Looking for release notes for Ollama itself, the model runtime? See Ollama release notes, every recent Ollama version summarized in plain language, or the official list on Ollama's GitHub releases page. This page tracks Ollama Herd, the router that runs Ollama across multiple machines.
0.12.4 · November 2026
A calm, live dashboard, in light or dark.
- Every page opens with its answer. The Overview starts with one line, Healthy, Needs attention or Critical, and the health score, then the figures that matter. Colour now means something: green, amber and red only appear when a machine is online, struggling or failing.
- The Overview is live. Each headline figure (requests per minute, tokens per second, time to first token, error rate, running and queued) has a small chart of its recent history. A wide activity chart shows tokens per second by model, refreshed every 15 seconds, and each machine's card plots its CPU, memory and speed on the same scale as the others, so the odd one out stands out. On a wide screen it uses the whole window.
- Eight tabs became five: Overview, Models, Usage, Health and Settings. Models now holds the per-model trends, Health holds the recommendations, and Benchmarks has its own page. Every old link still works.
- Light and dark. The dashboard follows your system, or pick one under Settings, Appearance. Text contrast is tested in both.
- Smaller fixes: the footer showed v0.1.0 on every release and now shows the real version, the Models summary now follows its time range, and turning on Reduce Motion takes effect without a reload.
0.12.3 · November 2026
Your machines find the router after a reboot.
- A router that starts before the network is up now announces its real address. The router tells your other machines where to find it over the local network, and it used to check its own address only once, when it started. If it started at boot before Wi-Fi or Ethernet came up, it kept announcing an address that only works on the router's own machine, so machines joining automatically could not connect until someone restarted it. It now re-checks every 30 seconds and updates the announcement when the address changes, which also covers a new address from your router or switching networks.
0.12.2 · October 2026
Fewer stalled Claude Code turns on gpt-oss, and one-shot requests spread across your machines.
- A tool call the model garbles no longer fails the turn.
gpt-osssometimes writes a tool call with one stray bracket in it. Ollama then gave up on the whole response, and Claude Code retried and got the same broken call back: one step failed eight times in a row while we recorded the demo video. Herd now repairs the call and works out which tool the model meant from the tools Claude Code offered, but only when exactly one fits. In three Claude Code runs afterwards, 91 requests all completed and 5 garbled calls were repaired. - One-shot requests from one machine no longer all go to one node. Herd kept every request from the same machine and model on the node that answered the first one, to reuse its warm cache. For a batch script or bot sending one-off requests there is no conversation to keep warm, so a machine that joined later got nothing. Now only a request that continues a conversation is kept on its node, and one-off requests spread by load with no setup.
0.12.1 · October 2026
Quieter Codex logs.
- Codex tool-call repairs that work now log at INFO. Local models often call Codex's file-editing tool by the wrong name, and Herd quietly fixes the call so the edit lands. Every Codex session on
gpt-oss:120bneeds that fix, so each one used to leave warnings in the log for something that had gone right. Now only a call Herd cannot fix, which Codex will reject, raises a warning. Verified with Codex 0.164 on release day: four failing tests fixed, every request completed, zero warnings.
0.12.0 · October 2026
Claude Code and Codex work on their newest versions again, and a prompt cut short by the model's context window is now reported instead of answered in silence.
- Codex CLI 0.164 works. Codex 0.164 changed how it sends tools, and older Herd passed none of them to the model: the run read nothing, edited nothing, and still reported success. Herd 0.12.0 reads the new format. Verified with a real
codex execongpt-oss:120b, which ran the tests, patched the code withapply_patch, and re-ran them to green. See the Codex guide. - Claude Code 2.1.220 and newer work. Newer Claude Code sends a system message inside the conversation, which older Herd rejected with a 400 on the first turn of every session. Verified with Claude Code 2.1.311 on a multi-file fix. See the Claude Code guide.
- Silent truncation is now caught. When a prompt is longer than the model's context window, Ollama cuts it and still answers normally. We measured it dropping an entire message and answering from 24 tokens of a 6,000-token prompt. Herd now compares what the model actually read against what was sent, logs the loss, and raises a health card naming the model and its window.
- Lighter install. Code for a coordination platform that was never launched is gone, about 2,000 lines, and with it the
cryptographydependency. Old start-command flags for it still work: they print a warning and are ignored until 0.14.0. - Removed
/api/embed-text, an endpoint that never answered (it returned 404). Text embeddings use/api/embedand/v1/embeddings, as before.
0.11.4 · October 2026
MLX tool calling works again, MLX coding models run up to nearly twice as fast, and a cut-off answer is now reported as one.
- Tool calling on MLX works again. MLX servers configured with
kv_bitsquantized the KV cache from the very first token, and Qwen3-Coder could no longer read its own prompt: asked for the weather in Paris, it called the tool for other cities or not at all. That was 0 of 5 correct, against 8 of 8 now that herd uses mlx-lm's own default and keeps the prompt at full precision. If you run Claude Code or Codex on an MLX model, this is the upgrade to take. - MLX coding models are much faster. mlx-lm moved to 0.32.2. On our M3 Ultra after warm-up, Qwen3-Coder-30B-A3B went from 65 to 124 tokens per second and GLM-4.7-Flash from 45 to 64. The version lives in
setup-mlx.sh, so re-run it after upgrading;uv tool list | grep mlx-lmshows what you actually have. - A truncated answer is now reported as a failure, on every API. When a backend's stream ended early, herd used to tell the client the answer had finished normally, so Claude Code and Codex could act on half a response with no reason to retry. Each API now sends its own real error after the partial text, and a non-streaming request gets a 502. We found this by checking herd for a bug Ollama fixed in 0.40.9.
- A backend that fails mid-response gets an explanation, not a dropped connection. Before, the connection just closed, and on non-streaming requests that could also break the client's next request on the same connection.
- OpenAI function calling with gpt-oss works. Through
/v1/chat/completions, every tool call gpt-oss made was dropped: clients sawfinish_reason: "tool_calls"and no call. Fixed and verified live in both streaming and non-streaming modes. - OpenAI clients now see
finish_reason: "length"when they hitmax_tokens, instead of being told the model chose to stop. - Two herds can share one network.
FLEET_MDNS_SERVICE_TYPEandFLEET_NODE_MDNS_SERVICE_TYPEwere documented but never applied, so a staging herd and a production herd on the same LAN could not be kept apart. They work now. - Windows nodes: pair herd 0.11.3 or newer with Ollama 0.40.6 or newer, which stops orphaned model runners on Windows.
0.11.3 · October 2026
Windows nodes start for the first time, and every dependency that parses untrusted input moved to a patched floor.
herd-nodenow runs on Windows. It registered its shutdown handlers with a call that exists only on macOS and Linux, so on Windows it stopped before sending its first heartbeat and the router showed an empty fleet. That was true of every release before this one. Reported as issue 8 on Windows 11, diagnosed and fixed the same day. If you have a Windows PC sitting idle, it can join the herd now. One thing worth knowing: a Windows node that stops does not deregister yet, so the router keeps it listed until its heartbeat goes stale.herd --versionandherd-node --versionprint the version and exit. They run before anything reads your config or binds a port, so they are safe to use as the check right after installing, even on a machine already serving a fleet. The download page uses it.- The vision embedding server decodes only JPEG, PNG and WebP. It accepts image bytes from any client on your network, and it used to hand them to every decoder Pillow has. Anything else now gets a
400naming the three accepted formats, before a decoder runs. That closes off the EPS path, which shells out to Ghostscript, along with the TIFF, GIF, MSP, MIC and XBM decoders this service never needed. - Patched floors for the three packages that read untrusted input on a path herd actually reaches:
zeroconf(DNS packet parsing, open to anyone on the LAN),python-multipart(reached through transcription uploads), andstarlette, now declared directly because the CORS layer imports it. The full suite passes with all of them pinned exactly at those minimums. - The dashboard pins Chart.js to one version with a Subresource Integrity hash, instead of tracking a floating major with none, so a compromised publish upstream cannot run code in the dashboard's origin.
- Vision embedding traffic now shows up in traces.
/api/embed-imageand the vision branch of/api/embedrecorded nothing, so their latency and failures were invisible to the dashboard and to every health check. Each request now leaves one row, taggedembedandvision-embed. - Tested against the latest release of every dependency, 57 packages forward, with deprecation warnings treated as errors.
pip install ollama-herdalready resolved the newest versions; what changed is that the code is now verified on them. A new vulnerability in any locked dependency also fails the release health check now.
0.11.2 · October 2026
Housekeeping, and a sharp edge removed from the CLI.
- Importing a CLI module no longer injects your whole env file into the process.
load_env_file()ran at import time, so merely importing the module mutatedos.environ. model_load_timeoutnow says which of its two causes fired, instead of always describing eviction.- A throughput runbook, because per-request median tokens per second is unstable by construction on a bandwidth-bound machine. Plus tests for the last two untested modules, and two tests that were asserting nothing.
0.11.1 · October 2026
OpenAI-format requests were missing context protection entirely.
/v1/chat/completionsgot no context protection. It was applied only on the Ollama, Anthropic and Responses paths, so the single most common client format went without it. If you drive the fleet with an OpenAI-format client, this one matters.- An
mlx:model no node hosts is no longer sent to whatever happens to hold the legacy port. - Three more health checks, 49 in total.
duplicate_runnerscatches one model name resident twice under different digests, which an Ollama upgrade can cause, and names the GB you get back. The other two catch a node running MLX while the router has it disabled, and traffic for anmlx:model nothing can serve. - Rate-based health checks now need a minimum sample. A 5% error-rate threshold fired on a single failure out of twenty.
- Ollama 0.40 picks a runner per model, with MLX the default on Apple Silicon, so "served by Ollama" stopped meaning one thing. Herd now reports which runner actually ran.
- GLM-4.7-Flash is confirmed MLX-only on Apple Silicon, and a July note claiming the opposite has been withdrawn. Measured across all-time traces: through Ollama, 89% of attempts never completed. The MLX vs Ollama guide has been corrected.
0.11.0 · October 2026
Multi-turn tool calling works, and herd can see four context problems it used to miss.
- Multi-turn tool calling no longer fails. OpenAI's format encodes tool-call arguments as a JSON string, Ollama's chat API expects an object, and a client replaying its own history sent the string straight back. Ollama rejected the whole request, so every follow-up turn after a tool call failed. If you tried an agent against the fleet and it died on turn two, this was why. Fixed by @pettersandvand in PR #5, our first outside code contribution.
- Four context-window conditions are now visible, prompted by a field report from someone running an unrelated fleet. They are the kind of problem that shows up as "it got slower" with every dashboard green.
- A per-model context override now survives. Sending no
num_ctxis not neutral: Ollama fills it from its own default, so the override was being reverted, and reverting it reloaded the model. - The KV-cache bloat check was measuring weights as cache. It estimated weights at Q4_0 rates while Ollama's common default is Q4_K_M, so a
gemma3:27bcame out 3.7 GB light and the missing weight was reported as cache overhead. It now uses the true on-disk size the node already reports, attributes cause per model, and says "at most, and possibly none" about what you would actually recover. Our own sizing guidance carried the same low estimate and has been corrected.
0.10.2 · October 2026
Long-running requests stopped being killed for taking a while.
- Both slot-release paths now watch progress instead of the clock. The reaper's only signal was elapsed time, so on 2026-10-06 it declared a request stuck at 645 seconds while that request was still streaming tokens. A request is now reaped when it stops producing, not when it takes longer than average. The first fix missed a second timeout on the queue worker, which this closes.
- A
num_ctx override cannot applywarning no longer fires for overrides that did apply.
0.10.1 · October 2026
Swap visibility, and 8 GiB per model that nothing was reporting.
- llama-server keeps a prompt cache of up to 8 GiB per loaded model, and nothing counts it. Ollama runs each model in its own upstream
llama-server, which defaults to--cache-ram 8192MiB. Ollama never sets the flag,/api/psexcludes it, and so did our own estimates. On a Mac mini agemma3:27breporting 17.7 GB had a 29 GB footprint, 7.8 GB of it cache, with 0 hits in 100 lookups. Measure your hit rate, and if it is near zero,LLAMA_ARG_CACHE_RAM=0took that model to 21.8 GiB. Now written up on the memory ceiling guide. - Heartbeats report swap and compressed memory, with a
swap_usagehealth check, so a node that is quietly swapping shows up before you go looking for it. - A timed-out image render is now actually stopped rather than left running to write its output after cleanup.
- An embedding or rerank request that finds no node says why, retries when it can, and leaves a trace. The reranker's dashboard card is also no longer labelled "undefined".
0.10.0 · October 2026
Two things need action on upgrade. Everything else is drop-in.
- MLX users: re-run
./scripts/setup-mlx.sh. The pinned mlx-lm moved to 0.32.0, which ships--kv-bitsnatively, so the local patch the supervisor needed has been retired. Skip this and MLX auto-start fails withunrecognized arguments: --kv-bits. - Embedding inputs over 2,048 tokens are truncated, where they used to run at up to 8,192. That is the fix for a 28 GB memory high-water mark in the node agent, and it changes results for long documents, so chunk before embedding if you relied on the old limit.
"truncate": falsenow returns 400, matching Ollama.
New:
POST /v1/rerank, which Ollama has no equivalent for. It is the rerank step of a RAG pipeline (Open WebUI, Dify, LangChain), served from the fastembed server nodes with theembeddingextra already run, so it adds no dependency. Jina/Cohere request and response shape. See the API reference.- herd can finally see a client talking straight to Ollama. Every routing signal is derived from what herd itself dispatched, so a second process on
:11434does not just go unmeasured, it invalidates the arithmetic. One co-located daemon contributed about 27% of a node's load and cost 15% of fleet decode with zero errors and a clean dashboard. Nodes now report any non-herd process holding a connection, naming the pid and full argv, and the remedy is to repoint it at the router rather than kill it. - herd measures its own process memory, including the peak. Heartbeats carried system memory only, which on a 512 GB box is dominated by model weights, so a 20 GB leak inside the agent was invisible in it.
Fixed:
FLEET_HOSTandFLEET_PORTwere silent no-ops. Both were documented as working while an explicit CLI default shadowed the environment lookup. Precedence is now flag, then env var, then default.- macOS memory pressure is read from the kernel now. The old check searched command output for words that output could never contain, so it returned "normal" unconditionally. Everything built on it had never fired, including on a machine sitting at 51 GB of swap. Critical pressure now withholds cold loads for models that are not already resident, rather than refusing requests a node could have served.
0.9.6 · September 2026
Desktop and web chat clients that used to fail now work.
- Browser-based clients can connect, if you let them. Hollama, TypingMind, Page Assist and the web builds of Chatbox and Msty all failed because a web page may only call another origin when that server sends CORS headers. Set
FLEET_CORS_ORIGINSto the origins you trust and they work. It is empty by default, so no middleware is installed at all and an existing router behaves exactly as before. Unlike Ollama, Herd adds no implicit origins of its own: the router listens on the whole LAN, so you name every origin rather than inheriting a list. POST /v1/embeddings, the OpenAI-shaped endpoint. This is what AnythingLLM's Generic OpenAI embedder and Chatbox's Knowledge Base need, and it routes through the same dispatcher as/api/embed, so it inherits fastembed routing and retries. An empty string ininputis rejected rather than skipped, because the backend drops empties and every later index would shift, handing you vectors quietly attached to the wrong inputs.- Strict Ollama clients load the model list again. Enchanted, Ollamac and Reins decode fields that
/api/tagsdid not return; it now sends Ollama's full field set. The digest is synthesized from the model name to satisfy those decoders, and is not a content hash. /api/showandHEAD /, which Ollama's own desktop app calls before every chat and which some clients use to probe that a server is alive.
One thing this does not change: an HTTPS page still cannot call a plain-HTTP router. That is a browser rule about the scheme, not about permission. See desktop and web chat apps.
0.9.5 · September 2026
The fleet survives a reboot, and three failures that used to be silent now say what went wrong.
- Autostart on macOS. The router and node now ship launchd agents, so a restart brings the fleet back without anyone noticing it was gone. A crash respawns; a clean exit stays stopped. This changes how you restart Herd on a Mac: use
launchctl kickstart -kto restart andlaunchctl bootoutto stop, because a plainpkillis respawned within about 30 seconds. - Embedding several images at once works with fixed-batch models. Sending two or more images to a model whose export expects exactly one (such as
dinov2-vit-s14) failed the whole request. Those models are now run one image at a time; models that can batch keep the single fast call. - A context-size override that can never apply is now a health card.
FLEET_NUM_CTX_OVERRIDESonly takes effect on a cold load, so a model already resident at a larger context stayed there indefinitely, silently allocating up to 4x the intended memory. The newnum_ctx_override_inertcheck says so, quantifies the waste, and gives the fix. - Pre-warming large models no longer gives up mid-load. The timeout was 120 seconds, shorter than a cold load of a 65GB model, and the failure message printed nothing after "error:". It now waits up to 15 minutes and logs the error type.
- A documented gotcha: Ollama's
OLLAMA_CONTEXT_LENGTHoverrides the per-request context Herd sends. Setting it too low quietly cut prefix-cache reuse and doubled latency while throughput looked healthy. Pin it to the largest context your fleet needs and set per-model sizes withFLEET_NUM_CTX_OVERRIDES.
0.9.4 · August 2026
You get told the leaderboard exists, and unnamed herds get a spot on it.
- The first-run notice now mentions the leaderboard. It listed only herds that had set a name, which was one of them, ours, while four others reported in unnamed. That looked like people declining to be named. They had not declined: nothing ever told them naming was possible. The notice now says so, next to the opt-out it already carried.
- Unnamed herds are listed under a generated handle like
herd-7f3a2b, derived on our side from a hash of your random install id. It holds a leaderboard position without saying anything about you or your hardware, and setting a name replaces it. Only installs running v0.9.4 or newer are ever listed: older ones were told the leaderboard would never know they existed, so they stay invisible until they upgrade and read the new notice. See the telemetry page. - The dashboard stopped calling it anonymous. A persistent random id that links one day to the next is pseudonymous, and the toggle used to promise "nothing identifies it publicly", which listing would have made false. It now says what it does.
- Being counted but unlisted is not yet a setting. Turning telemetry off entirely is the only way to stay off the board today, and that also stops you contributing to the totals. We would rather say that than imply a finer switch exists; a proper one needs the server to accept the field before any client sends it.
0.9.3 · August 2026
Two fixes for running nodes off Apple Silicon, both reported by someone who actually did it.
- A configured Ollama host is respected again. If you set
FLEET_NODE_OLLAMA_HOST, the router was discarding it and rebuilding the address from the node's own LAN IP. Inside a container that is the Docker bridge address, which the router cannot reach, so every routed request failed while the dashboard displayed the unreachable URL as though it were healthy. The router now prefers an explicitly configured address, then the address the heartbeat actually arrived from, and only then the node's self-report. See running a node in a container. - Node architecture is detected instead of assumed. Every node reported
apple_silicon, including x86 boxes with NVIDIA GPUs, because the field defaulted to it and nothing ever set it. That feeds device-aware scoring, so it was skewing routing rather than just the display. It is detected now, withunknownas the honest fallback.
Both bugs needed a fleet that was not a single co-located Mac, which is exactly what we had been testing on. Thanks to the reporter of issue #1.
0.9.2 · August 2026
Anonymous telemetry, and session affinity you can actually see.
- Anonymous community telemetry, on by default. One daily summary per herd, not per machine: which models ran and how often, error counts by category, and a line per Mac covering chip, memory and runtime versions. No prompts, no hostnames, no file paths. Every field was published on the telemetry page before the code existed, and
FLEET_NODE_TELEMETRY=falsestops it before anything is sent, including on first run. The aggregates come back to you on the stats page. X-Fleet-Affinityon every scored route.matchedwhen a follow-up went back to the node already holding that conversation,newotherwise. Session affinity shipped in 0.9.1 but was invisible from outside; now you can watch it work. It reports the routing decision rather than a backend cache hit.- Real cache numbers where they exist. MLX-backed responses now carry
prompt_tokens_details.cached_tokens, the part of your prompt that was skipped because the prefix was already cached. Ollama cannot measure this, so the field is omitted rather than set to zero: absent means "cannot measure", zero would claim every request missed. See the API reference. - Affinity now yields under load. The bonus decays with queue depth, so a warm node that is backing up stops winning against an idle one. The MLX prompt cache also moved up to upstream's default, which makes a matched follow-up worth more.
- Four more tested models. Qwen3-VL 32B and Muse Glimmer 30B (both vision, both verified end to end on the fleet), plus Qwen3.6 27B and 35B-A3B for coding. The router knows Qwen3-VL is a vision model, so images route to it automatically, and it knows Muse Glimmer is a thinking model, so it gets the larger token budget those need. See tested models.
- Ollama and MLX versions collected per node, so a future "requires Ollama 0.32.7+" is a decision rather than a guess. Dashboard toggles persist across restarts.
0.9.1 · July 2026
An 8th scoring signal, two health checks, and reliability fixes.
- Session affinity, the 8th scoring signal. Multi-turn conversations now stick to the node holding their warm prefix cache, so turn N+1 re-uses the cache instead of re-encoding the whole history. It also keeps a fresh prefill from interfering with other streams on the same node. See the Routing Engine guide.
- Two new health checks. One flags a node whose decode speed has collapsed, the other flags a pin request that cannot physically fit, so both surface before they cost you a slow session.
- Steadier under load. Queue concurrency is now capped at the backend's real admission limit, plus fixes for a node-id shadowing bug that orphaned pins, image requests that could fall back to a text-only model, and a stale dashboard tab that could wedge the router.
No breaking changes and no new dependencies, so upgrading from 0.9.0 is a plain pip install ollama-herd --upgrade.
0.9.0 · July 2026
OpenAI Codex support, image routing, and the first release since 0.7.0.
Nothing between 0.7.0 and this shipped publicly, so upgrading lands 0.8.0, 0.8.1, and 0.8.2 all at once. It contains breaking changes, so read the upgrade notes below before deploying. Soaked for 26 hours across every code change in it: 11,925 requests at 99.85% success, including 302 requests to the new Codex endpoint with zero failures.
- OpenAI Codex runs on your fleet. A native Responses API at
/v1/responses, which is the only protocol current Codex speaks after it dropped Chat Completions in February 2026. Agentic coding is verified end to end: Codex read sources, ran pytest, created a module from scratch, patched files, and reached green on its own. Zero config, because model ids auto-route. Works with both the CLI and the macOS desktop app. See the Codex guide. - No more model map to maintain. The built-in default map is gone. An unmapped
claude-*or Codex model id now routes to the best coding model you actually have loaded, instead of a hard-coded name your machine may never have pulled. Explicit pins still win. - Images route to a model that can see them. On every endpoint. An image-bearing request auto-selects a vision-capable model even when your conversation model is code-tuned. A dropped image used to produce a fluent, specific, wrong answer while every metric reported success.
- Fleet control API.
GET /fleet/limitsreports serving capacity, andPOST /fleet/pinpre-warms a model, optionally blocking until it is actually resident. - Consistent
X-Fleet-*headers on every proxied response, plus per-request strict mode (X-Fleet-No-Fallback) and an optional per-client concurrency cap. - Reliability. Failed requests are now recorded reliably, so dashboard success rates stop hiding failures. Models are sized by what they actually cost in RAM (weights plus KV cache) rather than on-disk weights, which had under-counted a 30B model by more than 5x.
Upgrading from 0.7.0, six things change: legacy single-server FLEET_NODE_MLX_* env vars are removed (migrate to a one-entry FLEET_NODE_MLX_SERVERS array); X-Fleet-Model is retired for X-Fleet-Served-Model; queue-full returns 429 instead of 503; POST /fleet/pin can now refuse a set that cannot co-reside; image requests fail loudly rather than being answered by a blind model; and backend 4xx errors surface as themselves instead of an opaque 500.
0.7.0 · June 2026
Native text embeddings and a full-fleet dashboard.
- Native text embedding server.
nomic-embed-textnow runs on a dedicated fastembed server (port 11439) via ONNX Runtime, so embedding requests never queue behind LLM inference. On a real fleet: 573 embed requests over 24 hours, 0.0% error rate, no timeouts. - "Node Models" dashboard. The dashboard now shows a card for every backend receiving traffic, Ollama, MLX, native fastembed, and vision embedding, each with live per-model stats.
- 4 new health checks (36 total) covering embedding routing and contention.
0.6.2 · May 2026
Reliability hardening for the trace store.
- Trace-store resilience. Dedicated read connections and periodic WAL checkpoints so heavy dashboard use can no longer stall background trace writes. Verified under a 30-write plus 120-poll burst, WAL peak dropped from 103 MB to 410 KB.
- New health check surfaces trace write failures instead of letting them hide.
- Fixed a daily log-rotation race between the router and node processes.
0.6.1 · April 2026
MLX stability and packaging.
- MLX supervisor hardening. Orphaned
mlx_lm.serverprocesses are now reaped on startup, and a crash-looping server is quarantined instead of restarted forever. brew install ollama-herdnow works cleanly.- Tunable per-model MLX concurrency, and speculative decoding on the dedicated context compactor for faster Claude Code summarization passes.
- Dashboard color semantics fixed so utilization reads as usage, not a warning.
0.6.0 · April 2026
MLX backend and the fix for "Claude Code breaks at 30K tokens."
- MLX backend, first-class alongside Ollama. Run
mlx_lm.serveras a routing target, with multiple MLX servers per node so a single machine can host a main coding model and a dedicated compactor side by side. - Three-layer context management. The systematic fix for long agentic sessions on local coding models: mechanical tool-result clearing, an LLM-based compactor, and a pre-inference cap that returns a clean error instead of wedging the model.
- Tool-use reliability. Tool-schema fixup and malformed-JSON tool-call repair so agentic loops hold together on local models.
Earlier releases
- 0.5.x · April 2026. Routing and dashboard refinements on top of the multimodal foundation.
- 0.4.1 · April 2026. Thinking-model support (auto-detects chain-of-thought models and inflates token budgets), image-generation fixes, and streaming reliability improvements.
- 0.1.0 · March 2025. The first release: mDNS auto-discovery and 7-signal scoring across a fleet of Ollama nodes.
Get the latest
Ollama Herd ships often. Upgrade with one command, or watch the repo for release notes:
pip install ollama-herd --upgrade
# or: brew upgrade ollama-herd
Full technical release notes live in the GitHub releases. New to Herd? Start with the Quickstart.