JustVugg/colibri

Website · Discord · English · 简体中文 · 繁體中文 · Italiano

Tiny engine, immense model. Run frontier MoE models — 744B to 2.8T parameters — on consumer and heterogeneous hardware, in pure C with zero engine dependencies, by treating storage, RAM, and VRAM as a single inference hierarchy (AI memory multitiering).

Nine families run today: GLM-5.2/5.3 (744B), GLM-5.3-Flash (321B, with vision), Inkling (975B), Kimi K3 (2.8T), DeepSeek V4 Flash (284B), DeepSeek V4.1 Flash (552B, with vision), Qwen3.8-Flash-Next (125B + 51B n-gram), Qwen3.6 (35B-A3B) and OLMoE (7B) — one C file each, the same coli chat / coli serve / coli web front end. Full roster ↓

Colibrì is an inference engine you can run today, and an open research platform. Its primary goal is to pursue inference-side performance across the entire software/hardware boundary — model formats, memory hierarchy, storage I/O, placement, scheduling, kernels, speculation, and CPU/GPU overlap — so large models depend less on scarce hardware and cost less to run.

Colibrì treats VRAM, RAM, and storage as a single multitier hierarchy, and it is deliberately a place to test aggressive systems ideas — so there is no SLA on speed, and a hard guarantee on semantics: experiments must earn their place through reproducible end-to-end measurements, and the default policy never silently changes model precision or router semantics. Insufficient fast memory may reduce speed; it must not quietly redefine the model.

$ ./coli chat
  🐦 colibri v1.12.0 — GLM-5.2 · 744B MoE · int4 · streaming CPU
  ✓ ready in 32s · resident 9.9 GB
  › ciao!
  ◆ Ciao! 😊 Come posso aiutarti oggi?

See it running

The web dashboard (./coli web), redesigned in 1.12.0: a workspace with a dock for the chat, Brio mode, the Brain page and Profiling, in a light or a dark theme. Here Qwen3.6 answering on a CPU box, experts streamed from disk.

Brio mode: the same model, told to stop writing. Give it a document and the only answers it may pick; it reads the probability of each one, generates nothing, and reports an entropy that says when it is not sure. Here: request changes at 99.9%, entropy 0.005, 4 tokens read, 0 generated.

The Brain page, Explore: the measured expert atlas of GLM-5.2 drawn as a cortex. 13,260 characterised experts in ten regions (Python, SQL, mathematics, poetry, law, Chinese…); position is measured routing affinity, not a learned embedding. Choose a region to enter it. Live routing switches to the model actually running: one cell per expert, colour is the storage tier, and every expert routed in a turn flashes white.

Inside the Python region: 1,142 experts as a constellation, each labelled by layer and index. The panel shows one of them, layer 17 expert 178: a generalist with entropy 3.13, whose measured affinity is 20.2% Python, 14.6% JSON, 14.2% conversation, 13.3% SQL.

The Profiling page: where the engine spends each turn, by phase, with the last 30 turns as a trend. Here Qwen3.6 on a CPU box: 19.0 s of wall time for 36 prompt and 55 generated tokens, 2.9 tok/s, 11.4 s of disk service overlapped with compute.

The research mission

With Colibrì, private frontier model access is not limited by availability of hyperscaler-class hardware.

With its multitiering features Colibrì removes proprietary hardware dependencies aggressively optimizing functional inference engine pipelines.

Our operational mission includes changing how weights are represented and moved, deciding what lives in VRAM, RAM, or storage, overlapping heterogeneous compute, reducing launch and synchronization overhead, exploiting sparsity and reuse, and testing new decoding algorithms. Nothing is protected merely because it is conventional; nothing is adopted merely because a microbenchmark looks fast. The deciding result is end-to-end inference on real machines, with correctness and quality measured alongside throughput, latency, memory, and cost.

The practical consequence is accessibility: run a 744B-parameter model on hardware you already own, watch every expert fire in real time, and change the code that does it. Not renting intelligence behind an API — holding it: probing it, measuring it, improving it. The engine is deliberately small enough that the next useful optimization can come from anyone willing to measure it.

Core techniques and measured findings

• One hierarchy, not limited by tier capacity. VRAM, RAM, and NVMe are placement tiers for the same weights; limited fast memory changes speed, not model semantics.
• A JIT for weights. Measured routing heat drives a per-layer LRU, a learned pinned hot-store, and one-layer-ahead prefetch instead of loading every expert. It wins on repeatable workloads; history can overfit, and lookahead can lose on some hosts, so both remain measurable policies rather than promises.
• I/O is part of the engine. Batched expert unions, overlapped reads and compute, O_DIRECT, and weighted dual-SSD striping attack the streaming path rather than pretending storage latency is free. O_DIRECT is drive-dependent, and dual-SSD still needs broader end-to-end community A/Bs.
• Heterogeneous execution. CPU, CUDA, Metal, NUMA memory, and partial or full expert residency share one runtime and can be combined according to the machine; the profitable combination depends on compute, bandwidth, residency, and workload.
• Compressed state without a different model. Token-exact forward validation, 57× smaller MLA KV state, persistent warm conversations, and faithful DSA keep optimization tied to correctness. These are memory, latency, and correctness properties — not a blanket throughput claim.
• Speculation that must earn its keep. Native MTP and grammar-forced drafts are measured end to end and can be disabled when acceptance does not repay verification.

Open hypotheses, experiments, and how to help

Colibrì treats an optimization as a hypothesis until a controlled end-to-end A/B shows otherwise. These are the main questions now:

Want to help? Pick one row and publish the negative results too. Record the hardware, commit, model/container, exact command, prompt, cache state, throughput, TTFT, expert hit rate, bytes read, and quality check; change one variable, repeat the run, and attach raw logs. Start with CONTRIBUTING.md, compare against the benchmark protocol, then open an experiment issue. A well-controlled failure is more valuable here than an unexplained fast number.

The idea

A 744B Mixture-of-Experts model activates only ~40B parameters per token — and only ~11 GB of those change from token to token (the routed experts):

So the model doesn't need to fit in fast memory — it needs to be placed:

• the dense part (attention, shared experts, embeddings — ~17B params) stays resident in RAM at int4 (~9.9 GB);
• the 19,456 routed experts (75 MoE layers × 256 + the MTP head, ~19 MB each at int4) live on disk (~370 GB) and are streamed on demand, with a per-layer LRU cache, a learned pinned hot-store, and an optional VRAM tier.

Think of the core algorithm as a JIT, but for weights. A compiler JIT never compiles the whole program — it watches what actually runs and compiles the hot paths, just in time. colibrì makes the same bet about a 744B parameter space: parameters are not resident state to be held, they are data to be staged across a heterogeneous storage hierarchy (VRAM / RAM / NVMe), exactly when the router proves they are needed. Measured routing heat decides which experts earn which tier, the router runs a layer ahead so prefetch hides the staging latency, and — like a JIT — the engine learns your workload: the more you run, the hotter the right experts get. It works because routing has measurable structure (see the expert atlas) — and structure is cacheable.

The engine is a single C file (c/colibri.c) plus small headers. No BLAS, no Python at runtime, no GPU required.

Local cluster mode

The coordinator keeps token generation, routing, and KV state local while disk-backed expert workers execute routed FFNs on other Macs. A layer's routed batch-union is sent as one persistent TCP request, so a token does not incur one round trip per expert.

Start the optional registration service:

./coli cluster coordinator --host 0.0.0.0 --port 8765

On each worker, with the same converted model available locally:

./coli cluster worker --model /nvme/glm52_i4 --port 9100 \
  --coordinator http://COORDINATOR:8765 --advertise-host WORKER_IP

Run the coordinator with discovery, or provide --cluster-workers HOST:PORT,... for a static setup:

./coli serve --model /nvme/glm52_i4 \
  --cluster-coordinator http://127.0.0.1:8765

The transport is disabled unless workers are configured, so the existing single-machine path remains unchanged. Dense-layer sharding and browser/WebGPU workers are separate follow-up seams.

How it works

The per-token path

Every layer of every token walks the same five steps. The design goal is that placement only ever decides speed — the router's decisions and the weights' precision are the same whether an expert answered from VRAM or from disk.

One memory hierarchy instead of one memory requirement

Dual-SSD: two copies of the model, twice the read bandwidth

Decode is disk-bound on most machines, and expert reads are read-only — so if you have a second SSD, put a full copy of the model on it and let the engine stream from both drives at once:

COLI_MODEL=/fast/glm52_i4 COLI_MODEL_MIRROR=/second/glm52_i4 ./coli chat
COLI_DISK_WEIGHTS=9,3 ...   # optional: primary,mirror bandwidth ratio (else measured at startup)

Each expert is routed to one drive by a deterministic hash, weighted by the two drives' measured (or declared) bandwidth, so readahead/PILOT prefetch and the demand read always hit the same drive and nothing is cached twice. The aggregate bandwidth is the sum of both drives — a 9 GB/s + 3 GB/s pair reads experts ~33% faster than the fast drive alone, and the OMP-parallel pin/warmup load streams from both. Details worth knowing:

• the mirror is validated at startup (per-file size + safetensors header must be byte-identical to the primary); divergent or missing files silently stay on the primary, so a partial mirror is fine — a smaller second SSD holding only some shards still helps;
• the mirror is never written: .coli_usage, .coli_kv and all sidecars stay on the primary;
• a read error on the mirror falls back to the primary (one warning, no crash), so unplugging the second drive mid-run degrades instead of killing the server;
• routing never changes tokens — both copies are byte-identical, and the per-run MIRROR: stats line shows GB served per drive.

The same engine spans the whole range: on a 25 GB laptop everything streams from disk (slow but correct); on a large host the entire expert set becomes resident (CUDA_EXPERT_GB=auto PIN_GB=all) and disk drops out of the decode path entirely. Between the tiers sits a learning cache: the engine records which experts your workload routes to (.coli_usage, updated every turn) and pins the hottest ones automatically — colibrì literally gets faster the more you use it. On multi-socket hosts, COLI_NUMA=1 interleaves the resident weights across memory controllers (#82).

For a second drive that cannot hold the whole model, Colibri can rank a partial mirror from the expert history it already learns. Run a few representative prompts first so .coli_usage reflects the workload, then plan, stage, and verify the mirror:

./c/coli mirror plan  --model /fast/glm52_i4 --mirror /second/glm52_i4 \
  --budget-gib 200 --reserve-gib 20
./c/coli mirror stage --model /fast/glm52_i4 --mirror /second/glm52_i4 \
  --budget-gib 200 --reserve-gib 20
./c/coli mirror verify --model /fast/glm52_i4 --mirror /second/glm52_i4

The planner reads safetensors headers directly, follows split-model directories from COLI_MODEL_DIRS, and prioritizes shards that can serve the hottest routed experts. Staging never changes the primary model: it copies through temporary files, preserves the requested free-space reserve, verifies every shard with SHA-256, never deletes an existing mirror shard, and atomically publishes a receipt only after the selected mirror is ready.

Never wait for the disk twice

Misses are expensive, so the engine spends most of its cleverness avoiding and overlapping them: each expert's three matrices are stored adjacent and read in one pread; a bounded async I/O pool (PIPE=1, default) loads missing experts while resident ones compute; batched positions read each unique expert once (batch-union); and a router-lookahead thread (PILOT=1) prefetches the next layer's experts — routing is measurably 71.6% predictable one layer ahead. On GPUs, the resident pipeline (COLI_CUDA_PIPE=2) keeps the residual stream on-device across layers so the CPU expert loop runs uninterrupted; on Apple Silicon an experimental Metal backend does the batched expert math on the unified-memory GPU; and a Vulkan backend brings the expert tier, dense projections, and the MLA attention core to any GPU with a Vulkan 1.2 driver — including AMD cards via Mesa/RADV (the only backend for cards the vendor stacks no longer support, like the RX 580, and competitive with ROCm on RDNA4 — see the benchmarking notes).

On real NVMe, measure DIRECT=1. O_DIRECT bypasses the page cache and is often a large win on drives with DRAM cache and bandwidth headroom (+34% decode measured with PIPE=1 on a Blackwell/Windows box; 4.25→9.69 GB/s in iobench on a GB10) — but it is drive-dependent: QLC/DRAM-less or virtualised disks can be neutral to negative. Try it first; keep what your hardware rewards.

Faithful model, compressed state

The forward pass is validated against a transformers oracle (teacher-forcing typically 30-32/32; two tiny-oracle positions are floating-point near-ties and toolchain-dependent). MLA attention stores a compressed KV state — 576 floats/token instead of 32,768 (57× smaller) — and persists it across restarts (.coli_kv): conversations reopen warm with zero re-prefill, byte-identical to an uninterrupted session. DSA sparse attention (GLM-5.2's lightning indexer) is implemented faithfully and validated by forcing full-key selection to reproduce dense attention exactly.

Speculative decoding, honestly

GLM-5.2's native MTP head drafts tokens that the main model verifies in one batched forward — 2.2–2.8 tokens/forward when it pays. Two hard-won rules ship as defaults: the MTP head must be int8 (int4 heads collapse to 0–4% acceptance, #8), and draft and verify must compute the same function — SPEC_PIN=1 pins both to one kernel family (#163 is the full forensic story). Grammar-forced drafts (GRAMMAR=file.gbnf) add ~free acceptance on constrained JSON output. Whether speculation is a net win depends on your cache temperature — measure, and use DRAFT=0 when it doesn't pay.

What it achieves

Same engine, same int4 container — the hardware only changes where the experts live. Highlights from the full benchmark tables:

• 6× RTX 5090, full residency: 5.8–6.8 tok/s decode, TTFT ~13 s (experiment log);
• 128 GB CPU-only desktop: ~1.8 tok/s warm (#200);
• single RTX 5070 Ti laptop-class box: 1.07 tok/s via the GPU-resident pipeline (#273);
• 25 GB dev box: 0.05–0.1 tok/s cold — the proven floor where this project started, and still the honest baseline.

Quality is measured, not assumed: the int4 container's quantization cost and the scale-granularity/rotation ablations live in docs/benchmarks.md and #108/#81.

Get started

You need two things: the program (a few hundred KB) and the model (372 GB). Step-by-step for every platform in the Quick Start guide.

1. Get colibri

Download a prebuilt release — Linux, macOS and Windows, no compiler needed. Take the archive for your platform from Releases and unpack it:

mkdir colibri && tar xzf colibri-v1.8.0-linux-x86_64.tar.gz -C colibri && cd colibri
python3 coli info                         # engine ready ✓

Inside you get the engine (colibri, colibri.exe on Windows), the coli launcher and its Python helpers. Nothing to rename or configure — coli finds the engine next to itself. You only need Python 3 installed: the launcher and the API gateway are Python scripts, while the engine itself is pure C with zero dependencies.

Or build from source — needs gcc (or clang) with OpenMP:

git clone https://github.com/JustVugg/colibri && cd colibri/c
./setup.sh                                # checks gcc/OpenMP, builds, self-tests

Want coli on your PATH? From a checkout, pip install -e . registers it (the engine still lives in c/ — an editable install from the clone, not a wheel).

2. Get the model

A pre-converted GLM-5.2 int4 container is on Hugging Face — use the group-scaled (gs64) build with the int8 MTP head. It is about 372 GB, so put it on a disk with the room, ideally a fast one:

https://huggingface.co/mastouri/GLM-5.2-colibri-int4-g64-with-int8-mtp

GLM-5.3 is the same family and loads with the same engine. It has its own container, also group-scaled (gs64), about 419 GB. It ships without the MTP head, so speculative decoding stays off:

https://huggingface.co/Justvugg/GLM-5.3-colibri-int4-g64

⚠️ Use the gs64 container above, not the older per-row int4 mirrors (mateogrgic/…, jlnsrk/…): those measure ~9pp worse on quality and are the root cause of the original think-mode loops and never-terminating generations in #455. The gs64 container fixed those controlled per-row A/Bs, but it is not a general repetition or EOS-starvation guard. The MTP head must also be int8, not int4 (int4 → 0% draft acceptance, #8): ls -l <model>/out-mtp-* — int8 (correct) is 3527131672 / 5366238584 / 1065950496 as three files, or a single out-mtp-00000.safetensors of 9959321520 bytes (the current upload of the recommended container ships it as one file: same int8 tensors, 777 of them at one byte per element).

Or convert from the FP8 source yourself — one resumable command that never needs the full 756 GB on disk at once:

./coli convert --model /nvme/glm52_i4     # download+convert shard by shard (python, one-time)

Other supported models

GLM-5.2 is the reference model, but the same streaming approach runs six more families. Each is a sibling engine — one C file, its own architecture, the same coli chat / coli serve / coli web front end (the launcher picks the binary from the model's config.json):

What each one needs. These differ a lot, and reading two of them together has confused people into thinking the requirements contradict each other (#191). They do not — they are different models. None of them needs a GPU.

Model	Disk for the weights	RAM	GPU
OLMoE	~7 GB (int8 container)	8 GB	not needed
GLM-5.2/5.3	~372 GB (5.2) / ~419 GB (5.3)	16 GB min, 24 GB comfortable	not needed
GLM-5.3-Flash	~195 GB converted	25 GB (12 GB weights at int4 + expert cache)	not needed
Inkling	~469 GB	25 GB with the int4 dense container, ~120 GB without	not needed
Kimi K3	~1.6 TB	32 GB+	not needed
DeepSeek V4 Flash	~167 GB (REAP 150B: ~85 GB)	16 GB min, 32 GB comfortable	optional; any NVIDIA card from the GTX 10 series up (Pascal/Turing via CUDA_ARCH=portable-pre-ampere NO_TC=1, best on RTX 50) makes prefill 5-10x and decode ~2.5x faster
Qwen3.8-Flash-Next	~185.5 GB (official FP8 checkpoint)	16 GB min, 24 GB comfortable at the default context	not supported; CPU only
Qwen3.6-35B-A3B	~20 GB (int4-gs64 container)	24 GB (needs full RAM residency)	optional; the CUDA VRAM expert tier measured 1.44 -> 10.05 tok/s (7.0x) on two 8 GB cards, output bit-identical to CPU

A GPU only ever makes it faster. Speed is set by your disk, because the experts are streamed from it — expect a fraction of a token per second on a slow drive and a few per second on a fast one with the cache warm.

Qwen3.6 ships three pre-converted containers: int4-gs64 (recommended — measured cosine to the int8 anchor 0.98777 → 0.99313 and KL 0.109 → 0.080 against per-row, i.e. ~44% less quantization error), int4 per-row as the A/B baseline, and KAT-Coder v2.5, which the same engine runs unchanged — any architecture-identical checkpoint works without a code path of its own. With CUDA=1 the VRAM expert tier measured 1.44 → 10.05 tok/s (7.0×) on two 8 GB cards, output bit-identical to the CPU path.

Kimi K3 needs no conversion: its QAT-trained MXFP4 experts are streamed straight from the original Hugging Face shards, and the bf16 dense set is quantized at load time. Long agent sessions can opt into recurrent-state checkpoints (COLI_K3_CKPT=N slots in RAM, or parked on disk with COLI_K3_CKPT_DIR): an edited or follow-up prompt restores the deepest surviving checkpoint and re-prefills only the tail, instead of replaying the whole conversation through the SSM layers. On Vulkan hosts K3_VK_UP=auto sizes the expert tier upload from measured bandwidth. The engine's KDA and MLA paths are validated token-exact in CI against the vendor implementation.

Inkling ships int4 experts but bf16 dense weights (49.4 GB resident); on a host that cannot hold those, inkling.md has a one-pass tool that brings the dense set to 15.3 GB and lets the 975B run on a 25 GB box — with the honest trade-off written down.

3. Run it

COLI_MODEL=/nvme/glm52_i4 ./coli chat     # RAM budget, cache and MTP auto-detected
COLI_MODEL=/nvme/glm52_i4 ./coli plan     # inspect the planned VRAM/RAM/disk placement
COLI_MODEL=/nvme/glm52_i4 ./coli doctor   # read-only readiness check
COLI_MODEL=/nvme/glm52_i4 ./coli doctor --deep  # strict tensors/shards/index/mirror preflight
COLI_MODEL=/nvme/glm52_i4 ./coli tune     # measure and save this machine's fastest safe execution profile
./coli web  --model /nvme/glm52_i4        # API + dashboard, and opens a browser
./coli serve --model /nvme/glm52_i4       # API + dashboard, no browser (headless)

Brio mode: ask a closed question

Most of what people ask a model for is a choice, not a paragraph: which queue, which verdict, which of the four values a field may take. Brio mode hands the engine the options and reads the probability of each one instead of generating: completion_tokens is 0, no answer can fall outside your list, and every answer comes with an entropy, so "the model is not sure" is a number you can put a threshold on. It runs on all nine families, on the same server, and it is opt-in per request: chat is byte-identical for everyone who does not ask for it.

# in the TUI: the same model, told to stop writing
./coli chat --model /nvme/qwen36_i4_gs64
> /brio merge | request changes | close
> 340 lines, 8 files, no tests. CI is green but nothing covers that path.

# from anywhere: one JSON request on the running server
curl -s http://127.0.0.1:8000/v1/brio -H 'Content-Type: application/json' -d '{
  "model": "qwen36",
  "state": "340 lines, 8 files, no tests. CI is green but nothing covers that path.",
  "question": "What should the reviewer do?",
  "options": ["merge", "request changes", "close"]}'

questions asks many things about one document read once, and schema fills a JSON object one field at a time, valid by construction. Measured on Qwen3.6 against generating the same answer on the same CPU box: 2.4x on a four-field schema, 5.7x on four questions about one document. The whole mode, the request and reply shapes, and where it does not help: docs/brio.md. The dashboard has a Brio page as well.

On Windows a release archive ships coli.cmd: double-click it for the quick start, or run coli.cmd chat --model D:\glm52_i4 from cmd or PowerShell. From a source checkout the same commands work with python coli chat --model D:\glm52_i4. The .exe files are the engines, not the launcher: started on their own they have no model to load and exit immediately. The engine at runtime is pure C — python is only used by the one-time converter and the optional API gateway.

The same commands run any of the models

coli reads the model's config.json, picks the matching engine binary, and renders that family's chat template — so nothing about the command line changes between models. Build the engine you want once, then just point COLI_MODEL at the right directory:

make -C c glm                                     # GLM-5.2
make -C c inkling                                 # Inkling
make -C c kimi_k3                                 # Kimi K3

COLI_MODEL=/nvme/glm52_i4      ./coli chat        # TUI
COLI_MODEL=/nvme/inkling_i4    ./coli chat
COLI_MODEL=/nvme/kimi_k3       ./coli chat

./coli web --model /nvme/inkling_i4               # API + dashboard, opens a browser
./coli web --model /nvme/kimi_k3
./coli serve --model /nvme/inkling_i4             # API + dashboard, no browser

For the non-GLM engines coli chat starts the gateway locally and attaches the TUI to it, so the TUI, the API and the dashboard all go through the same arch-aware chat template — you never have to pass the template yourself.

Two things that differ per model, both documented in the per-model page:

• Inkling on a RAM-tight host needs the int4 dense container and a small expert cache: ./coli chat --model /nvme/inkling_i4 --cap 2 (see inkling.md — the default --cap 8 wants ~14 GB of cache on top of the resident set).
• Kimi K3 streams its MXFP4 experts from the original checkpoint, so there is nothing to convert — but the snapshot is ~1.6 TB (see kimi_k3.md).

4. Go deeper

DeepSeek V4

DeepSeek V4 Flash streams the official checkpoint with no conversion: routed experts stay native fp4, the dense set stays fp8-e4m3 with UE8M0 block scales. MLA + DSA sparse attention, 43 layers, 256 routed experts plus one shared, top-6. Supported on x86-64/aarch64 Linux and Windows/MSYS2 (CPU), with an optional CUDA tier (Windows runtime DLL; Linux CUDA=1 direct link, verified under WSL2) that keeps every stage CPU-canonical and falls back per stage.

cd c
make deepseek-v4
python ./coli chat --model /path/to/DeepSeek-V4-Flash --ram 32
# also: coli run / coli serve / coli web
# Windows CUDA tier: make cuda-dsv4-dll CUDA_ARCH=portable  (+ make cuda-dsv4-dg-dll on RTX 50)

Two opt-in GPU levers are new and looking for community numbers, both default off and byte-identical when unset: DSV4_HYBRID=1 splits VRAM-tier misses between the GPU fill branch and the CPU branch using bandwidths measured at runtime, and COLI_CUDA_MOE_DOUBLE=1 (on top of COLI_CUDA_MOE_BATCH=1) prefetches the next layer's full expert set into a second VRAM bank while the current layer computes, falling back to the single bank when VRAM is short. The CUDA tier also runs on Pascal and Turing cards now (GTX 10 / RTX 20 series): build with CUDA_ARCH=portable-pre-ampere NO_TC=1.

Greedy decode and one KV slot. Tool calling is wired through the HTTP gateway with V4's native prompt and DSML call blocks; grammar is not supported. See the per-engine API matrix. Prefix checkpoints (in memory + on disk) make agent sessions and follow-up turns start in seconds after the first prefill of a system prompt. Measured on an RTX 5080 + 2 NVMe: 3324-token prefill 90 s, 8.3k-token first turn ~4 min once, later sessions/turns 6-9 s, decode ~1.6 tok/s at 3k context — see docs/deepseek-v4.md.

Give it RAM. 43 × 256 routed experts are ~137 GiB on disk and a token touches 301 of them, so the expert cache hit rate is what sets tok/s — --ram is the single most valuable knob, and it changes speed only, never output.

Speculative drafting exists and is off. DSpark's markov drafter and full MTP are both implemented and verified: a draft can save forward passes but never change a token, because every accepted token is still the target's own argmax. Measured on real multi-turn chat, they accepted 1 in 15 and 10 in 24, and the rejected-suffix replay of this engine's recurrent attention state cost more than the drafts saved — one 14-token answer took 495 seconds. So V4_DRAFT and V4_MTP default to 0 and the code stays, with the numbers beside it, for whoever retries this on faster storage.

See docs/deepseek-v4.md for the CUDA tier (build, DLL selection, GPU coverage), the environment reference, performance numbers, checkpoint validation, and the generated tiny independent oracle.

What's next

• Inference-systems research is the product. The current hierarchy is LRU + a learned pin set; active work spans model formats, compression, placement, scheduling, I/O, CPU/GPU kernels, heterogeneous overlap, KV state, and routing-aware speculation. The objective is lower hardware requirements and lower cost per useful token. Everything lands the way this project works: measured end to end, reviewed, and developed in the open.
• More open models. The tiering algorithm is model-agnostic: any MoE with routed experts can be staged the same way. Nine families run today (GLM-5.2, GLM-5.3-Flash, Inkling, Kimi K3, DeepSeek V4 Flash, DeepSeek V4.1 Flash, Qwen3.8-Flash-Next, Qwen3.6, OLMoE); further open-weight families — MiniMax among the candidates — earn an engine the way the first eight did: when someone measures one end to end.

Supporting the project

colibrì started as a one-person project on a 12-core laptop with 25 GB of RAM; today its numbers come from a community of real machines. If it's useful to you:

• ⭐ star the repo and share it;
• 🐛 open issues with benchmark numbers from your hardware — datapoints move this project more than anything else;
• 💬 join the Discord community to discuss experiments, hardware results, and research directions;
• 💬 reach out via GitHub issues to sponsor development or donate hardware.

Repo layout

Makefile                  root build/check entry point
c/
├── colibri.c             GLM-5.2 engine  (make glm)
├── inkling.c             Inkling engine  (make inkling)
├── kimi_k3.c             Kimi K3 engine  (make kimi_k3)
├── deepseek_v4.c         DeepSeek V4 Flash engine  (make deepseek-v4)
├── qwen38.c              Qwen3.8-Flash-Next text engine  (make qwen38)
├── qwen36.c              Qwen3.6 engine  (make qwen36)
├── olmoe.c               OLMoE engine  (make olmoe)
│
├── st.h                  safetensors index and range reads
├── quant.h               canonical container decoders
├── expert_ffn.h          routed-expert FFN kernel shared by the MoE engines (planar int4, layer runner)
├── tok.h, json.h         tokenizer and JSON parser
├── compat.h              Windows/macOS shims (POSIX names, one place)
├── expert_store.h        streaming expert cache
├── route_trace.h         routing telemetry and .coli_usage, engine-agnostic
├── kv_prefix.h           KV prefix reuse across turns
│
├── backend_cuda.*        optional CUDA tier   (CUDA=1)
├── backend_metal.*       optional Metal tier  (METAL=1)
├── backend_vulkan.*      optional Vulkan tier (VULKAN=1)
│
├── Makefile              build and local checks
├── coli                  user-facing CLI
├── openai_server.py      OpenAI-compatible HTTP gateway
├── resource_plan.py      RAM/VRAM planner behind `coli plan` and `coli doctor`
├── tools/                offline conversion, fixtures and benchmarks
├── scripts/              long-running conversion helpers
└── tests/                dependency-free C and Python tests
web/                      browser UI (pure OpenAI-API client)
desktop/                  Tauri v2 desktop shell wrapping the web UI
docker/                   container images
docs/                     reference docs, experiments, media

One .c per model family, over shared single headers. An engine owns its architecture and nothing else; anything two engines both need — the safetensors reader, the container decoders, the tokenizer, the expert cache — lives in a header they both include, so a fix reaches all of them at once. That rule is not decorative: the defects that keep recurring here are the ones where a mechanism landed in one engine and never reached its siblings.

From the repository root, make, make check and make clean delegate to the engine Makefile.

Why "colibrì"

The hummingbird weighs a few grams, hovers in place, and visits a thousand flowers a day. This engine keeps a 744-billion-parameter giant alive on hummingbird rations: 25 GB of RAM, twelve CPU cores, and a lot of disk patience.

Acknowledgements

colibrì is an engine; the minds it runs are a gift. Thank you to the teams releasing frontier-class weights in the open — Z.ai (GLM), Moonshot AI (Kimi), Alibaba Qwen, MiniMax, and Allen AI (OLMoE) — and to every contributor who benchmarked, bisected, replicated an atlas run, or sent a patch. This project is proof of what open weights make possible.

The project's expert placement, compression, and routing experiments also build on ideas and evidence from the following open research and systems work:

• REAP and EASY-EP for output-aware and domain-specific expert importance.
• SERE for similarity-based expert re-routing, and ReMoE for cache-locality-aware router fine-tuning.
• MC-SMoE for routing-guided expert merging and compression.
• MoBE and D²-MoE for shared expert bases and low-rank expert deltas.
• HybriMoE for hybrid CPU/GPU expert scheduling, ScMoE for overlapping expert communication with computation, and OD-MoE for distributed on-demand expert loading.
• vLLM, llama.cpp, and kTransformers for the open inference systems and expert-offload work that make comparisons reproducible.

The engine also stands on concrete engineering work, not only ideas. Each of these is used or reimplemented in the tree today:

• safetensors — the container every engine reads (c/st.h), including its fp8 and I64 dtypes.
• tiktoken — c/tok.h reimplements its byte_pair_encode exactly, merging the adjacent pair whose concatenation has the lowest vocab id, so a tiktoken-derived vocabulary needs no merges list.
• llama.cpp — the GBNF grammar subset in c/grammar.h follows its syntax and its set-of-stacks PDA, and the Metal path borrows its newBufferWithBytesNoCopy residency trick.
• vLLM — the reference for output semantics the engine matches position by position (e.g. where the final norm lands relative to the LM head).
• transformers — the oracle: CI reproduces a random-init model token for token against it.
• DietGPU — the GPU ANS codec behind the experimental compressed expert tier (COLI_ANS).
• rocWMMA — the HIP backend maps CUDA's nvcuda::wmma fragment/mma_sync API onto it (c/backend_gpu_compat.h), which is what lets one .cu source compile for both vendors.

License

Apache 2.0, Copyright 2026 Vincenzo Fornaro. See LICENSE and NOTICE. GLM-5.2 weights are released by Z.ai under MIT.