---
title: "2026.04.28"
date: 2026-04-28
extra:
  version: "2026.04.28"
  build: 12
  change_id: uonuvsqnvwmx
  highlights:
    - text: "Replies stream in token by token, and Stop or Retry is one tap away"
      path: "/conversations"
    - text: "Attach images to a message on vision-capable models"
      path: "/conversations"
    - text: "Anthropic and llama.cpp join the provider types you can connect"
      path: "/settings/providers"
    - text: "Every reply reports its tokens, cost, and timing"
      path: "/conversations"
---

The biggest BYLM release to date: responses stream in live and can be
stopped or retried, messages carry image attachments, two new provider
types join the list (Anthropic and llama.cpp), and every reply reports
what it cost — in tokens, money, and time.

## Live, controllable responses

- Assistant replies stream token-by-token instead of appearing only when
  the model finishes.
- A Stop button cancels a runaway response immediately, keeping whatever
  was produced so far.
- If a response fails mid-stream (network drop, provider error), a Retry
  button on the failed message re-runs the turn — no retyping the prompt.
- You can keep typing while the assistant answers: sends queue up as the
  next turn instead of being blocked, and your draft survives the
  streaming transitions.
- The list auto-scrolls to follow a long reply while you're at the bottom,
  and stops fighting you the moment you scroll up to read something.
- Text you select in a finished message stays selected while new tokens
  stream in.

## Reasoning, visible and controllable

- Reasoning models get a proper thinking tile: expandable mid-stream, full
  message width, flowing prose, with a live seconds/words counter that
  stops the moment the model stops thinking.
- Each bot has a reasoning-effort setting, translated to the matching
  wire-level control for Anthropic, OpenRouter, Ollama, OpenAI, and
  llama.cpp — and "Off" genuinely disables thinking on Ollama and
  llama.cpp instead of being silently ignored.
- The waiting indicator is stage-aware: it tells you whether the backend
  is loading the model, processing the prompt, thinking, or generating,
  each with its own live timer.

## Images in conversations

- On vision-capable models, a paperclip menu attaches images to a message
  — from a file or straight from the clipboard (Ctrl+V / Cmd+V works too).
- Attached images render as inline thumbnails with a tap-to-zoom
  fullscreen viewer, both in the composer and in sent messages.
- Image data is stored encrypted on disk, like the rest of your local
  data, with small previews generated so image-heavy conversations scroll
  smoothly.

## New providers, better model lists

- **Anthropic**: talk to Claude models with your own API key, with
  per-model capabilities (image input, thinking, context length) read from
  the native model catalog.
- **llama.cpp**: point BYLM at your own llama.cpp server as a first-class
  provider type, including reasoning control and per-model detail fetch.
- Providers accept custom auth and request headers, so OpenAI-compatible
  backends with non-standard schemes work.
- The model list was redesigned: modality and capability icons, context
  and pricing at a glance, expandable details, filters by capability,
  multi-word token search, and newest-first sorting.
- Model lists are cached and refreshable, load without blocking the page,
  and the bot form warns inline when the typed model id isn't in the
  selected provider's catalog.

## Know what each reply cost

- Per-message token usage (prompt / completion / thinking), updated live
  during streaming and persisted.
- Per-message cost in dollars when the provider reports it (OpenRouter).
- Conversation totals and a context-window usage bar in the page header.
- A per-message performance breakdown: model load, prompt processing, and
  generation time with tokens-per-second.
- Each response records which provider and model produced it, so
  switching a bot's model later doesn't lose attribution on past turns.

## Richer message rendering

- LaTeX rendering for `$...$` and `$$...$$` blocks.
- Syntax-highlighted fenced code blocks with a working Copy button (even
  mid-stream), and unlabeled code fences render instead of erroring.
- A copy button on every message.
- Markdown tables render and scroll smoothly instead of freezing the page.
- On wide screens, messages stay in a comfortable reading column.

## Reliability and polish

- A migration framework runs one-time data upgrades behind a progress
  screen on first launch after an update, instead of revealing a
  half-migrated database.
- Image decryption moved to a hardware-accelerated backend — image-heavy
  conversations open without multi-second freezes.
- Navigation cleanups throughout: Back after saving a provider, bot, or
  tool returns to the list; switching sections from the drawer resets the
  stack; deleting a conversation no longer leaves a stale page behind.
- Providers and tools can be deleted straight from their detail pages,
  bots gain a one-tap "Start conversation" shortcut, and bot parameters
  accept precise numeric input alongside the sliders.