Ollama: Speculative Decoding Gets a New Draft Architecture
A multi-commit effort from Jesse Gross lands DFlash, a block-diffusion speculative decoding system that reshapes how the MLX runner handles drafting and caches, while a cluster of smaller fixes cleans up terminal UI, config parsing, and API edge cases.
Duration: PT2M34S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-08-08T13:00:36Z
- Audio duration: PT2M34S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
Good morning. It's August 8th, and today's codebase news centers on one big architectural addition and a wave of small reliability fixes trailing behind it.
The headline is DFlash, a new speculative decoding approach merged in PR seventeen-five-seven-one. Unlike the existing one-token-per-call drafting chain, DFlash proposes a whole block of tokens in a single forward pass, borrowing its embedding table and output head from the target model. Getting there took real…
Second theme: input validation and edge-case hardening across the CLI and API surface. PR seventeen-six-two-four fixes a null integration entry in user-edited config that was silently crashing convenience callers. PR seventeen-six-two-three lets Claude Code's context-window suffix pass Ollama's readiness check…
Also notable: PR seventeen-five-nine-three fixes namespaced tool declarations in the Responses API, which were previously being silently dropped, and PR seventeen-six-one-six tightens the tool-call parser so ordinary JSON content stops getting misclassified as a tool call.
What's next: watch for review activity on the Intel SYCL GPU backend, PR seventeen-six-two-one, and the MLX preflight profile in PR…
That's the…
Nearby episodes from Ollama
- Model Coverage and Tool-Call Reliability
- Vision Support Lands in the MLX Runner
- Weekly Recap - Vision Runners and Robustness Cleanup
- Edge Case Cleanup Across the Stack
- Crash Fixes and the Launcher Ecosystem Grows
- Silent Data Corruption Fixes
- Hardening the Thinking and Tool-Call Pipeline
- Closing the Gaps Where Failures Go Silent