[THE RESEARCH ARM OF REKA]
Frontier models.
Released openly.
Reka Labs trains multimodal foundation models from scratch - and ships the weights, the eval, and the code in the same release. No moats. No roadmap theater.
Reka Edge
REKA MODEL REGISTRY
CORE
67B
multimodal
HIGH-PERFORMANCE
FLASH
21B
multimodal
COST-EFFICIENT
EDGE
7B
FASTEST
SPARK
1B
ULTRA-COMPACT
VIBE-EVAL
benchmark
REKA QUANT
toolkit
5 artifacts loaded
in sync
VIBE-EVAL / CORE 67B
2026.05.15 · 10:22
MMMU
83.1
VIDEO-MME
79.4
DOCVQA
91.2
CHARTQA
86.4
AUIDOBENCH
74.0
LONGVID-32K
71.8
MATHVISTA
68.2
269 / 269 tasks
eval ok
[ONE TOOLKIT - EVERY SURFACE]
Same models.
Wherever you work.
Pull the weights from the CLI. Inspect them on the dashboard. Fine-tune in a notebook. Ship to the edge from a snippet. The artifact is one model — the surfaces are yours to pick.
REKA CLI
HUGGING FACE
EVAL DASHBOARD
~/REKA-LABS
ZSH
# Pull weights, quantize, eval — three commands.
$
reka pull
core-67b
--quant int4
→ fetching shards…
[████████░░] 80%
→ verifying signatures…
ok
$
reka eval
vibe-eval
--model core-67b
running 269 tasks · ~14m est.
overall:
81.3
vision:
83.1
audio:
79.4
$
reka serve
core-67b
—port 7860
→ binding…
ready at http://0.0.0.0:7860
$
[BENCHMARKS]
A benchmark we can't game.
Vibe-Eval is 269 hand-written multimodal tasks, kept private until release. No web scrape can memorize them. No data flywheel can shortcut them. The score on the table is the score you'll get.
TASKS
269
Hand-written, hand-verified, never indexed
MODALITIES
4
Image, video, audio, and long-form text
MODELS TESTED
38+
Across every major vendor, with reproducible logs
[TRAINING ROOM]
Watch the lab work.
A real-time view into the eval queue. Every run, every regression, every release tagged and reproducible.
MISSION CONTROL
REKA-LABS
EVAL QUEUE / LIVE
7 RUNNING
14:22
core-67b · vibe-eval · 269/269 tasks
✓ 81.3
14:18
flash-21b · long-video-bench · 1024/2048
… running
14:09
spark-1b · int4-quant · edge-arm64
✓ ready
14:04
core-67b · doc-vqa-v2
✓ 91.2
13:58
vibe-eval-v2 · regenerate gold
✓ signed
13:42
spark-1b · audio-bench-tiny
✗ retry
13:30
core-67b · chart-qa
✓ 86.4
# Tagging a release.
$
reka release tag
core-67b
--weights ./checkpoints/9a4f2c
→ generating model card…
ok
→ uploading shards (3/6)…
shard-0001.safetensors
2.1 GB ✓
shard-0002.safetensors
2.1 GB ✓
shard-0003.safetensors
2.1 GB ✓
shard-0004.safetensors
… 67%
→ writing eval bundle
vibe-eval.json
→ signing release
ed25519:9a4f2c
$
[BUILT FOR OPEN RESEARCH]
No moats.
No mystery sauce.
Every release ships with the bargain a researcher actually needs: a paper, the weights, the recipe, the eval, the seeds. Nothing held back for the next funding round.
Open weights, real licenses.
Permissive terms across the family. Deploy, fine-tune, quantize, distill — without asking permission.
Paste SVG
Reproducible by default.
Training recipes, eval scripts, and seeds shipped with the paper. The number in the table is the number you'll get.
Paste SVG
Honest evaluation.
Benchmarks we can't game, written by hand and kept private until release. Reward models that don't reward shortcuts.
Paste SVG
One architecture, every size.
From 1B Spark to 67B Core, the same backbone scales. Pick the size that fits the workload, not the headline.
Paste SVG
One team, one stack.
Research, infra, and evaluation under one roof. The model you download is the model we use ourselves.
Paste SVG
Multimodal from the ground up.
Image, video, audio, and text in one shared latent space. No bolt-on adapters, no second-class modalities.
[JOIN THE LAB]
Build the field.
Not just the model.
We're hiring researchers and engineers across vision, audio, and long-context reasoning. If you'd rather ship the weights than write the press release — that's us.
Reka Edge
