Unified Sciences · typography review
Pick the blog's voice
The same excerpt from the LLM post, rendered five ways. Headings, a serif or sans reading body, and the mono used for every number. Thinking Machines pairs GT America (grotesque) with Iowan Old Style (serif) and a Chakra Petch accent; both premium faces are commercial, so C uses free stand-ins with the same feel. Tell me a letter and I'll wire it into every post.
A · CurrentSpace Grotesk · Inter · IBM Plex Mono
The look the design skill flags as AI-default. Here for reference.
Inference
2.34× faster LLM decoding on Gemma-4
Serving one user at a time, Gemma-4-26B decodes at 127 tokens/s on an A100 out of the box. We got it to 297 with the exact same output.
Decoding one token means reading every weight it needs from memory. For Gemma-4 that's 7.36 GB per token, which on the A100's 2.04 TB/s of HBM sets a hard floor near 3.6 ms. The regime is bandwidth-bound, so the only levers that help are reading fewer bytes and taking fewer passes. Switching on the model's own MTP head does the second.
The win grows with reply length
Speculative decoding removes whole weight-read passes, and that saving amortizes over a longer generation. The per-token speedup climbs toward a steady-state ceiling of 2.45×.
2.34×faster than the vLLM default127.4 tok/s → 297.2 tok/s · 3.13 ms/token
B · Keep Space Grotesk, serif bodySpace Grotesk · Source Serif 4 · IBM Plex Mono
Your idea: reuse the slide font for headings, add a serif reading body.
Inference
2.34× faster LLM decoding on Gemma-4
Serving one user at a time, Gemma-4-26B decodes at 127 tokens/s on an A100 out of the box. We got it to 297 with the exact same output.
Decoding one token means reading every weight it needs from memory. For Gemma-4 that's 7.36 GB per token, which on the A100's 2.04 TB/s of HBM sets a hard floor near 3.6 ms. The regime is bandwidth-bound, so the only levers that help are reading fewer bytes and taking fewer passes. Switching on the model's own MTP head does the second.
The win grows with reply length
Speculative decoding removes whole weight-read passes, and that saving amortizes over a longer generation. The per-token speedup climbs toward a steady-state ceiling of 2.45×.
2.34×faster than the vLLM default127.4 tok/s → 297.2 tok/s · 3.13 ms/token
C · Editorial (Thinking Machines style)Archivo · Newsreader · IBM Plex Mono
Free stand-ins for GT America + Iowan Old Style. Grotesque heads, serif body.
Inference
2.34× faster LLM decoding on Gemma-4
Serving one user at a time, Gemma-4-26B decodes at 127 tokens/s on an A100 out of the box. We got it to 297 with the exact same output.
Decoding one token means reading every weight it needs from memory. For Gemma-4 that's 7.36 GB per token, which on the A100's 2.04 TB/s of HBM sets a hard floor near 3.6 ms. The regime is bandwidth-bound, so the only levers that help are reading fewer bytes and taking fewer passes. Switching on the model's own MTP head does the second.
The win grows with reply length
Speculative decoding removes whole weight-read passes, and that saving amortizes over a longer generation. The per-token speedup climbs toward a steady-state ceiling of 2.45×.
2.34×faster than the vLLM default127.4 tok/s → 297.2 tok/s · 3.13 ms/token
D · Technical brand markChakra Petch · Inter · IBM Plex Mono
Chakra Petch (the TM accent face) as the display voice. Squared, engineered.
Inference
2.34× faster LLM decoding on Gemma-4
Serving one user at a time, Gemma-4-26B decodes at 127 tokens/s on an A100 out of the box. We got it to 297 with the exact same output.
Decoding one token means reading every weight it needs from memory. For Gemma-4 that's 7.36 GB per token, which on the A100's 2.04 TB/s of HBM sets a hard floor near 3.6 ms. The regime is bandwidth-bound, so the only levers that help are reading fewer bytes and taking fewer passes. Switching on the model's own MTP head does the second.
The win grows with reply length
Speculative decoding removes whole weight-read passes, and that saving amortizes over a longer generation. The per-token speedup climbs toward a steady-state ceiling of 2.45×.
2.34×faster than the vLLM default127.4 tok/s → 297.2 tok/s · 3.13 ms/token
E · Characterful editorialFraunces · Source Serif 4 · IBM Plex Mono
A variable old-style serif with real personality for display. Most brand-forward.
Inference
2.34× faster LLM decoding on Gemma-4
Serving one user at a time, Gemma-4-26B decodes at 127 tokens/s on an A100 out of the box. We got it to 297 with the exact same output.
Decoding one token means reading every weight it needs from memory. For Gemma-4 that's 7.36 GB per token, which on the A100's 2.04 TB/s of HBM sets a hard floor near 3.6 ms. The regime is bandwidth-bound, so the only levers that help are reading fewer bytes and taking fewer passes. Switching on the model's own MTP head does the second.
The win grows with reply length
Speculative decoding removes whole weight-read passes, and that saving amortizes over a longer generation. The per-token speedup climbs toward a steady-state ceiling of 2.45×.
2.34×faster than the vLLM default127.4 tok/s → 297.2 tok/s · 3.13 ms/token