Morph MORPH
SIGN IN SIGN UP
dhruvhead / toshllm UNCLAIMED
0 0 0 JavaScript

COMMITS

main
October 3, 2026
E
release: 0.87.14
Engelbert committed
0fbdd28
E
engine: Flash-Next asks for its per-layer embedding rows before reading them and uses the fused hyper-connection kernels on 64-lane cards
Engelbert committed
2f994e4
E
ci: the math runtime is imported once before its tests, so a translated first import does not run into the helper's call timeout
Engelbert committed
7c3167e
E
chat: after a refused math call the model can only correct the call or ask, and a second refusal ends the turn with a fixed message instead of an answer from memory
Engelbert committed
724f549
E
tools: math calls carry the user's own message and are refused before running when they drop, change or invent part of the problem; every result shows what was computed
Engelbert committed
95cd404
October 2, 2026
E
tools: with both math tool sets on, one routing rule keeps exact requests in SymPy, and the numerical helper refuses or flags calls that cannot mean what they say
Engelbert committed
c12c244
E
tools: optional numerical tools backed by NumPy and SciPy in the bundled math runtime, off by default
Engelbert committed
78892c3
E
tools: SymPy says when a result is numerical or timed out, takes e as a variable when declared, and its definitions cost 1017 tokens instead of 2332
Engelbert committed
4fc31b6
E
tools: optional symbolic math tools backed by a bundled SymPy runtime, off by default
Engelbert committed
ab46d16
October 1, 2026
E
release: 0.87.13
Engelbert committed
8af5d37
E
chat: the conversation memory tools archive only when the context is filling up
Engelbert committed
7db08f7
E
models: a model with its MTP head built in is listed even when its file name says MTP
Engelbert committed
938faf2
E
docs: Dynamic MoE performance tables and version 0.87.13
Engelbert committed
2c4d185
E
app: a Save RAM switch, separate MTP heads under Dynamic MoE, model publish dates and a 10 minute start limit
Engelbert committed
81bef71
E
engine: Dynamic MoE can save RAM, counts the MTP head and starts without it when it does not fit, copies evicted experts back from VRAM and frees slots a prompt left half loaded
Engelbert committed
5dc2efc
E
engine: the bounded Dynamic MoE cache keeps filling after the server is ready
Engelbert committed
db62835
E
engine: Dynamic MoE takes batches of 9 tokens and up on the GPU
Engelbert committed
58b0834
E
engine: a state save with nothing to read from the GPU no longer takes a Metal command buffer
Engelbert committed
cd68bdf
E
engine: hybrid models cache their recurrent state in 8 bits and the plan counts their checkpoints
Engelbert committed
3ebe4e4
E
servers: Start uses the settings at the click, the header keeps its layout and model tags wrap
Engelbert committed
1783819
E
engine: Dynamic MoE plans its locked memory by what is free, without a fixed unwired floor
Engelbert committed
43acc04
September 30, 2026
E
Revert "servers: a Dynamic MoE restart never shows the previous run's memory plan"
Engelbert committed
3a93550
E
servers: default reasoning and response limit for requests that bring none
Engelbert committed
0f0a6cb
E
benchmarks: a context recall test against the running server
Engelbert committed
d82e39f
E
servers: a starting server shows its phase and how long it has taken
Engelbert committed
9da49b0
E
servers: a Dynamic MoE restart never shows the previous run's memory plan
Engelbert committed
ad5e8e7
E
patches: fold 0114-0127 into one patch per area
Engelbert committed
74e8a45
E
engine: Dynamic MoE uses the multi-row CPU dots
Engelbert committed
d81b699
E
cpu: a few tokens on one expert share each unpacked block
Engelbert committed
927d1ea
E
cpu: the Prism group-128 dot products run in AVX2
Engelbert committed
f82946d