and all the important settings have good defaults now. Also; Ornith 1.5 is no longer the front runner, my text only version of Qwen 3.6 27B beats it in my benchmarks.
github.com/tijs/mei
github.com/tijs/mei
GitHub - tijs/mei: Native Swift/MLX OpenAI-compatible inference server for Apple Silicon
Native Swift/MLX OpenAI-compatible inference server for Apple Silicon - tijs/mei
github.com
September 13, 2026 at 8:20 AM
@cocore.dev this also means cocore on mlx is likely viable again 😛 I think most of these optimizations would translate fine, though Mei is single user and I guess cocore is not so that might need some tinkering still.
With the release of Mei 0.5.0 it has become much easier to run things. You can brew install and all the important settings have good defaults now. Also; Ornith 1.5 is no longer the front runner, my text only version of Qwen 3.6 27B beats it in my benchmarks.
github.com/tijs/mei
github.com/tijs/mei
GitHub - tijs/mei: Native Swift/MLX OpenAI-compatible inference server for Apple Silicon
Native Swift/MLX OpenAI-compatible inference server for Apple Silicon - tijs/mei
github.com
September 13, 2026 at 8:16 AM
With the release of Mei 0.5.0 it has become much easier to run things. You can brew install and all the important settings have good defaults now. Also; Ornith 1.5 is no longer the front runner, my text only version of Qwen 3.6 27B beats it in my benchmarks.
github.com/tijs/mei
github.com/tijs/mei
Reposted by Tijs Teulings 🦑
De feiten liegen niet! De rijkste Nederlanders gingen er in de afgelopen 8 jaar veel meer op vooruit dan de armste!
Laat de rijken en zeker de zeer vermogenden méér bijdragen aan Defensie; haal de bezuinigingen op SZW van tafel!
Bron: Centraal Planbureau (CPB)
Laat de rijken en zeker de zeer vermogenden méér bijdragen aan Defensie; haal de bezuinigingen op SZW van tafel!
Bron: Centraal Planbureau (CPB)
September 11, 2026 at 4:10 PM
De feiten liegen niet! De rijkste Nederlanders gingen er in de afgelopen 8 jaar veel meer op vooruit dan de armste!
Laat de rijken en zeker de zeer vermogenden méér bijdragen aan Defensie; haal de bezuinigingen op SZW van tafel!
Bron: Centraal Planbureau (CPB)
Laat de rijken en zeker de zeer vermogenden méér bijdragen aan Defensie; haal de bezuinigingen op SZW van tafel!
Bron: Centraal Planbureau (CPB)
Reposted by Tijs Teulings 🦑
I'll spare you the hassle: this Xcode does not contain the iPhone Duo Simulator. You'll need to wait for Xcode 27.1 beta. developer.apple.com/documentatio...
Xcode 27 RC Release Notes | Apple Developer Documentation
Update your apps to use new features, and test your apps against API changes.
developer.apple.com
September 9, 2026 at 8:01 PM
I'll spare you the hassle: this Xcode does not contain the iPhone Duo Simulator. You'll need to wait for Xcode 27.1 beta. developer.apple.com/documentatio...
We already peaked, for agentic software engineering, with Opus 4.8 (5 acts like a confused older gentleman most of the time) and Sol 5.6. Or actually with Sonnet and Terra (or Luna at max effort which is my fav now).
Astra is really, really cool but I cannot current trust it for my present day engineering. At least until I have adjusted. Some random thoughts including code samples of what I recovered from my traces. lucumr.pocoo.org/2026/9/7/ast...
Astra for Coding: Why Are We Doing This Again?
Some thoughts on Astra and newfangled long-horizon models.
lucumr.pocoo.org
September 9, 2026 at 7:35 PM
We already peaked, for agentic software engineering, with Opus 4.8 (5 acts like a confused older gentleman most of the time) and Sol 5.6. Or actually with Sonnet and Terra (or Luna at max effort which is my fav now).
I just finished reading Central Station!
Find it on BookHive: https://bookhive.buzz/books/bk_PtnFbtS5vzhY4a0LUoPW
Find it on BookHive: https://bookhive.buzz/books/bk_PtnFbtS5vzhY4a0LUoPW
bookhive.buzz
September 9, 2026 at 9:14 AM
I just finished reading Central Station!
Find it on BookHive: https://bookhive.buzz/books/bk_PtnFbtS5vzhY4a0LUoPW
Find it on BookHive: https://bookhive.buzz/books/bk_PtnFbtS5vzhY4a0LUoPW
I'm basically left with Ornith 1.5 and Qwen 3.6 as viable models for this machine. Since Ornith is text only i also tested a version of Qwen without vision which is a nice boost to memory usage. So if you don't need vision that's a nice option. If you *do* want your local model to have vision..
Mei 0.3.0 is out with a small speed increase and some compatibility fixes github.com/tijs/mei/rel...
Release Mei 0.3.0 · tijs/mei
Re-pins vmlx-swift onto the Mei fork synced with 37 upstream commits, and makes the MLXPress cold-weight tier reachable.
Changed
vmlx-swift re-pinned 91fed8be -> e37d1d59 — the fork's main merged w...
github.com
September 7, 2026 at 2:46 PM
I'm basically left with Ornith 1.5 and Qwen 3.6 as viable models for this machine. Since Ornith is text only i also tested a version of Qwen without vision which is a nice boost to memory usage. So if you don't need vision that's a nice option. If you *do* want your local model to have vision..
Mei 0.3.0 is out with a small speed increase and some compatibility fixes github.com/tijs/mei/rel...
Release Mei 0.3.0 · tijs/mei
Re-pins vmlx-swift onto the Mei fork synced with 37 upstream commits, and makes the MLXPress cold-weight tier reachable.
Changed
vmlx-swift re-pinned 91fed8be -> e37d1d59 — the fork's main merged w...
github.com
September 7, 2026 at 12:20 PM
Mei 0.3.0 is out with a small speed increase and some compatibility fixes github.com/tijs/mei/rel...
I’m stripping the 333 vision tensors from Qwen 3.6 and it feels like I’m rather unqualified to poke the poor models eyes out 🥹 sorry please don’t have your robot friends run me over when they spot me on the streets.
September 6, 2026 at 3:23 PM
I’m stripping the 333 vision tensors from Qwen 3.6 and it feels like I’m rather unqualified to poke the poor models eyes out 🥹 sorry please don’t have your robot friends run me over when they spot me on the streets.
Ornith 1.5 seems to hit a sweet spot for my M1 + Mei (MLX) setup. It’s a fairly large 35B model, but its MoE design only activates about 3B parameters for each token. Its hybrid design also does less full attention than a conventional model, so it doesn’t hammer memory and the KV cache as hard
September 6, 2026 at 11:10 AM
Ornith 1.5 seems to hit a sweet spot for my M1 + Mei (MLX) setup. It’s a fairly large 35B model, but its MoE design only activates about 3B parameters for each token. Its hybrid design also does less full attention than a conventional model, so it doesn’t hammer memory and the KV cache as hard
Reposted by Tijs Teulings 🦑
Frontier AI models are converging on similar intelligence, yet cost per task can still vary 10x-60x across vendors.
That seems unsustainable as models become more substitutable.
A lot of frontier lab economics depend on customers paying huge premiums for increasingly marginal intelligence gains.
That seems unsustainable as models become more substitutable.
A lot of frontier lab economics depend on customers paying huge premiums for increasingly marginal intelligence gains.
September 5, 2026 at 4:10 AM
Frontier AI models are converging on similar intelligence, yet cost per task can still vary 10x-60x across vendors.
That seems unsustainable as models become more substitutable.
A lot of frontier lab economics depend on customers paying huge premiums for increasingly marginal intelligence gains.
That seems unsustainable as models become more substitutable.
A lot of frontier lab economics depend on customers paying huge premiums for increasingly marginal intelligence gains.
Reposted by Tijs Teulings 🦑
5 more days! Filling out this application is extremely good practice in thinking about what you're building and why!
⏰Get your Entrepreneur in Residence applications in by September 9th!
This virtual residency is for anyone who has an impactful tech idea to improve social connection and/or local information environments — but needs the resources, community, and organizational support to make it real!
This virtual residency is for anyone who has an impactful tech idea to improve social connection and/or local information environments — but needs the resources, community, and organizational support to make it real!
Enterpreneur in Residence
A virtual residency for founders with bold ideas for the internet
newpublic.org
September 4, 2026 at 5:17 PM
5 more days! Filling out this application is extremely good practice in thinking about what you're building and why!
And if you just want a nice Mac app instead of messing around with alpha software and command line stuff (no worries I get it) you should checkout osaurus.ai Mei is based on some of their open source work (but Mei is more memory efficient and faster 😛) and it’s quite good and has a cute logo
September 3, 2026 at 8:33 PM
And if you just want a nice Mac app instead of messing around with alpha software and command line stuff (no worries I get it) you should checkout osaurus.ai Mei is based on some of their open source work (but Mei is more memory efficient and faster 😛) and it’s quite good and has a cute logo
If you have a Mac you can just use the passwords app. It’s perfectly fine now. It has the family sharing, the one time token. It works. You don’t need an extra app for this never mind paying a subscription fee. Save that money and donate it to whatever is your local version of the ACLU instead.
The controversy surrounds David Heinemeier Hansson, who has posted strident anti-vax and anti-immigrant views. www.theverge.com/tech/988536/...
1Password wades into a right-wing mess after funding a Linux project
The controversy surrounds David Heinemeier Hansson, who has posted strident anti-vax and anti-immigrant views.
www.theverge.com
September 3, 2026 at 8:23 PM
If you have a Mac you can just use the passwords app. It’s perfectly fine now. It has the family sharing, the one time token. It works. You don’t need an extra app for this never mind paying a subscription fee. Save that money and donate it to whatever is your local version of the ACLU instead.
tldr; you can run Ornith 1.5 35B on a 32GB M1 Mac Studio at 50 tokens/sec and get Luna (or Qwen 3.8 27B) level intelligence at proper speeds. And you can do it with an MLX based inference engine you could embed in your own swift app. So much ground to still discover.
If you want to try this out for yourself, all very alpha, you are welcome to do so the engine is available here github.com/tijs/mei I’m sure I’m still missing a lot of tricks and ideas as well so input is always welcome.
GitHub - tijs/mei: Native Swift/MLX OpenAI-compatible inference server for Apple Silicon
Native Swift/MLX OpenAI-compatible inference server for Apple Silicon - tijs/mei
github.com
September 3, 2026 at 8:11 PM
tldr; you can run Ornith 1.5 35B on a 32GB M1 Mac Studio at 50 tokens/sec and get Luna (or Qwen 3.8 27B) level intelligence at proper speeds. And you can do it with an MLX based inference engine you could embed in your own swift app. So much ground to still discover.
If you were following progress on my little quest getting a decent model working on my M1 Mac with only 32GB ram I have some updates. Phase 1 of github.com/tijs/local-m... is basically done. There I figured out a few viable models and which inference engine to run them on. Llama.ccp ended up..
GitHub - tijs/local-model-bench: Benchmark harness for picking a local LLM + inference engine to power a coding agent (Hermes) on Apple Silicon — built and run autonomously by Claude.
Benchmark harness for picking a local LLM + inference engine to power a coding agent (Hermes) on Apple Silicon — built and run autonomously by Claude. - tijs/local-model-bench
github.com
September 3, 2026 at 8:04 PM
If you were following progress on my little quest getting a decent model working on my M1 Mac with only 32GB ram I have some updates. Phase 1 of github.com/tijs/local-m... is basically done. There I figured out a few viable models and which inference engine to run them on. Llama.ccp ended up..
This is the updated summary after the lastest round of testing btw (small but useful models running locally, M1 32GB) github.com/tijs/local-m...
August 29, 2026 at 11:08 AM
This is the updated summary after the lastest round of testing btw (small but useful models running locally, M1 32GB) github.com/tijs/local-m...
Ok so I’ve now tried three different MLX implementations for a local inference engine and all of them collapse when handed more than a few thousand tokens to <1 token/sec. All in the server wrapper though so I see some potential for making a native swift version to resolve this..
August 29, 2026 at 8:32 AM
Ok so I’ve now tried three different MLX implementations for a local inference engine and all of them collapse when handed more than a few thousand tokens to <1 token/sec. All in the server wrapper though so I see some potential for making a native swift version to resolve this..
@devingaffney.com so ehm.. how viable would a llama based cocore be? bsky.app/profile/tijs...
I’m sure there is both higher quality and speed to be gained from Qwen 3.8 though so that’s what I’m tinkering with currently. Also llama.ccp runs any model much faster than any mlx engine I tried. Factor 5 faster so I’ve given up on mlx until someone shows me how to set things up to approach llama
August 26, 2026 at 8:56 PM
@devingaffney.com so ehm.. how viable would a llama based cocore be? bsky.app/profile/tijs...
Ok so Claude has been hard at work running benchmarks for almost a full week now, with minor prodding from my end here and there. And initial results are in github.com/tijs/local-m...
Summary of that summary. Qwen 3.8 27B, Qwen 3.6 and Ornith 1.5 35B can be run on an M1 Mac with 32GB ram…
Summary of that summary. Qwen 3.8 27B, Qwen 3.6 and Ornith 1.5 35B can be run on an M1 Mac with 32GB ram…
August 26, 2026 at 8:31 PM
Ok so Claude has been hard at work running benchmarks for almost a full week now, with minor prodding from my end here and there. And initial results are in github.com/tijs/local-m...
Summary of that summary. Qwen 3.8 27B, Qwen 3.6 and Ornith 1.5 35B can be run on an M1 Mac with 32GB ram…
Summary of that summary. Qwen 3.8 27B, Qwen 3.6 and Ornith 1.5 35B can be run on an M1 Mac with 32GB ram…
Reposted by Tijs Teulings 🦑
Beste volgers, graag breng ik deze petitie onder jullie aandacht, waarin wordt gevraagd om het behoud van het programma #Zomergasten in 2027.
Tekenen jullie ook (voor zover dat al niet is gebeurd)?
B.v.d.
petities.nl/petitions/la...
Tekenen jullie ook (voor zover dat al niet is gebeurd)?
B.v.d.
petities.nl/petitions/la...
Laat Zomergasten in '27 gewoon doorgaan
Zomergasten stopt in 2027 door bezuinigingen. Daarmee verdwijnt een uniek podium voor diepgaande gesprekken op televisie.
Behoud dit VPRO-programma! Help mij door te tekenen voor het voortbestaan van ...
petities.nl
August 24, 2026 at 5:23 PM
Beste volgers, graag breng ik deze petitie onder jullie aandacht, waarin wordt gevraagd om het behoud van het programma #Zomergasten in 2027.
Tekenen jullie ook (voor zover dat al niet is gebeurd)?
B.v.d.
petities.nl/petitions/la...
Tekenen jullie ook (voor zover dat al niet is gebeurd)?
B.v.d.
petities.nl/petitions/la...
📍 Kräuterhotel & Restaurant Heidejäger
Restaurant · Rotenburg, Lower Saxony, DE
☁️ 14°C
★★★★
Breakfast before the last leg home
#BeaconBits
Restaurant · Rotenburg, Lower Saxony, DE
☁️ 14°C
★★★★
Breakfast before the last leg home
#BeaconBits
Kräuterhotel & Restaurant Heidejäger | Beacon Bits
Beacon dropped by @tijs.org
beaconbits.app
August 22, 2026 at 7:22 AM
📍 Kräuterhotel & Restaurant Heidejäger
Restaurant · Rotenburg, Lower Saxony, DE
☁️ 14°C
★★★★
Breakfast before the last leg home
#BeaconBits
Restaurant · Rotenburg, Lower Saxony, DE
☁️ 14°C
★★★★
Breakfast before the last leg home
#BeaconBits
Anything you feel should be in this list for sure?
Current crop of models under test is LFM2.5-2.6B, LFM2.5-8B-A1B, Qwen3.8-27B, Qwen3.5-9B, Ternary-Bonsai-27B, Ornith-1.5-35B-A3B, Muse-Glimmer-30B, Qwen3-Coder-30B-A3B, Laguna-XS-2.1 verified against hosted GPT 5.6 Luna for reference
August 20, 2026 at 8:53 AM
Anything you feel should be in this list for sure?
Will be releasing results and benchmarks at some point when I’m reasonably sure it’s not donn in g anything really stupid anymore, not there yet
In the search for the perfect model fit for my Hermes agent on a Mac Studio M1 Max with 32GB I have given up trying to feel the vibes and am now building and running full on benchmarks for about 10 models that might fit. Each runs on MLX, llama.ccp/gguf and when available DFlash 2 with several tests
August 20, 2026 at 8:49 AM
Will be releasing results and benchmarks at some point when I’m reasonably sure it’s not donn in g anything really stupid anymore, not there yet
In the search for the perfect model fit for my Hermes agent on a Mac Studio M1 Max with 32GB I have given up trying to feel the vibes and am now building and running full on benchmarks for about 10 models that might fit. Each runs on MLX, llama.ccp/gguf and when available DFlash 2 with several tests
August 20, 2026 at 8:45 AM
In the search for the perfect model fit for my Hermes agent on a Mac Studio M1 Max with 32GB I have given up trying to feel the vibes and am now building and running full on benchmarks for about 10 models that might fit. Each runs on MLX, llama.ccp/gguf and when available DFlash 2 with several tests